Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus...
Transcript of Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus...
![Page 1: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/1.jpg)
Climate Assimilation of Data and ModelsWhen Data Fail Us
JUAN M RESTREPOGroup Leader
Uncertainty Quantification Group
University of Arizona
June 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 1 88
UQG Group
Uncertainty Quantification Group
FacultyJMR (Math Physics Atmospheric Sciences)Shankar Venkataramani Kevin Lin Hermann Flaschka (Math)Shlomo Neuman Larry Winter (Hydrology)Rabi Bhattacharya Walt Piegorsch (Math Statistics)Kobus Barnard Alon Efrat (Computer Science)
Post-DocsP Krause (estimation)J Ramırez (probability)P Dostert (scientific computing)T-T Shieh (variational methods)
Graduate Students 15 (Brad Weir Darin Comeau Suz Tolwinski)Undergraduate Students 2 (Jason Ditmann)
UQGQG
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 2 88
UQG Group
Focus ProblemsClimateWeather Hydrogeology Computer Vision
Complex Problems Problems in which progress will result from acombination of statistical methods and deterministic and stochasticmathematics
UQGQG
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 3 88
UQG Group
Focus ProblemsClimateWeather Hydrogeology Computer Vision
Complex Problems Problems in which progress will result from acombination of statistical methods and deterministic and stochasticmathematics
UQGQG
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 3 88
UQG Group
Focus ProblemsClimateWeather Hydrogeology Computer Vision
Complex Problems Problems in which progress will result from acombination of statistical methods and deterministic and stochasticmathematics
UQGQG
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 3 88
UQG Group
Earthrsquos Climate ForcedDissipativeThermodynamicSystem
TIME
SP
AC
E
WEATHER
300 years1 day 1 season
Climate
10 years
continent
country
city
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 4 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
When data fool us
0 100 200 300 400 500 600minus2
minus1
0
1
2
time
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 6 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
When data fool ussame data zoomed in
0 5 10 15 20minus2
minus1
0
1
2
time
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 7 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
use our understanding of the dynamics
dx = 4x(1minus x2)dt +κdWt
x(0) = x0
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 8 88
Time Dependent Estimation
Data Assimilation in Climate Studies
Combine information derived from data and models
Bayes Theorem
P(X|Y) prop likelihoodtimesprior
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 9 88
Time Dependent Estimation LinearGaussian Estimation
Least Squares and Kalman FilterSmoother
W(m)xminusV = Θ
ΘsimN (0σ)
Find x mean such that E(θgtθ) is minimizedFind the uncertainty U = E[(xminus x)(xminus x)gt]
Alternatively can find x and U using a sequential approach the KalmanFilterSmoother (RTS Algorithm)
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 10 88
Time Dependent Estimation LinearGaussian Estimation
Kalman Filter
Forecast
Xlowast = MX(t)+BP(t) t = 01
Ulowast = MU(t)Mgt
Analysis
X(t +1) = Xlowast+K(t +1)[Y(t +1)minusH(t +1)Xlowast]U(t +1) = UlowastK(t +1)H(t +1)Ulowast
where the Kalman Gain Matrix is
K(t +1) = UlowastH(t +1)gt[H(t +1)UlowastH(t +1)gt]minus1
X(0) and U(0) are known
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 11 88
Time Dependent Estimation LinearGaussian Estimation
Nonlinear Non-Gaussian Problems
Forecast not much of a problem
X(t +1) = N(X(t)BP(t))
But not clear how to propagate uncertainty U(t +1)
Extended Kalman Filter used extensively on nonlinear problems linearizeabout X(t) and use closure ideas for moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 12 88
Nonlinear Non-Gaussian Estimation
A Nonlinear Non-Gaussian Example
Let x(t) isinR1 0 lt t le T
Model
dx = 4x(1minus x2)dt +κdWt
x(0) = x0 known distrib
Data
y(tm) = x(tm)+η(tm)E(ηmηl) = Rδtmtl
m = 12 M
minus2 minus15 minus1 minus05 0 05 1 15 2minus2
0
2
4
6
8
10
12
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 13 88
Nonlinear Non-Gaussian Estimation
Double-Well Stationary Distribution
dx = 4x(1minus x2)dt +κdWt
The Stationary distribution for the double-well problem (κ = 05)
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 14 88
Nonlinear Non-Gaussian Estimation
The Observations
0 1 2 3 4 5 6 7minus15
minus1
minus05
0
05
1
15
time
minus2
minus1
5minus
1minus
05
00
51
15
2minus
202468
10
12
A single realization of a random process with known statistics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 15 88
Nonlinear Non-Gaussian Estimation
Goals of Traditional Data Assimilation
Find mean history xS(t) conditioned on observations E(x(t)|y1 yM)and the uncertainty Cs(t) = E[(x(t)minus xs(t))(x(t)minus xs(t))T |y1 yM]xs(t) should minimize the tr(Cs)Guarantee statistical convergence of moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 16 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 2: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/2.jpg)
UQG Group
Uncertainty Quantification Group
FacultyJMR (Math Physics Atmospheric Sciences)Shankar Venkataramani Kevin Lin Hermann Flaschka (Math)Shlomo Neuman Larry Winter (Hydrology)Rabi Bhattacharya Walt Piegorsch (Math Statistics)Kobus Barnard Alon Efrat (Computer Science)
Post-DocsP Krause (estimation)J Ramırez (probability)P Dostert (scientific computing)T-T Shieh (variational methods)
Graduate Students 15 (Brad Weir Darin Comeau Suz Tolwinski)Undergraduate Students 2 (Jason Ditmann)
UQGQG
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 2 88
UQG Group
Focus ProblemsClimateWeather Hydrogeology Computer Vision
Complex Problems Problems in which progress will result from acombination of statistical methods and deterministic and stochasticmathematics
UQGQG
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 3 88
UQG Group
Focus ProblemsClimateWeather Hydrogeology Computer Vision
Complex Problems Problems in which progress will result from acombination of statistical methods and deterministic and stochasticmathematics
UQGQG
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 3 88
UQG Group
Focus ProblemsClimateWeather Hydrogeology Computer Vision
Complex Problems Problems in which progress will result from acombination of statistical methods and deterministic and stochasticmathematics
UQGQG
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 3 88
UQG Group
Earthrsquos Climate ForcedDissipativeThermodynamicSystem
TIME
SP
AC
E
WEATHER
300 years1 day 1 season
Climate
10 years
continent
country
city
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 4 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
When data fool us
0 100 200 300 400 500 600minus2
minus1
0
1
2
time
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 6 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
When data fool ussame data zoomed in
0 5 10 15 20minus2
minus1
0
1
2
time
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 7 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
use our understanding of the dynamics
dx = 4x(1minus x2)dt +κdWt
x(0) = x0
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 8 88
Time Dependent Estimation
Data Assimilation in Climate Studies
Combine information derived from data and models
Bayes Theorem
P(X|Y) prop likelihoodtimesprior
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 9 88
Time Dependent Estimation LinearGaussian Estimation
Least Squares and Kalman FilterSmoother
W(m)xminusV = Θ
ΘsimN (0σ)
Find x mean such that E(θgtθ) is minimizedFind the uncertainty U = E[(xminus x)(xminus x)gt]
Alternatively can find x and U using a sequential approach the KalmanFilterSmoother (RTS Algorithm)
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 10 88
Time Dependent Estimation LinearGaussian Estimation
Kalman Filter
Forecast
Xlowast = MX(t)+BP(t) t = 01
Ulowast = MU(t)Mgt
Analysis
X(t +1) = Xlowast+K(t +1)[Y(t +1)minusH(t +1)Xlowast]U(t +1) = UlowastK(t +1)H(t +1)Ulowast
where the Kalman Gain Matrix is
K(t +1) = UlowastH(t +1)gt[H(t +1)UlowastH(t +1)gt]minus1
X(0) and U(0) are known
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 11 88
Time Dependent Estimation LinearGaussian Estimation
Nonlinear Non-Gaussian Problems
Forecast not much of a problem
X(t +1) = N(X(t)BP(t))
But not clear how to propagate uncertainty U(t +1)
Extended Kalman Filter used extensively on nonlinear problems linearizeabout X(t) and use closure ideas for moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 12 88
Nonlinear Non-Gaussian Estimation
A Nonlinear Non-Gaussian Example
Let x(t) isinR1 0 lt t le T
Model
dx = 4x(1minus x2)dt +κdWt
x(0) = x0 known distrib
Data
y(tm) = x(tm)+η(tm)E(ηmηl) = Rδtmtl
m = 12 M
minus2 minus15 minus1 minus05 0 05 1 15 2minus2
0
2
4
6
8
10
12
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 13 88
Nonlinear Non-Gaussian Estimation
Double-Well Stationary Distribution
dx = 4x(1minus x2)dt +κdWt
The Stationary distribution for the double-well problem (κ = 05)
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 14 88
Nonlinear Non-Gaussian Estimation
The Observations
0 1 2 3 4 5 6 7minus15
minus1
minus05
0
05
1
15
time
minus2
minus1
5minus
1minus
05
00
51
15
2minus
202468
10
12
A single realization of a random process with known statistics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 15 88
Nonlinear Non-Gaussian Estimation
Goals of Traditional Data Assimilation
Find mean history xS(t) conditioned on observations E(x(t)|y1 yM)and the uncertainty Cs(t) = E[(x(t)minus xs(t))(x(t)minus xs(t))T |y1 yM]xs(t) should minimize the tr(Cs)Guarantee statistical convergence of moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 16 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 3: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/3.jpg)
UQG Group
Focus ProblemsClimateWeather Hydrogeology Computer Vision
Complex Problems Problems in which progress will result from acombination of statistical methods and deterministic and stochasticmathematics
UQGQG
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 3 88
UQG Group
Focus ProblemsClimateWeather Hydrogeology Computer Vision
Complex Problems Problems in which progress will result from acombination of statistical methods and deterministic and stochasticmathematics
UQGQG
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 3 88
UQG Group
Focus ProblemsClimateWeather Hydrogeology Computer Vision
Complex Problems Problems in which progress will result from acombination of statistical methods and deterministic and stochasticmathematics
UQGQG
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 3 88
UQG Group
Earthrsquos Climate ForcedDissipativeThermodynamicSystem
TIME
SP
AC
E
WEATHER
300 years1 day 1 season
Climate
10 years
continent
country
city
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 4 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
When data fool us
0 100 200 300 400 500 600minus2
minus1
0
1
2
time
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 6 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
When data fool ussame data zoomed in
0 5 10 15 20minus2
minus1
0
1
2
time
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 7 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
use our understanding of the dynamics
dx = 4x(1minus x2)dt +κdWt
x(0) = x0
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 8 88
Time Dependent Estimation
Data Assimilation in Climate Studies
Combine information derived from data and models
Bayes Theorem
P(X|Y) prop likelihoodtimesprior
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 9 88
Time Dependent Estimation LinearGaussian Estimation
Least Squares and Kalman FilterSmoother
W(m)xminusV = Θ
ΘsimN (0σ)
Find x mean such that E(θgtθ) is minimizedFind the uncertainty U = E[(xminus x)(xminus x)gt]
Alternatively can find x and U using a sequential approach the KalmanFilterSmoother (RTS Algorithm)
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 10 88
Time Dependent Estimation LinearGaussian Estimation
Kalman Filter
Forecast
Xlowast = MX(t)+BP(t) t = 01
Ulowast = MU(t)Mgt
Analysis
X(t +1) = Xlowast+K(t +1)[Y(t +1)minusH(t +1)Xlowast]U(t +1) = UlowastK(t +1)H(t +1)Ulowast
where the Kalman Gain Matrix is
K(t +1) = UlowastH(t +1)gt[H(t +1)UlowastH(t +1)gt]minus1
X(0) and U(0) are known
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 11 88
Time Dependent Estimation LinearGaussian Estimation
Nonlinear Non-Gaussian Problems
Forecast not much of a problem
X(t +1) = N(X(t)BP(t))
But not clear how to propagate uncertainty U(t +1)
Extended Kalman Filter used extensively on nonlinear problems linearizeabout X(t) and use closure ideas for moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 12 88
Nonlinear Non-Gaussian Estimation
A Nonlinear Non-Gaussian Example
Let x(t) isinR1 0 lt t le T
Model
dx = 4x(1minus x2)dt +κdWt
x(0) = x0 known distrib
Data
y(tm) = x(tm)+η(tm)E(ηmηl) = Rδtmtl
m = 12 M
minus2 minus15 minus1 minus05 0 05 1 15 2minus2
0
2
4
6
8
10
12
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 13 88
Nonlinear Non-Gaussian Estimation
Double-Well Stationary Distribution
dx = 4x(1minus x2)dt +κdWt
The Stationary distribution for the double-well problem (κ = 05)
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 14 88
Nonlinear Non-Gaussian Estimation
The Observations
0 1 2 3 4 5 6 7minus15
minus1
minus05
0
05
1
15
time
minus2
minus1
5minus
1minus
05
00
51
15
2minus
202468
10
12
A single realization of a random process with known statistics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 15 88
Nonlinear Non-Gaussian Estimation
Goals of Traditional Data Assimilation
Find mean history xS(t) conditioned on observations E(x(t)|y1 yM)and the uncertainty Cs(t) = E[(x(t)minus xs(t))(x(t)minus xs(t))T |y1 yM]xs(t) should minimize the tr(Cs)Guarantee statistical convergence of moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 16 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 4: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/4.jpg)
UQG Group
Focus ProblemsClimateWeather Hydrogeology Computer Vision
Complex Problems Problems in which progress will result from acombination of statistical methods and deterministic and stochasticmathematics
UQGQG
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 3 88
UQG Group
Focus ProblemsClimateWeather Hydrogeology Computer Vision
Complex Problems Problems in which progress will result from acombination of statistical methods and deterministic and stochasticmathematics
UQGQG
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 3 88
UQG Group
Earthrsquos Climate ForcedDissipativeThermodynamicSystem
TIME
SP
AC
E
WEATHER
300 years1 day 1 season
Climate
10 years
continent
country
city
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 4 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
When data fool us
0 100 200 300 400 500 600minus2
minus1
0
1
2
time
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 6 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
When data fool ussame data zoomed in
0 5 10 15 20minus2
minus1
0
1
2
time
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 7 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
use our understanding of the dynamics
dx = 4x(1minus x2)dt +κdWt
x(0) = x0
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 8 88
Time Dependent Estimation
Data Assimilation in Climate Studies
Combine information derived from data and models
Bayes Theorem
P(X|Y) prop likelihoodtimesprior
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 9 88
Time Dependent Estimation LinearGaussian Estimation
Least Squares and Kalman FilterSmoother
W(m)xminusV = Θ
ΘsimN (0σ)
Find x mean such that E(θgtθ) is minimizedFind the uncertainty U = E[(xminus x)(xminus x)gt]
Alternatively can find x and U using a sequential approach the KalmanFilterSmoother (RTS Algorithm)
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 10 88
Time Dependent Estimation LinearGaussian Estimation
Kalman Filter
Forecast
Xlowast = MX(t)+BP(t) t = 01
Ulowast = MU(t)Mgt
Analysis
X(t +1) = Xlowast+K(t +1)[Y(t +1)minusH(t +1)Xlowast]U(t +1) = UlowastK(t +1)H(t +1)Ulowast
where the Kalman Gain Matrix is
K(t +1) = UlowastH(t +1)gt[H(t +1)UlowastH(t +1)gt]minus1
X(0) and U(0) are known
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 11 88
Time Dependent Estimation LinearGaussian Estimation
Nonlinear Non-Gaussian Problems
Forecast not much of a problem
X(t +1) = N(X(t)BP(t))
But not clear how to propagate uncertainty U(t +1)
Extended Kalman Filter used extensively on nonlinear problems linearizeabout X(t) and use closure ideas for moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 12 88
Nonlinear Non-Gaussian Estimation
A Nonlinear Non-Gaussian Example
Let x(t) isinR1 0 lt t le T
Model
dx = 4x(1minus x2)dt +κdWt
x(0) = x0 known distrib
Data
y(tm) = x(tm)+η(tm)E(ηmηl) = Rδtmtl
m = 12 M
minus2 minus15 minus1 minus05 0 05 1 15 2minus2
0
2
4
6
8
10
12
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 13 88
Nonlinear Non-Gaussian Estimation
Double-Well Stationary Distribution
dx = 4x(1minus x2)dt +κdWt
The Stationary distribution for the double-well problem (κ = 05)
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 14 88
Nonlinear Non-Gaussian Estimation
The Observations
0 1 2 3 4 5 6 7minus15
minus1
minus05
0
05
1
15
time
minus2
minus1
5minus
1minus
05
00
51
15
2minus
202468
10
12
A single realization of a random process with known statistics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 15 88
Nonlinear Non-Gaussian Estimation
Goals of Traditional Data Assimilation
Find mean history xS(t) conditioned on observations E(x(t)|y1 yM)and the uncertainty Cs(t) = E[(x(t)minus xs(t))(x(t)minus xs(t))T |y1 yM]xs(t) should minimize the tr(Cs)Guarantee statistical convergence of moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 16 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 5: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/5.jpg)
UQG Group
Focus ProblemsClimateWeather Hydrogeology Computer Vision
Complex Problems Problems in which progress will result from acombination of statistical methods and deterministic and stochasticmathematics
UQGQG
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 3 88
UQG Group
Earthrsquos Climate ForcedDissipativeThermodynamicSystem
TIME
SP
AC
E
WEATHER
300 years1 day 1 season
Climate
10 years
continent
country
city
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 4 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
When data fool us
0 100 200 300 400 500 600minus2
minus1
0
1
2
time
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 6 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
When data fool ussame data zoomed in
0 5 10 15 20minus2
minus1
0
1
2
time
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 7 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
use our understanding of the dynamics
dx = 4x(1minus x2)dt +κdWt
x(0) = x0
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 8 88
Time Dependent Estimation
Data Assimilation in Climate Studies
Combine information derived from data and models
Bayes Theorem
P(X|Y) prop likelihoodtimesprior
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 9 88
Time Dependent Estimation LinearGaussian Estimation
Least Squares and Kalman FilterSmoother
W(m)xminusV = Θ
ΘsimN (0σ)
Find x mean such that E(θgtθ) is minimizedFind the uncertainty U = E[(xminus x)(xminus x)gt]
Alternatively can find x and U using a sequential approach the KalmanFilterSmoother (RTS Algorithm)
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 10 88
Time Dependent Estimation LinearGaussian Estimation
Kalman Filter
Forecast
Xlowast = MX(t)+BP(t) t = 01
Ulowast = MU(t)Mgt
Analysis
X(t +1) = Xlowast+K(t +1)[Y(t +1)minusH(t +1)Xlowast]U(t +1) = UlowastK(t +1)H(t +1)Ulowast
where the Kalman Gain Matrix is
K(t +1) = UlowastH(t +1)gt[H(t +1)UlowastH(t +1)gt]minus1
X(0) and U(0) are known
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 11 88
Time Dependent Estimation LinearGaussian Estimation
Nonlinear Non-Gaussian Problems
Forecast not much of a problem
X(t +1) = N(X(t)BP(t))
But not clear how to propagate uncertainty U(t +1)
Extended Kalman Filter used extensively on nonlinear problems linearizeabout X(t) and use closure ideas for moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 12 88
Nonlinear Non-Gaussian Estimation
A Nonlinear Non-Gaussian Example
Let x(t) isinR1 0 lt t le T
Model
dx = 4x(1minus x2)dt +κdWt
x(0) = x0 known distrib
Data
y(tm) = x(tm)+η(tm)E(ηmηl) = Rδtmtl
m = 12 M
minus2 minus15 minus1 minus05 0 05 1 15 2minus2
0
2
4
6
8
10
12
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 13 88
Nonlinear Non-Gaussian Estimation
Double-Well Stationary Distribution
dx = 4x(1minus x2)dt +κdWt
The Stationary distribution for the double-well problem (κ = 05)
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 14 88
Nonlinear Non-Gaussian Estimation
The Observations
0 1 2 3 4 5 6 7minus15
minus1
minus05
0
05
1
15
time
minus2
minus1
5minus
1minus
05
00
51
15
2minus
202468
10
12
A single realization of a random process with known statistics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 15 88
Nonlinear Non-Gaussian Estimation
Goals of Traditional Data Assimilation
Find mean history xS(t) conditioned on observations E(x(t)|y1 yM)and the uncertainty Cs(t) = E[(x(t)minus xs(t))(x(t)minus xs(t))T |y1 yM]xs(t) should minimize the tr(Cs)Guarantee statistical convergence of moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 16 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 6: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/6.jpg)
UQG Group
Earthrsquos Climate ForcedDissipativeThermodynamicSystem
TIME
SP
AC
E
WEATHER
300 years1 day 1 season
Climate
10 years
continent
country
city
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 4 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
When data fool us
0 100 200 300 400 500 600minus2
minus1
0
1
2
time
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 6 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
When data fool ussame data zoomed in
0 5 10 15 20minus2
minus1
0
1
2
time
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 7 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
use our understanding of the dynamics
dx = 4x(1minus x2)dt +κdWt
x(0) = x0
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 8 88
Time Dependent Estimation
Data Assimilation in Climate Studies
Combine information derived from data and models
Bayes Theorem
P(X|Y) prop likelihoodtimesprior
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 9 88
Time Dependent Estimation LinearGaussian Estimation
Least Squares and Kalman FilterSmoother
W(m)xminusV = Θ
ΘsimN (0σ)
Find x mean such that E(θgtθ) is minimizedFind the uncertainty U = E[(xminus x)(xminus x)gt]
Alternatively can find x and U using a sequential approach the KalmanFilterSmoother (RTS Algorithm)
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 10 88
Time Dependent Estimation LinearGaussian Estimation
Kalman Filter
Forecast
Xlowast = MX(t)+BP(t) t = 01
Ulowast = MU(t)Mgt
Analysis
X(t +1) = Xlowast+K(t +1)[Y(t +1)minusH(t +1)Xlowast]U(t +1) = UlowastK(t +1)H(t +1)Ulowast
where the Kalman Gain Matrix is
K(t +1) = UlowastH(t +1)gt[H(t +1)UlowastH(t +1)gt]minus1
X(0) and U(0) are known
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 11 88
Time Dependent Estimation LinearGaussian Estimation
Nonlinear Non-Gaussian Problems
Forecast not much of a problem
X(t +1) = N(X(t)BP(t))
But not clear how to propagate uncertainty U(t +1)
Extended Kalman Filter used extensively on nonlinear problems linearizeabout X(t) and use closure ideas for moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 12 88
Nonlinear Non-Gaussian Estimation
A Nonlinear Non-Gaussian Example
Let x(t) isinR1 0 lt t le T
Model
dx = 4x(1minus x2)dt +κdWt
x(0) = x0 known distrib
Data
y(tm) = x(tm)+η(tm)E(ηmηl) = Rδtmtl
m = 12 M
minus2 minus15 minus1 minus05 0 05 1 15 2minus2
0
2
4
6
8
10
12
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 13 88
Nonlinear Non-Gaussian Estimation
Double-Well Stationary Distribution
dx = 4x(1minus x2)dt +κdWt
The Stationary distribution for the double-well problem (κ = 05)
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 14 88
Nonlinear Non-Gaussian Estimation
The Observations
0 1 2 3 4 5 6 7minus15
minus1
minus05
0
05
1
15
time
minus2
minus1
5minus
1minus
05
00
51
15
2minus
202468
10
12
A single realization of a random process with known statistics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 15 88
Nonlinear Non-Gaussian Estimation
Goals of Traditional Data Assimilation
Find mean history xS(t) conditioned on observations E(x(t)|y1 yM)and the uncertainty Cs(t) = E[(x(t)minus xs(t))(x(t)minus xs(t))T |y1 yM]xs(t) should minimize the tr(Cs)Guarantee statistical convergence of moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 16 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 7: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/7.jpg)
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
When data fool us
0 100 200 300 400 500 600minus2
minus1
0
1
2
time
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 6 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
When data fool ussame data zoomed in
0 5 10 15 20minus2
minus1
0
1
2
time
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 7 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
use our understanding of the dynamics
dx = 4x(1minus x2)dt +κdWt
x(0) = x0
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 8 88
Time Dependent Estimation
Data Assimilation in Climate Studies
Combine information derived from data and models
Bayes Theorem
P(X|Y) prop likelihoodtimesprior
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 9 88
Time Dependent Estimation LinearGaussian Estimation
Least Squares and Kalman FilterSmoother
W(m)xminusV = Θ
ΘsimN (0σ)
Find x mean such that E(θgtθ) is minimizedFind the uncertainty U = E[(xminus x)(xminus x)gt]
Alternatively can find x and U using a sequential approach the KalmanFilterSmoother (RTS Algorithm)
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 10 88
Time Dependent Estimation LinearGaussian Estimation
Kalman Filter
Forecast
Xlowast = MX(t)+BP(t) t = 01
Ulowast = MU(t)Mgt
Analysis
X(t +1) = Xlowast+K(t +1)[Y(t +1)minusH(t +1)Xlowast]U(t +1) = UlowastK(t +1)H(t +1)Ulowast
where the Kalman Gain Matrix is
K(t +1) = UlowastH(t +1)gt[H(t +1)UlowastH(t +1)gt]minus1
X(0) and U(0) are known
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 11 88
Time Dependent Estimation LinearGaussian Estimation
Nonlinear Non-Gaussian Problems
Forecast not much of a problem
X(t +1) = N(X(t)BP(t))
But not clear how to propagate uncertainty U(t +1)
Extended Kalman Filter used extensively on nonlinear problems linearizeabout X(t) and use closure ideas for moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 12 88
Nonlinear Non-Gaussian Estimation
A Nonlinear Non-Gaussian Example
Let x(t) isinR1 0 lt t le T
Model
dx = 4x(1minus x2)dt +κdWt
x(0) = x0 known distrib
Data
y(tm) = x(tm)+η(tm)E(ηmηl) = Rδtmtl
m = 12 M
minus2 minus15 minus1 minus05 0 05 1 15 2minus2
0
2
4
6
8
10
12
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 13 88
Nonlinear Non-Gaussian Estimation
Double-Well Stationary Distribution
dx = 4x(1minus x2)dt +κdWt
The Stationary distribution for the double-well problem (κ = 05)
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 14 88
Nonlinear Non-Gaussian Estimation
The Observations
0 1 2 3 4 5 6 7minus15
minus1
minus05
0
05
1
15
time
minus2
minus1
5minus
1minus
05
00
51
15
2minus
202468
10
12
A single realization of a random process with known statistics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 15 88
Nonlinear Non-Gaussian Estimation
Goals of Traditional Data Assimilation
Find mean history xS(t) conditioned on observations E(x(t)|y1 yM)and the uncertainty Cs(t) = E[(x(t)minus xs(t))(x(t)minus xs(t))T |y1 yM]xs(t) should minimize the tr(Cs)Guarantee statistical convergence of moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 16 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 8: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/8.jpg)
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
When data fool us
0 100 200 300 400 500 600minus2
minus1
0
1
2
time
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 6 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
When data fool ussame data zoomed in
0 5 10 15 20minus2
minus1
0
1
2
time
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 7 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
use our understanding of the dynamics
dx = 4x(1minus x2)dt +κdWt
x(0) = x0
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 8 88
Time Dependent Estimation
Data Assimilation in Climate Studies
Combine information derived from data and models
Bayes Theorem
P(X|Y) prop likelihoodtimesprior
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 9 88
Time Dependent Estimation LinearGaussian Estimation
Least Squares and Kalman FilterSmoother
W(m)xminusV = Θ
ΘsimN (0σ)
Find x mean such that E(θgtθ) is minimizedFind the uncertainty U = E[(xminus x)(xminus x)gt]
Alternatively can find x and U using a sequential approach the KalmanFilterSmoother (RTS Algorithm)
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 10 88
Time Dependent Estimation LinearGaussian Estimation
Kalman Filter
Forecast
Xlowast = MX(t)+BP(t) t = 01
Ulowast = MU(t)Mgt
Analysis
X(t +1) = Xlowast+K(t +1)[Y(t +1)minusH(t +1)Xlowast]U(t +1) = UlowastK(t +1)H(t +1)Ulowast
where the Kalman Gain Matrix is
K(t +1) = UlowastH(t +1)gt[H(t +1)UlowastH(t +1)gt]minus1
X(0) and U(0) are known
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 11 88
Time Dependent Estimation LinearGaussian Estimation
Nonlinear Non-Gaussian Problems
Forecast not much of a problem
X(t +1) = N(X(t)BP(t))
But not clear how to propagate uncertainty U(t +1)
Extended Kalman Filter used extensively on nonlinear problems linearizeabout X(t) and use closure ideas for moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 12 88
Nonlinear Non-Gaussian Estimation
A Nonlinear Non-Gaussian Example
Let x(t) isinR1 0 lt t le T
Model
dx = 4x(1minus x2)dt +κdWt
x(0) = x0 known distrib
Data
y(tm) = x(tm)+η(tm)E(ηmηl) = Rδtmtl
m = 12 M
minus2 minus15 minus1 minus05 0 05 1 15 2minus2
0
2
4
6
8
10
12
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 13 88
Nonlinear Non-Gaussian Estimation
Double-Well Stationary Distribution
dx = 4x(1minus x2)dt +κdWt
The Stationary distribution for the double-well problem (κ = 05)
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 14 88
Nonlinear Non-Gaussian Estimation
The Observations
0 1 2 3 4 5 6 7minus15
minus1
minus05
0
05
1
15
time
minus2
minus1
5minus
1minus
05
00
51
15
2minus
202468
10
12
A single realization of a random process with known statistics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 15 88
Nonlinear Non-Gaussian Estimation
Goals of Traditional Data Assimilation
Find mean history xS(t) conditioned on observations E(x(t)|y1 yM)and the uncertainty Cs(t) = E[(x(t)minus xs(t))(x(t)minus xs(t))T |y1 yM]xs(t) should minimize the tr(Cs)Guarantee statistical convergence of moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 16 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 9: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/9.jpg)
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
When data fool us
0 100 200 300 400 500 600minus2
minus1
0
1
2
time
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 6 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
When data fool ussame data zoomed in
0 5 10 15 20minus2
minus1
0
1
2
time
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 7 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
use our understanding of the dynamics
dx = 4x(1minus x2)dt +κdWt
x(0) = x0
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 8 88
Time Dependent Estimation
Data Assimilation in Climate Studies
Combine information derived from data and models
Bayes Theorem
P(X|Y) prop likelihoodtimesprior
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 9 88
Time Dependent Estimation LinearGaussian Estimation
Least Squares and Kalman FilterSmoother
W(m)xminusV = Θ
ΘsimN (0σ)
Find x mean such that E(θgtθ) is minimizedFind the uncertainty U = E[(xminus x)(xminus x)gt]
Alternatively can find x and U using a sequential approach the KalmanFilterSmoother (RTS Algorithm)
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 10 88
Time Dependent Estimation LinearGaussian Estimation
Kalman Filter
Forecast
Xlowast = MX(t)+BP(t) t = 01
Ulowast = MU(t)Mgt
Analysis
X(t +1) = Xlowast+K(t +1)[Y(t +1)minusH(t +1)Xlowast]U(t +1) = UlowastK(t +1)H(t +1)Ulowast
where the Kalman Gain Matrix is
K(t +1) = UlowastH(t +1)gt[H(t +1)UlowastH(t +1)gt]minus1
X(0) and U(0) are known
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 11 88
Time Dependent Estimation LinearGaussian Estimation
Nonlinear Non-Gaussian Problems
Forecast not much of a problem
X(t +1) = N(X(t)BP(t))
But not clear how to propagate uncertainty U(t +1)
Extended Kalman Filter used extensively on nonlinear problems linearizeabout X(t) and use closure ideas for moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 12 88
Nonlinear Non-Gaussian Estimation
A Nonlinear Non-Gaussian Example
Let x(t) isinR1 0 lt t le T
Model
dx = 4x(1minus x2)dt +κdWt
x(0) = x0 known distrib
Data
y(tm) = x(tm)+η(tm)E(ηmηl) = Rδtmtl
m = 12 M
minus2 minus15 minus1 minus05 0 05 1 15 2minus2
0
2
4
6
8
10
12
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 13 88
Nonlinear Non-Gaussian Estimation
Double-Well Stationary Distribution
dx = 4x(1minus x2)dt +κdWt
The Stationary distribution for the double-well problem (κ = 05)
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 14 88
Nonlinear Non-Gaussian Estimation
The Observations
0 1 2 3 4 5 6 7minus15
minus1
minus05
0
05
1
15
time
minus2
minus1
5minus
1minus
05
00
51
15
2minus
202468
10
12
A single realization of a random process with known statistics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 15 88
Nonlinear Non-Gaussian Estimation
Goals of Traditional Data Assimilation
Find mean history xS(t) conditioned on observations E(x(t)|y1 yM)and the uncertainty Cs(t) = E[(x(t)minus xs(t))(x(t)minus xs(t))T |y1 yM]xs(t) should minimize the tr(Cs)Guarantee statistical convergence of moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 16 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 10: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/10.jpg)
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
When data fool us
0 100 200 300 400 500 600minus2
minus1
0
1
2
time
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 6 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
When data fool ussame data zoomed in
0 5 10 15 20minus2
minus1
0
1
2
time
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 7 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
use our understanding of the dynamics
dx = 4x(1minus x2)dt +κdWt
x(0) = x0
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 8 88
Time Dependent Estimation
Data Assimilation in Climate Studies
Combine information derived from data and models
Bayes Theorem
P(X|Y) prop likelihoodtimesprior
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 9 88
Time Dependent Estimation LinearGaussian Estimation
Least Squares and Kalman FilterSmoother
W(m)xminusV = Θ
ΘsimN (0σ)
Find x mean such that E(θgtθ) is minimizedFind the uncertainty U = E[(xminus x)(xminus x)gt]
Alternatively can find x and U using a sequential approach the KalmanFilterSmoother (RTS Algorithm)
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 10 88
Time Dependent Estimation LinearGaussian Estimation
Kalman Filter
Forecast
Xlowast = MX(t)+BP(t) t = 01
Ulowast = MU(t)Mgt
Analysis
X(t +1) = Xlowast+K(t +1)[Y(t +1)minusH(t +1)Xlowast]U(t +1) = UlowastK(t +1)H(t +1)Ulowast
where the Kalman Gain Matrix is
K(t +1) = UlowastH(t +1)gt[H(t +1)UlowastH(t +1)gt]minus1
X(0) and U(0) are known
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 11 88
Time Dependent Estimation LinearGaussian Estimation
Nonlinear Non-Gaussian Problems
Forecast not much of a problem
X(t +1) = N(X(t)BP(t))
But not clear how to propagate uncertainty U(t +1)
Extended Kalman Filter used extensively on nonlinear problems linearizeabout X(t) and use closure ideas for moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 12 88
Nonlinear Non-Gaussian Estimation
A Nonlinear Non-Gaussian Example
Let x(t) isinR1 0 lt t le T
Model
dx = 4x(1minus x2)dt +κdWt
x(0) = x0 known distrib
Data
y(tm) = x(tm)+η(tm)E(ηmηl) = Rδtmtl
m = 12 M
minus2 minus15 minus1 minus05 0 05 1 15 2minus2
0
2
4
6
8
10
12
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 13 88
Nonlinear Non-Gaussian Estimation
Double-Well Stationary Distribution
dx = 4x(1minus x2)dt +κdWt
The Stationary distribution for the double-well problem (κ = 05)
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 14 88
Nonlinear Non-Gaussian Estimation
The Observations
0 1 2 3 4 5 6 7minus15
minus1
minus05
0
05
1
15
time
minus2
minus1
5minus
1minus
05
00
51
15
2minus
202468
10
12
A single realization of a random process with known statistics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 15 88
Nonlinear Non-Gaussian Estimation
Goals of Traditional Data Assimilation
Find mean history xS(t) conditioned on observations E(x(t)|y1 yM)and the uncertainty Cs(t) = E[(x(t)minus xs(t))(x(t)minus xs(t))T |y1 yM]xs(t) should minimize the tr(Cs)Guarantee statistical convergence of moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 16 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 11: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/11.jpg)
UQG Group
The Meteorology vs Climate Problem
Climate poor andor sparse data Very large spatio-temporal ranges(oceans) Challenging interations oceanrsquos interaction with theatmosphere and ice
Aim is to discern variability and understanding dynamicsBayesian data assimilation is viable strategyNeed to make more use of models need to develop parameterizationsNeed to focus more on non-Gaussian issues as well as the use of dataassimilation combined with time of transit dynamics
Meteorology lots of data very nonlinear very stiffSocietal imperative forecastsEnsemble forecastingReduced state space representationLarge deviation theory could be explored here
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 5 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
When data fool us
0 100 200 300 400 500 600minus2
minus1
0
1
2
time
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 6 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
When data fool ussame data zoomed in
0 5 10 15 20minus2
minus1
0
1
2
time
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 7 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
use our understanding of the dynamics
dx = 4x(1minus x2)dt +κdWt
x(0) = x0
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 8 88
Time Dependent Estimation
Data Assimilation in Climate Studies
Combine information derived from data and models
Bayes Theorem
P(X|Y) prop likelihoodtimesprior
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 9 88
Time Dependent Estimation LinearGaussian Estimation
Least Squares and Kalman FilterSmoother
W(m)xminusV = Θ
ΘsimN (0σ)
Find x mean such that E(θgtθ) is minimizedFind the uncertainty U = E[(xminus x)(xminus x)gt]
Alternatively can find x and U using a sequential approach the KalmanFilterSmoother (RTS Algorithm)
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 10 88
Time Dependent Estimation LinearGaussian Estimation
Kalman Filter
Forecast
Xlowast = MX(t)+BP(t) t = 01
Ulowast = MU(t)Mgt
Analysis
X(t +1) = Xlowast+K(t +1)[Y(t +1)minusH(t +1)Xlowast]U(t +1) = UlowastK(t +1)H(t +1)Ulowast
where the Kalman Gain Matrix is
K(t +1) = UlowastH(t +1)gt[H(t +1)UlowastH(t +1)gt]minus1
X(0) and U(0) are known
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 11 88
Time Dependent Estimation LinearGaussian Estimation
Nonlinear Non-Gaussian Problems
Forecast not much of a problem
X(t +1) = N(X(t)BP(t))
But not clear how to propagate uncertainty U(t +1)
Extended Kalman Filter used extensively on nonlinear problems linearizeabout X(t) and use closure ideas for moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 12 88
Nonlinear Non-Gaussian Estimation
A Nonlinear Non-Gaussian Example
Let x(t) isinR1 0 lt t le T
Model
dx = 4x(1minus x2)dt +κdWt
x(0) = x0 known distrib
Data
y(tm) = x(tm)+η(tm)E(ηmηl) = Rδtmtl
m = 12 M
minus2 minus15 minus1 minus05 0 05 1 15 2minus2
0
2
4
6
8
10
12
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 13 88
Nonlinear Non-Gaussian Estimation
Double-Well Stationary Distribution
dx = 4x(1minus x2)dt +κdWt
The Stationary distribution for the double-well problem (κ = 05)
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 14 88
Nonlinear Non-Gaussian Estimation
The Observations
0 1 2 3 4 5 6 7minus15
minus1
minus05
0
05
1
15
time
minus2
minus1
5minus
1minus
05
00
51
15
2minus
202468
10
12
A single realization of a random process with known statistics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 15 88
Nonlinear Non-Gaussian Estimation
Goals of Traditional Data Assimilation
Find mean history xS(t) conditioned on observations E(x(t)|y1 yM)and the uncertainty Cs(t) = E[(x(t)minus xs(t))(x(t)minus xs(t))T |y1 yM]xs(t) should minimize the tr(Cs)Guarantee statistical convergence of moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 16 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 12: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/12.jpg)
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
When data fool us
0 100 200 300 400 500 600minus2
minus1
0
1
2
time
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 6 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
When data fool ussame data zoomed in
0 5 10 15 20minus2
minus1
0
1
2
time
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 7 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
use our understanding of the dynamics
dx = 4x(1minus x2)dt +κdWt
x(0) = x0
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 8 88
Time Dependent Estimation
Data Assimilation in Climate Studies
Combine information derived from data and models
Bayes Theorem
P(X|Y) prop likelihoodtimesprior
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 9 88
Time Dependent Estimation LinearGaussian Estimation
Least Squares and Kalman FilterSmoother
W(m)xminusV = Θ
ΘsimN (0σ)
Find x mean such that E(θgtθ) is minimizedFind the uncertainty U = E[(xminus x)(xminus x)gt]
Alternatively can find x and U using a sequential approach the KalmanFilterSmoother (RTS Algorithm)
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 10 88
Time Dependent Estimation LinearGaussian Estimation
Kalman Filter
Forecast
Xlowast = MX(t)+BP(t) t = 01
Ulowast = MU(t)Mgt
Analysis
X(t +1) = Xlowast+K(t +1)[Y(t +1)minusH(t +1)Xlowast]U(t +1) = UlowastK(t +1)H(t +1)Ulowast
where the Kalman Gain Matrix is
K(t +1) = UlowastH(t +1)gt[H(t +1)UlowastH(t +1)gt]minus1
X(0) and U(0) are known
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 11 88
Time Dependent Estimation LinearGaussian Estimation
Nonlinear Non-Gaussian Problems
Forecast not much of a problem
X(t +1) = N(X(t)BP(t))
But not clear how to propagate uncertainty U(t +1)
Extended Kalman Filter used extensively on nonlinear problems linearizeabout X(t) and use closure ideas for moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 12 88
Nonlinear Non-Gaussian Estimation
A Nonlinear Non-Gaussian Example
Let x(t) isinR1 0 lt t le T
Model
dx = 4x(1minus x2)dt +κdWt
x(0) = x0 known distrib
Data
y(tm) = x(tm)+η(tm)E(ηmηl) = Rδtmtl
m = 12 M
minus2 minus15 minus1 minus05 0 05 1 15 2minus2
0
2
4
6
8
10
12
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 13 88
Nonlinear Non-Gaussian Estimation
Double-Well Stationary Distribution
dx = 4x(1minus x2)dt +κdWt
The Stationary distribution for the double-well problem (κ = 05)
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 14 88
Nonlinear Non-Gaussian Estimation
The Observations
0 1 2 3 4 5 6 7minus15
minus1
minus05
0
05
1
15
time
minus2
minus1
5minus
1minus
05
00
51
15
2minus
202468
10
12
A single realization of a random process with known statistics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 15 88
Nonlinear Non-Gaussian Estimation
Goals of Traditional Data Assimilation
Find mean history xS(t) conditioned on observations E(x(t)|y1 yM)and the uncertainty Cs(t) = E[(x(t)minus xs(t))(x(t)minus xs(t))T |y1 yM]xs(t) should minimize the tr(Cs)Guarantee statistical convergence of moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 16 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 13: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/13.jpg)
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
When data fool ussame data zoomed in
0 5 10 15 20minus2
minus1
0
1
2
time
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 7 88
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
use our understanding of the dynamics
dx = 4x(1minus x2)dt +κdWt
x(0) = x0
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 8 88
Time Dependent Estimation
Data Assimilation in Climate Studies
Combine information derived from data and models
Bayes Theorem
P(X|Y) prop likelihoodtimesprior
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 9 88
Time Dependent Estimation LinearGaussian Estimation
Least Squares and Kalman FilterSmoother
W(m)xminusV = Θ
ΘsimN (0σ)
Find x mean such that E(θgtθ) is minimizedFind the uncertainty U = E[(xminus x)(xminus x)gt]
Alternatively can find x and U using a sequential approach the KalmanFilterSmoother (RTS Algorithm)
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 10 88
Time Dependent Estimation LinearGaussian Estimation
Kalman Filter
Forecast
Xlowast = MX(t)+BP(t) t = 01
Ulowast = MU(t)Mgt
Analysis
X(t +1) = Xlowast+K(t +1)[Y(t +1)minusH(t +1)Xlowast]U(t +1) = UlowastK(t +1)H(t +1)Ulowast
where the Kalman Gain Matrix is
K(t +1) = UlowastH(t +1)gt[H(t +1)UlowastH(t +1)gt]minus1
X(0) and U(0) are known
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 11 88
Time Dependent Estimation LinearGaussian Estimation
Nonlinear Non-Gaussian Problems
Forecast not much of a problem
X(t +1) = N(X(t)BP(t))
But not clear how to propagate uncertainty U(t +1)
Extended Kalman Filter used extensively on nonlinear problems linearizeabout X(t) and use closure ideas for moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 12 88
Nonlinear Non-Gaussian Estimation
A Nonlinear Non-Gaussian Example
Let x(t) isinR1 0 lt t le T
Model
dx = 4x(1minus x2)dt +κdWt
x(0) = x0 known distrib
Data
y(tm) = x(tm)+η(tm)E(ηmηl) = Rδtmtl
m = 12 M
minus2 minus15 minus1 minus05 0 05 1 15 2minus2
0
2
4
6
8
10
12
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 13 88
Nonlinear Non-Gaussian Estimation
Double-Well Stationary Distribution
dx = 4x(1minus x2)dt +κdWt
The Stationary distribution for the double-well problem (κ = 05)
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 14 88
Nonlinear Non-Gaussian Estimation
The Observations
0 1 2 3 4 5 6 7minus15
minus1
minus05
0
05
1
15
time
minus2
minus1
5minus
1minus
05
00
51
15
2minus
202468
10
12
A single realization of a random process with known statistics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 15 88
Nonlinear Non-Gaussian Estimation
Goals of Traditional Data Assimilation
Find mean history xS(t) conditioned on observations E(x(t)|y1 yM)and the uncertainty Cs(t) = E[(x(t)minus xs(t))(x(t)minus xs(t))T |y1 yM]xs(t) should minimize the tr(Cs)Guarantee statistical convergence of moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 16 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 14: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/14.jpg)
UQG Group
What Models tell Us about Data the Non-Gaussian Issue
use our understanding of the dynamics
dx = 4x(1minus x2)dt +κdWt
x(0) = x0
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 8 88
Time Dependent Estimation
Data Assimilation in Climate Studies
Combine information derived from data and models
Bayes Theorem
P(X|Y) prop likelihoodtimesprior
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 9 88
Time Dependent Estimation LinearGaussian Estimation
Least Squares and Kalman FilterSmoother
W(m)xminusV = Θ
ΘsimN (0σ)
Find x mean such that E(θgtθ) is minimizedFind the uncertainty U = E[(xminus x)(xminus x)gt]
Alternatively can find x and U using a sequential approach the KalmanFilterSmoother (RTS Algorithm)
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 10 88
Time Dependent Estimation LinearGaussian Estimation
Kalman Filter
Forecast
Xlowast = MX(t)+BP(t) t = 01
Ulowast = MU(t)Mgt
Analysis
X(t +1) = Xlowast+K(t +1)[Y(t +1)minusH(t +1)Xlowast]U(t +1) = UlowastK(t +1)H(t +1)Ulowast
where the Kalman Gain Matrix is
K(t +1) = UlowastH(t +1)gt[H(t +1)UlowastH(t +1)gt]minus1
X(0) and U(0) are known
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 11 88
Time Dependent Estimation LinearGaussian Estimation
Nonlinear Non-Gaussian Problems
Forecast not much of a problem
X(t +1) = N(X(t)BP(t))
But not clear how to propagate uncertainty U(t +1)
Extended Kalman Filter used extensively on nonlinear problems linearizeabout X(t) and use closure ideas for moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 12 88
Nonlinear Non-Gaussian Estimation
A Nonlinear Non-Gaussian Example
Let x(t) isinR1 0 lt t le T
Model
dx = 4x(1minus x2)dt +κdWt
x(0) = x0 known distrib
Data
y(tm) = x(tm)+η(tm)E(ηmηl) = Rδtmtl
m = 12 M
minus2 minus15 minus1 minus05 0 05 1 15 2minus2
0
2
4
6
8
10
12
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 13 88
Nonlinear Non-Gaussian Estimation
Double-Well Stationary Distribution
dx = 4x(1minus x2)dt +κdWt
The Stationary distribution for the double-well problem (κ = 05)
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 14 88
Nonlinear Non-Gaussian Estimation
The Observations
0 1 2 3 4 5 6 7minus15
minus1
minus05
0
05
1
15
time
minus2
minus1
5minus
1minus
05
00
51
15
2minus
202468
10
12
A single realization of a random process with known statistics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 15 88
Nonlinear Non-Gaussian Estimation
Goals of Traditional Data Assimilation
Find mean history xS(t) conditioned on observations E(x(t)|y1 yM)and the uncertainty Cs(t) = E[(x(t)minus xs(t))(x(t)minus xs(t))T |y1 yM]xs(t) should minimize the tr(Cs)Guarantee statistical convergence of moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 16 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 15: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/15.jpg)
Time Dependent Estimation
Data Assimilation in Climate Studies
Combine information derived from data and models
Bayes Theorem
P(X|Y) prop likelihoodtimesprior
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 9 88
Time Dependent Estimation LinearGaussian Estimation
Least Squares and Kalman FilterSmoother
W(m)xminusV = Θ
ΘsimN (0σ)
Find x mean such that E(θgtθ) is minimizedFind the uncertainty U = E[(xminus x)(xminus x)gt]
Alternatively can find x and U using a sequential approach the KalmanFilterSmoother (RTS Algorithm)
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 10 88
Time Dependent Estimation LinearGaussian Estimation
Kalman Filter
Forecast
Xlowast = MX(t)+BP(t) t = 01
Ulowast = MU(t)Mgt
Analysis
X(t +1) = Xlowast+K(t +1)[Y(t +1)minusH(t +1)Xlowast]U(t +1) = UlowastK(t +1)H(t +1)Ulowast
where the Kalman Gain Matrix is
K(t +1) = UlowastH(t +1)gt[H(t +1)UlowastH(t +1)gt]minus1
X(0) and U(0) are known
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 11 88
Time Dependent Estimation LinearGaussian Estimation
Nonlinear Non-Gaussian Problems
Forecast not much of a problem
X(t +1) = N(X(t)BP(t))
But not clear how to propagate uncertainty U(t +1)
Extended Kalman Filter used extensively on nonlinear problems linearizeabout X(t) and use closure ideas for moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 12 88
Nonlinear Non-Gaussian Estimation
A Nonlinear Non-Gaussian Example
Let x(t) isinR1 0 lt t le T
Model
dx = 4x(1minus x2)dt +κdWt
x(0) = x0 known distrib
Data
y(tm) = x(tm)+η(tm)E(ηmηl) = Rδtmtl
m = 12 M
minus2 minus15 minus1 minus05 0 05 1 15 2minus2
0
2
4
6
8
10
12
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 13 88
Nonlinear Non-Gaussian Estimation
Double-Well Stationary Distribution
dx = 4x(1minus x2)dt +κdWt
The Stationary distribution for the double-well problem (κ = 05)
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 14 88
Nonlinear Non-Gaussian Estimation
The Observations
0 1 2 3 4 5 6 7minus15
minus1
minus05
0
05
1
15
time
minus2
minus1
5minus
1minus
05
00
51
15
2minus
202468
10
12
A single realization of a random process with known statistics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 15 88
Nonlinear Non-Gaussian Estimation
Goals of Traditional Data Assimilation
Find mean history xS(t) conditioned on observations E(x(t)|y1 yM)and the uncertainty Cs(t) = E[(x(t)minus xs(t))(x(t)minus xs(t))T |y1 yM]xs(t) should minimize the tr(Cs)Guarantee statistical convergence of moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 16 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 16: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/16.jpg)
Time Dependent Estimation LinearGaussian Estimation
Least Squares and Kalman FilterSmoother
W(m)xminusV = Θ
ΘsimN (0σ)
Find x mean such that E(θgtθ) is minimizedFind the uncertainty U = E[(xminus x)(xminus x)gt]
Alternatively can find x and U using a sequential approach the KalmanFilterSmoother (RTS Algorithm)
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 10 88
Time Dependent Estimation LinearGaussian Estimation
Kalman Filter
Forecast
Xlowast = MX(t)+BP(t) t = 01
Ulowast = MU(t)Mgt
Analysis
X(t +1) = Xlowast+K(t +1)[Y(t +1)minusH(t +1)Xlowast]U(t +1) = UlowastK(t +1)H(t +1)Ulowast
where the Kalman Gain Matrix is
K(t +1) = UlowastH(t +1)gt[H(t +1)UlowastH(t +1)gt]minus1
X(0) and U(0) are known
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 11 88
Time Dependent Estimation LinearGaussian Estimation
Nonlinear Non-Gaussian Problems
Forecast not much of a problem
X(t +1) = N(X(t)BP(t))
But not clear how to propagate uncertainty U(t +1)
Extended Kalman Filter used extensively on nonlinear problems linearizeabout X(t) and use closure ideas for moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 12 88
Nonlinear Non-Gaussian Estimation
A Nonlinear Non-Gaussian Example
Let x(t) isinR1 0 lt t le T
Model
dx = 4x(1minus x2)dt +κdWt
x(0) = x0 known distrib
Data
y(tm) = x(tm)+η(tm)E(ηmηl) = Rδtmtl
m = 12 M
minus2 minus15 minus1 minus05 0 05 1 15 2minus2
0
2
4
6
8
10
12
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 13 88
Nonlinear Non-Gaussian Estimation
Double-Well Stationary Distribution
dx = 4x(1minus x2)dt +κdWt
The Stationary distribution for the double-well problem (κ = 05)
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 14 88
Nonlinear Non-Gaussian Estimation
The Observations
0 1 2 3 4 5 6 7minus15
minus1
minus05
0
05
1
15
time
minus2
minus1
5minus
1minus
05
00
51
15
2minus
202468
10
12
A single realization of a random process with known statistics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 15 88
Nonlinear Non-Gaussian Estimation
Goals of Traditional Data Assimilation
Find mean history xS(t) conditioned on observations E(x(t)|y1 yM)and the uncertainty Cs(t) = E[(x(t)minus xs(t))(x(t)minus xs(t))T |y1 yM]xs(t) should minimize the tr(Cs)Guarantee statistical convergence of moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 16 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 17: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/17.jpg)
Time Dependent Estimation LinearGaussian Estimation
Kalman Filter
Forecast
Xlowast = MX(t)+BP(t) t = 01
Ulowast = MU(t)Mgt
Analysis
X(t +1) = Xlowast+K(t +1)[Y(t +1)minusH(t +1)Xlowast]U(t +1) = UlowastK(t +1)H(t +1)Ulowast
where the Kalman Gain Matrix is
K(t +1) = UlowastH(t +1)gt[H(t +1)UlowastH(t +1)gt]minus1
X(0) and U(0) are known
Review in C Wunsch The Ocean Inverse Circulation Problem Cambridge U Press
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 11 88
Time Dependent Estimation LinearGaussian Estimation
Nonlinear Non-Gaussian Problems
Forecast not much of a problem
X(t +1) = N(X(t)BP(t))
But not clear how to propagate uncertainty U(t +1)
Extended Kalman Filter used extensively on nonlinear problems linearizeabout X(t) and use closure ideas for moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 12 88
Nonlinear Non-Gaussian Estimation
A Nonlinear Non-Gaussian Example
Let x(t) isinR1 0 lt t le T
Model
dx = 4x(1minus x2)dt +κdWt
x(0) = x0 known distrib
Data
y(tm) = x(tm)+η(tm)E(ηmηl) = Rδtmtl
m = 12 M
minus2 minus15 minus1 minus05 0 05 1 15 2minus2
0
2
4
6
8
10
12
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 13 88
Nonlinear Non-Gaussian Estimation
Double-Well Stationary Distribution
dx = 4x(1minus x2)dt +κdWt
The Stationary distribution for the double-well problem (κ = 05)
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 14 88
Nonlinear Non-Gaussian Estimation
The Observations
0 1 2 3 4 5 6 7minus15
minus1
minus05
0
05
1
15
time
minus2
minus1
5minus
1minus
05
00
51
15
2minus
202468
10
12
A single realization of a random process with known statistics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 15 88
Nonlinear Non-Gaussian Estimation
Goals of Traditional Data Assimilation
Find mean history xS(t) conditioned on observations E(x(t)|y1 yM)and the uncertainty Cs(t) = E[(x(t)minus xs(t))(x(t)minus xs(t))T |y1 yM]xs(t) should minimize the tr(Cs)Guarantee statistical convergence of moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 16 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 18: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/18.jpg)
Time Dependent Estimation LinearGaussian Estimation
Nonlinear Non-Gaussian Problems
Forecast not much of a problem
X(t +1) = N(X(t)BP(t))
But not clear how to propagate uncertainty U(t +1)
Extended Kalman Filter used extensively on nonlinear problems linearizeabout X(t) and use closure ideas for moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 12 88
Nonlinear Non-Gaussian Estimation
A Nonlinear Non-Gaussian Example
Let x(t) isinR1 0 lt t le T
Model
dx = 4x(1minus x2)dt +κdWt
x(0) = x0 known distrib
Data
y(tm) = x(tm)+η(tm)E(ηmηl) = Rδtmtl
m = 12 M
minus2 minus15 minus1 minus05 0 05 1 15 2minus2
0
2
4
6
8
10
12
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 13 88
Nonlinear Non-Gaussian Estimation
Double-Well Stationary Distribution
dx = 4x(1minus x2)dt +κdWt
The Stationary distribution for the double-well problem (κ = 05)
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 14 88
Nonlinear Non-Gaussian Estimation
The Observations
0 1 2 3 4 5 6 7minus15
minus1
minus05
0
05
1
15
time
minus2
minus1
5minus
1minus
05
00
51
15
2minus
202468
10
12
A single realization of a random process with known statistics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 15 88
Nonlinear Non-Gaussian Estimation
Goals of Traditional Data Assimilation
Find mean history xS(t) conditioned on observations E(x(t)|y1 yM)and the uncertainty Cs(t) = E[(x(t)minus xs(t))(x(t)minus xs(t))T |y1 yM]xs(t) should minimize the tr(Cs)Guarantee statistical convergence of moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 16 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 19: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/19.jpg)
Nonlinear Non-Gaussian Estimation
A Nonlinear Non-Gaussian Example
Let x(t) isinR1 0 lt t le T
Model
dx = 4x(1minus x2)dt +κdWt
x(0) = x0 known distrib
Data
y(tm) = x(tm)+η(tm)E(ηmηl) = Rδtmtl
m = 12 M
minus2 minus15 minus1 minus05 0 05 1 15 2minus2
0
2
4
6
8
10
12
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 13 88
Nonlinear Non-Gaussian Estimation
Double-Well Stationary Distribution
dx = 4x(1minus x2)dt +κdWt
The Stationary distribution for the double-well problem (κ = 05)
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 14 88
Nonlinear Non-Gaussian Estimation
The Observations
0 1 2 3 4 5 6 7minus15
minus1
minus05
0
05
1
15
time
minus2
minus1
5minus
1minus
05
00
51
15
2minus
202468
10
12
A single realization of a random process with known statistics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 15 88
Nonlinear Non-Gaussian Estimation
Goals of Traditional Data Assimilation
Find mean history xS(t) conditioned on observations E(x(t)|y1 yM)and the uncertainty Cs(t) = E[(x(t)minus xs(t))(x(t)minus xs(t))T |y1 yM]xs(t) should minimize the tr(Cs)Guarantee statistical convergence of moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 16 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 20: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/20.jpg)
Nonlinear Non-Gaussian Estimation
Double-Well Stationary Distribution
dx = 4x(1minus x2)dt +κdWt
The Stationary distribution for the double-well problem (κ = 05)
minus3 minus2 minus1 0 1 2 30
01
02
03
04
05
06
07
08SOLUTION
X
HOT
COLD
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 14 88
Nonlinear Non-Gaussian Estimation
The Observations
0 1 2 3 4 5 6 7minus15
minus1
minus05
0
05
1
15
time
minus2
minus1
5minus
1minus
05
00
51
15
2minus
202468
10
12
A single realization of a random process with known statistics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 15 88
Nonlinear Non-Gaussian Estimation
Goals of Traditional Data Assimilation
Find mean history xS(t) conditioned on observations E(x(t)|y1 yM)and the uncertainty Cs(t) = E[(x(t)minus xs(t))(x(t)minus xs(t))T |y1 yM]xs(t) should minimize the tr(Cs)Guarantee statistical convergence of moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 16 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 21: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/21.jpg)
Nonlinear Non-Gaussian Estimation
The Observations
0 1 2 3 4 5 6 7minus15
minus1
minus05
0
05
1
15
time
minus2
minus1
5minus
1minus
05
00
51
15
2minus
202468
10
12
A single realization of a random process with known statistics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 15 88
Nonlinear Non-Gaussian Estimation
Goals of Traditional Data Assimilation
Find mean history xS(t) conditioned on observations E(x(t)|y1 yM)and the uncertainty Cs(t) = E[(x(t)minus xs(t))(x(t)minus xs(t))T |y1 yM]xs(t) should minimize the tr(Cs)Guarantee statistical convergence of moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 16 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 22: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/22.jpg)
Nonlinear Non-Gaussian Estimation
Goals of Traditional Data Assimilation
Find mean history xS(t) conditioned on observations E(x(t)|y1 yM)and the uncertainty Cs(t) = E[(x(t)minus xs(t))(x(t)minus xs(t))T |y1 yM]xs(t) should minimize the tr(Cs)Guarantee statistical convergence of moments
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 16 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 23: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/23.jpg)
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 24: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/24.jpg)
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 25: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/25.jpg)
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 26: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/26.jpg)
Nonlinear Non-Gaussian Estimation
The EKF Results1
Figure 10 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 1
Figure 20 uncertainty ∆t = 025
1R Miller M Ghil P Gauthiez Advanced data assimilation in strongly nonlineardynamical systems J Atmo Sci 51 1037-1056 (1994)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 17 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 27: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/27.jpg)
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 28: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/28.jpg)
Nonlinear Non-Gaussian Estimation
Other Approaches on NonlinearNon-Gaussian Problems
Optimal (variance-minimizer) KSP (Kushner Stratonovich Pardoux) early 60rsquos
4D-VarAdjoint (Maximum Likelihood) (Wunsch McLaughlin Courtier late 80rsquos)
ensemble KF (Evensen rsquo97)
Mean Field Variational (Eyink Restrepo rsquo01)
Parametrized Resampling Particle Filter (Kim Eyink Restrepo Alexander Johnson rsquo02)
Path Integral Monte Carlo (Restrepo rsquo07 Alexander Eyink amp Restrepo rsquo05)
Diffusion Kernel Filter (Krause Restrepo rsquo09)
Langevin Sampler (A Stuart rsquo05)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 18 88
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 29: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/29.jpg)
Nonlinear Non-Gaussian Estimation
KSP Optimalso why not use it
Filter between tm and tm+1 solve
parttP = minuspartx[f (x)P]+12
κ2partxxP
P(x0) = Ps(x)
At measurement times tm the probability jumps
P(x t+) =1N
e1
R2 [ymminus x2m2 ]P(x tminus)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 19 88
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 30: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/30.jpg)
Nonlinear Non-Gaussian Estimation
Smoother between tm+1 and tm solve
parttA = minuspartx[f (x)]A +12
κ2partxxA
A (x tf ) = 1
At measurement times tm the probability jumps
A (x tminus) =1N
e1
R2 [ymminus x2m2 ]A (x t+)
Finally P(x t|ymm = 12 M) = A (x t)P(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 20 88
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 31: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/31.jpg)
Nonlinear Non-Gaussian Estimation
KSP Filter and Smoother Results
Figure KSP Filter
Figure KSP Smoother
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 21 88
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 32: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/32.jpg)
Nonlinear Non-Gaussian Estimation
enKF Most Favored in Practice
The enKF (rdquostate-of-the-artrdquo)
Use model for forecast
Update the uncertainty using Monte Carlo
Pros and Cons
Can handle legacy code easily
Linear (Gaussian) analysis
Requires full model runs
Ad-hoc
G Evensen Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast errorstatistics J Geophys Res 99 10143-10162
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 22 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 33: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/33.jpg)
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
PIMC The Path Integral Monte Carlo
Optimal on the discretized model
Simple to implement but very subtle
Can handle legacy code
Relies on sampling
Can yield a variety of different estimators
J Restrepo A Path Integral Method for Data Assimilation Physica D 2007F Alexander G Eyink J Restrepo Accelerated Monte-Carlo for Optimal Estimation of Time Series J Stat Phys 2005
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 23 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 34: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/34.jpg)
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Bayesian Statement
P(x|y) prop LikelihoodtimesPrior
Use data for likelihood
Use model for prior
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 24 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 35: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/35.jpg)
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Amodel
dx = f (x t)dt +[2D(x t)]12dW
is discretized
xn+1 = xn +∆tf (xn tn)+ [2D(xn tn)]12[Wn+1minusWn]
n = 01 Tminus1
Amodel asympT
sumn=1
[(xn+1minus xnminus∆tf (xn tn))gtD(xn tn)minus1 (xn+1minus xnminus∆tf (xn tn))]
if Prob(∆W) prop exp(minus∆W2D)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 25 88
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 36: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/36.jpg)
Nonlinear Non-Gaussian Estimation The Path Integral Monte Carlo (PIMC)
Adata
ym = H(xm)+ [2R[xm tm)]12ηm
m = 12 M
Adata =M
summ=1
[(ymminusH(xm))gtR(xn tn)minus1 (ymminusH(xm))]
if Prob(η) prop exp(minusη2R)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 26 88
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 37: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/37.jpg)
Nonlinear Non-Gaussian Estimation Samplers
MCMC Samplers
P(x|y) prop eminusAmodeleminusAdata = eminusA (x)
The Path Integral Monte Carlo practicality depends on fast sampling
Multigrid (UMC)
Langevin Sampler (LS)
Hybrid Monte Carlo (HMC)
Shadow Hybrid MC (sHMC)
Riemannian Manifold Hamiltonian Monte Carlo (RM-HMC)
generalized Hybrid Monte Carlo (gHMC)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 27 88
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 38: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/38.jpg)
Nonlinear Non-Gaussian Estimation Samplers
(HMC) Hybrid Markov Chain Monte Carlo
Proposals generated bysolving Hamiltoniansystem in fictitious timeτ
Acceptreject viaMetropolis Hastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 28 88
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 39: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/39.jpg)
Nonlinear Non-Gaussian Estimation Samplers
HMC Algorithm
Let qn(τ = 0) = xn
To each qn a conjugate generalized momemtum pn is assigned
The momenta pn give rise to a kinetic contributionK = sum
Tn=1 pgtn Mminus1pn2
The Hamiltonian of the system H = A (q)+K(p)The dynamics are
partqn
partτ= Mminus1pn
partpn
partτ= Fn where Fn =minusgrad(A (q))
Solve using Verlet integrator (detailed balance)
AcceptReject MetropolisHastings
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 29 88
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 40: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/40.jpg)
Nonlinear Non-Gaussian Estimation Samplers
Why does HMC work What are good HMC properties
Write probability Π(q) = 1ZΠ
eminusA (q)
Sampling π(qp) = 1Zπ
eminusH (qp) sim 1Z eminusA (q) samples Π(q)
Gradient dynamics makes system search through configuration spacemore efficiently
Moves in qn are linear in pn ie partqpartτ
= Mminus1p
A (q) and grad(A (q)) should be easily evaluated
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 30 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 41: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/41.jpg)
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Estimates
Sampler Efficiency key to choosing and tuning sampler
Computational Cost O(NT)r nmethod(pL)p =lt Pacc gt=lt min1exp[minus∆H ]gtprop erfc
(12 δτm(NT)12
)
c(L) =lt H (0)H (0+L) gt Depends on problem dimension and statespace characteristics
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 31 88
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 42: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/42.jpg)
Nonlinear Non-Gaussian Estimation Samplers
RM-HMC Algorithm2
Hamiltonian replaced by
H = A (q)+12
pgtG(q)minus1p
where the non-degenerate Fisher information matrix G = EnablaA nablaA gt
Challenges
find a time-reversiblevolume-preserving discrete integrator forHamiltonian problem
optimize its computational efficiency
2Girolami Calderhead Chin preprint 2009JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 32 88
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 43: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/43.jpg)
Nonlinear Non-Gaussian Estimation Samplers
Decrease decorrelation length L gHMC Algorithm
Hamiltonian dynamics replaced by
partqn
partτ= CMminus1pn
partpn
partτ= CgtF(qn)
where C isinRTtimesT matrix
Challenge find C that leads to a significant reduction in the sampledecorrelation length
We used the circulant matrix C = circ(1eminusα eminus2α eminusTα)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 33 88
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 44: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/44.jpg)
Nonlinear Non-Gaussian Estimation Samplers
Sampler Efficiency Comparison
Table T is the number of time steps (middot) is the standard deviation on the number ofsamples [α] used in C J is the number of τ time steps
T +1 HMC (J=1) HMC (J=8) UMC gHMC (J=1)8 900(125) 170(7) 800(40) 40(8) [020]16 5300(1600) 560(20) 1040(60) 60(10) [010]32 13300(8300) 2700 (140) 1430(100) 200(30) [005]64 30000(7800) 2800(400) 1570(100) 420(70) [00245]
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 34 88
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 45: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/45.jpg)
The DKF
The Diffusion Kernel Filter
Collaborator
Paul Krause UQG U Arizona
A particle-filter based method
p(x|ym tm) =p(ym|x) p(x tm)
p(ym)
SequentialSimple to implement provided an adjoint existsThe Clustered cDKF is competitive with EKFCan handle nonlinearnon-Gaussian problems
P Krause and J Restrepo The Diffusion Kernel Filter Applied to Lagrangian Data Assimilation Mon Wea Rev 2009P Krause and J Restrepo Sequential Estimation with Deterministic Models and Noisy Data SIAM J Sci Comp 2009
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 35 88
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 46: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/46.jpg)
The DKF
The Diffusion Kernel
Φ(middot) = φ(middot)+Φprime(middot)
Φprime(middot) is obtained by applyingDuhamelrsquos principle andexpectation projections
Φprime(ξ (tm i) t)sim
int t
tmG dw(sminus tm)
where
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
is the Diffusion Kernel
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 36 88
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 47: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/47.jpg)
The DKF
The Uncertainty Norm
The diffusion kernel is
G(ξ (tm i) ts) = nablaφ(ξ (tm i) t) g(Φ(ξ (tm i)s))
The uncertainty norm ||G||(t) for branches of prediction bounds the sup-normof the covariance matrix of Φprime(ξ (tm i) t)
1N||cov(Φprime)||infin le ||G||2 tm le t le tm+1
The key observation is that ||cov(Φprime)||infin a measure of entropy of the branchwill be small whenever ||G|| is small
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 37 88
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 48: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/48.jpg)
Application Lagrangian Data Assimilation
The Oceanic Lagrangian Setting
Figure 2 CONTROL run layer thickness displacement (h hrest) superimposed to the
velocity field in each layer (L1L2L3) at initial time (DAY0) and at the end (DAY120) of
the simulation Contour intervals are 25m and the maximum velocity field is approximately
50cms in L1 10cms in L2 5cms in L3
Figure 3 Left column shows successive drifters positions (one point each half day) in
each layer (L1L2L3) Right column shows the corresponding Lagrangian autocovariance
functions (in cm2s2) versus time lag (in days) The solid (dashed) line shows the meridional
(zonal) component
Figure Synthetic Eulerian Flow LagrangianTracks From Molcard et al J Atmos OceanTech (2005)
Figure Four MLFIIrsquos Ready forHurricane Isidore after Puget SoundTesting in July 2002 From EDrsquoAsarorsquos (Washington) Web Page
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 38 88
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 49: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/49.jpg)
Application Lagrangian Data Assimilation
2D VortexDrifter Problem
Measure the tracks of Np passive tracers ζi(t) = vm(t)+ iwm(t) give anoptimal estimate of tracks and of Nv point vortices zm(t) = xm(t)+ iym(t)
Vortices
dzm
dt=
i2π
Nv
suml=1lm
Γl
zlowastmminus zlowastl+η
V(t)
Drifters
dζn
dt=
i2π
Np
suml=1
Γl
ζ lowastn minus zlowastl+η
P(t)
minus2 minus1 0 1 2
minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
particlesvortexvortex
L Kusnetsov and K Ide and CKRT Jones A method for assimilation of Lagrangian data Mon Wea Rev 131 (2003)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 39 88
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 50: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/50.jpg)
Application Lagrangian Data Assimilation
EKF Lagrangian Estimation
Kusnetsov et al proposed using Constrained EKF
for reasonable data uncertainties andor frequent enough observationsEKF works reasonably well based on the computation of distancebetween the rdquotruerdquo path and the EKF estimate
if pushed too hard the EKF would fail saddles in the orbit paths wouldcause divergences in the forecast stage
reasonably efficient robust
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 40 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 51: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/51.jpg)
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and EKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Mean drifter path (v1w1) as estimated by DKF and benchmark and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 41 88
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 52: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/52.jpg)
Application Lagrangian Data Assimilation
Comparison of Bootstrap DKF and EKF
minus2 minus1 0 1 2minus2
minus15
minus1
minus05
0
05
1
15
2
X
Y
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 42 88
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 53: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/53.jpg)
Application Lagrangian Data Assimilation
(Short-time) Behavior of Bootstrap DKF and EKF
minus1 minus05 0 05 1 15minus05
0
05
1
15
2
t=0t1
DKF
EKF
Mean drifter path (v1w1) as estimated by DKF and particle filter and EKF(red dashed) rdquoTruthrdquo path (black) Measurements (green dashed)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 43 88
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 54: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/54.jpg)
Application Lagrangian Data Assimilation
Comparison of Benchmark DKF and enKF
minus2 minus15 minus1 minus05 0 05 1 15 2
minus15
minus1
minus05
0
05
1
15
X
Y
Figure Mean drifter path (v1w1) as estimated by DKF and benchmark and enKF(jagged dashed) The true path is superimposed (black)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 44 88
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 55: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/55.jpg)
Application Lagrangian Data Assimilation
SECOND MOMENT Comparison of Benchmark DKFenKF and EKF
(a)0 2 4 6 8 10
0
001
002
003
004
005
006
007
time
2n
d m
om
en
t
(b)0 2 4 6 8 10
0
005
01
time2
nd
mo
me
nt
Comparison of DKF benchmark filter for v1(t) Second moment (a) EKF (b)enKF
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 45 88
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 56: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/56.jpg)
Application Lagrangian Data Assimilation
THIRD MOMENT Comparison of Benchmark DKF andenKF
0 2 4 6 8 10minus002
minus001
0
001
002
003
004
time
3rd
mo
me
nt
Comparison of enKF DKF and benchmark filter moment estimates for v1(t)Third moment EKF does not generate a third moment
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 46 88
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 57: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/57.jpg)
Application Lagrangian Data Assimilation Computational Cost EKF vs Particle Filter vs DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)
Nx is the dimension of state variable
T is the number of deterministic time steps
αT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
EKF vs DKF Cost
In the prediction step both methods are comparable
In the analysis step EKF is minN3x N3
y for DKF it is Ik timesN2x
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 47 88
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 58: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/58.jpg)
Application Lagrangian Data Assimilation Computational Cost Clustered DKF
Computational Efficiency from tk to tk+1
Bootstrap Filter Cost
CβαTtimesN2x times I
I sample paths (5times105 in examples)Nx is the dimension of state variableT is the number of deterministic time stepsαT is the number of times steps taken in the stochastic differential equation integrator α 1 β gt 1 (SODE time step in examples10minus4 T = 10minus2)
DKF Cost
Ttimes (CprimeNx +CN3x )times Ik asymp TtimesCN3
x times Ik
Conditions whereby the cost of the bootstrap filter exceeds that of the DKF
Nx lt αβ IIk
Reflects the impact of non-Gaussianity in the cost of the DKF the less Gaussian the closer IIk is to 1
cDKF Cost
Can use Ik le 10minusr I r ge 1
In examples r = 2 Conditions whereby the cost of bootstrap exceeds that of rDKF
Nx lt αβ10r
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 48 88
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 59: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/59.jpg)
New Projects
In nonlinear problems there is no unique best predictor
Shown average entropy average likelihood average
from P Krause J M Restrepo to be submitted SIAM J Sci Comp 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 49 88
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 60: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/60.jpg)
New Projects
Lagrangian Data Blending
Data Blending Contour Dynamics
23
4
1
23
23
4
1
2
3
2
1
23
4
2
3
23
1
ClassicCont ourAnalysis
ψ ψ
ψ ψ
(a) (b)
(c) (d)
1 2
A competing approach B Beechler J Weiss G Duane J Tribbia to appear in J Atmos Sci 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 50 88
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 61: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/61.jpg)
New Projects
The Meteorology vs Climate Problem
the data assimilation problem will remain dimensionally-challenging
Develop statistically-convergent methods
Increase the availability and quality of the dataReduction of the state space
Exploit vastly different degrees of uncertaintyDevelop closures and stochastic parameterizations
all the while being clear about what it is that we are doing
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 51 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 62: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/62.jpg)
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 63: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/63.jpg)
State-Space Reduction
The Ensemble Bred Vectors
GOAL
Propose an alternative and robust methodology for sensitivity analysisand reduced-space representation
Provide a mathematical basis for the Bred Vector (BV)
COLLABORATORSNusret Balci (IMAUMN) Anna Mazzucato (PSU) George Sell (UMN)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 52 88
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 64: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/64.jpg)
State-Space Reduction
The Bred Vectors Defined
Given a model
Y(tn +δ tn) = M(Y(tn) tn) n = 012
Y(0) = Y0
Y(tn)asymp y(t = tn) while Y0 asymp y0
BVs computed as follows
δYn+1 = M(Yn +δYn tn)minusM(Yn tn)δYn+1 = RδYn+1
R =δY0δYn+1
the normalization factor
See Toth and Kalnay 1993 Bull Amer Met Soc Toth and Kalnay Mon Wea Rev 1997
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 53 88
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 65: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/65.jpg)
State-Space Reduction
The Lyapunov Vectors Defined
Let δΘ be the solution of the initial value problem
δΘ(tn +δ tn) = LnδΘ(tn) n = 012
δΘ(0) = δΘ0
Take nrarr infinAt each time step the Tangent Linear Model is
Ln = L(Yn tn) =partG(y t)
party
∣∣∣∣y=Ynt=tn
LV make sense in asymptotic time (the BV is local in time)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 54 88
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 66: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/66.jpg)
State-Space Reduction
The Lorenz63
0 5 10 15 20 25 30 35 40 45 50minus2
0
2
4
X
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Y
0 5 10 15 20 25 30 35 40 45 50minus5
0
5
Z
Time
dxdt
= σ(yminus x)
dydt
= x(rminus z)minus y
dzdt
= xyminusbz t gt 0 (1)
At t = 0 x = 08001 y = 04314 z = 09106 Here we set r = 28b = 83σ = 10
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 55 88
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 67: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/67.jpg)
State-Space Reduction
The Cahn Hilliard (Cessi-Young 92 Model)
partSpart t
= αpart 2
part z2 [f (z)+ microS(Sminus sin(z))2 +Sminus γpart 2Spart z2 ] t gt 0z isin [minusππ]
S(z0) = S0(z)
minus3 minus2 minus1 0 1 2 3
minus8
minus6
minus4
minus2
0
2
Zminus3 minus2 minus1 0 1 2 3
minus02
minus015
minus01
minus005
0
005
01
015
02
Z
Forcing function and perturbation
time
Z
0 10 20 30 40 50 60minus3
minus2
minus1
0
1
2
3
minus1
minus05
0
05
minus3 minus2 minus1 0 1 2 3
minus04
minus02
0
02
04
Z
LVBV
Solution of CY92 and LV vs BVJUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 56 88
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 68: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/68.jpg)
State-Space Reduction
Numerical Instability of Bred Vectors
dxdt
= Ax
where A is non-normal constant (Jordan) matrix of dim(20)
0 5 10 15 20 25 300
005
01
015
02
025
03
035
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
2
4
6
8
10
12
14
16
Time
standard BV computation
0 5 10 15 20 25 300
005
01
015
02
025
Time
Bred Vectors Norm=2
0 5 10 15 20 25 300
5
10
15
20
25
Time
Well-conditioned computation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 57 88
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 69: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/69.jpg)
State-Space Reduction
The Ensemble Bred Vector (EBV) Defined
The Ensemble Bred VectorPick a norm and its size for an ensemble of initial conditions
Advance the ensemble of initial conditions using the BV
find the element from the ensemble that grew the most and use it toresize the whole ensemble
repeat
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 58 88
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 70: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/70.jpg)
State-Space Reduction
The Lorenz63 using EBV
minus01minus005000501
minus01
0
01
minus01
minus005
0
005
01
XY
Z
minus010
01
minus01minus005000501
minus01
minus005
0
005
01
YX
Z
Superposition of an ensemble of 1500 runs over all times from 0 to 50 (Weshow 150 of these) normalized to the initial condition 2-norm set to 01
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 59 88
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 71: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/71.jpg)
State-Space Reduction
The Cahn Hilliard CY92 using EBV
BVrsquos
minus3 minus2 minus1 0 1 2 3minus03
minus02
minus01
0
01
02
03
Z
EBVrsquos
minus3 minus2 minus1 0 1 2 3
minus004
minus002
0
002
004
006
Z
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 60 88
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 72: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/72.jpg)
State-Space Reduction
The Non-Normal Linear Problem using EBV
dxdt
= Ax
where A is non-normal and dim(20)
time
N
0 10 20 30 40 50 60
2
4
6
8
10
12
14
16
18
20
minus18
minus16
minus14
minus12
minus1
minus08
minus06
minus04
minus02
5 10 15 20minus2
minus15
minus1
minus05
0
N
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 61 88
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 73: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/73.jpg)
New Projects
Bred Vectors Some of our Conclusions
BV are attractive because they can be applied to legacy code
Established that BV is a nonlinear generalization of the LV though it isnot clear how a finite-time quantity is to be compared to anasymptotically-defined quantity unless we are discussing a steady state
Established that BV is norm dependent
In applications the norm matters a great deal when considering what is tobe analyzed
But do they provide any practical information not contained in singularvectors or LV
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 62 88
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 74: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/74.jpg)
Stochastic Parameterization
Dimension Reduction via Stochastic Parameterization
dx = f (x)dt + sochastic term
x(0) = x0
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 63 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 75: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/75.jpg)
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 76: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/76.jpg)
Stochastic Parameterization
Stochastic Parameterization
GOAL
Produce stochastically-based subscale parameterization
Make greater contact with data (and data assimilation)
COLLABORATORSDarin Comeau (U Arizona)Brad Weir (U Arizona) Jorge Ramırez (U Nacional de Colombia) J CMcWilliams (UCLA) M Banner (UNSW)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 64 88
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 77: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/77.jpg)
Stochastic Parameterization
WAVE BREAKING DISSIPATION
Geophysical GoalsHow does dissipation at wave scales manifest itself at longer time scales
What is the correct stochastic parameterization
Can we parameterize dissipation in such a way that we use field datamore efficiently
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 65 88
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 78: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/78.jpg)
Stochastic Parameterization
Scale Range of the Model
10 secs-months
100m-basin scale
Speed waves gt currents
kH sim 1Applications
climate dynamics (transport)erodible bed dynamicsriver plume evolutionalgalplankton bloomsOil spillspollution
J McWilliams and J M Restrepo The Wave-driven OceanCirculation J Phys Oceanogr (1999)J Restrepo Wave-Current Interactions and Shore-connected Bars J Estuarine Sci (2001)J McWilliams J M Restrepo E Lane An asymptotic Theory for the Interaction of Waves and Currents in Coastal Waters J Fluid Mechanics(2004)E Lane J M Restrepo J McWilliams Wave-Current Interaction A comparison of radiation-stress and vortex-force representations J PhysOceanogr (2007)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 66 88
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 79: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/79.jpg)
Stochastic Parameterization
Lagrangian Motion
dX = Vdt
Figure Deterministic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 67 88
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 80: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/80.jpg)
Stochastic Parameterization
Lagrangian Motion Under White Capping
dXt = V(X t)dt +dWt(X t)
Figure Stochastic
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 68 88
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 81: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/81.jpg)
Stochastic Parameterization
What Stochastic Model
dXt = V(X t)dt +dWt(X t)
Experiments are needed to determine the right model
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW standardwhite noise
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedmean-reverting process
minus02 minus015 minus01 minus005 0 005 01 015 02 025minus02
minus01
0
01
X
Z
Figure dW with addedjump process
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 69 88
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 82: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/82.jpg)
Stochastic Parameterization ProjectionFiltering Strategy
LagrangianEulerian Projections in Multiscale Setting
BASIC STRATEGYMultiscale projection of systems of (mostly) hyperbolic equations betweenthe Eulerian and Lagrangian frames
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 70 88
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 83: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/83.jpg)
Stochastic Parameterization Model
CurrentWavesBreaking Model
The momentum with breaking-generated stresses and diffusion
partvc
partT= VtimesZminusnablaΦ+ 〈(btimesZ)+btimes (nablatimesb)+ [Vtimesnablatimesb]minus 1
2nabla|b|2〉+nabla middotR
where V = vc +ustokes Z = nablatimesv+2Ω and Φ = p0 + 12 |V|
2
The ustokes is the wave contributionThe break-stresses modify the vortex force and Bernoulli headThe diffusion accounts for mixed-layer boundary-layer effects
The tracers obeypartCpartT
+V middotnablaC =minusb middotnablaC+nabla middotQ
J M Restrepo Wave Breaking Dissipation in the Wave-Driven Ocean Circulation J Phys Oceanogr (2007)JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents J PhysOceanogr 2010
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 71 88
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 84: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/84.jpg)
Stochastic Parameterization Model
Wave Breaking Diffusion
The observation is that wave breaking increases the size of the mixing layerand this layer will then create a great deal of dissipation
Rv asymp νpartvh
part z Rh asymp νnablavh
Qv asymp κpartCpart z
Qh asymp κnablaC
We assume thatν sim 〈`b
∣∣wb∣∣〉 κ sim 〈`θ
∣∣wb∣∣〉
wb is the vertical component of the velocity associated with breaking and themixing length is
`b = γη(x t) `θ = αη(x t)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 72 88
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 85: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/85.jpg)
Stochastic Parameterization Model
Boundary Conditions
The surface boundary conditions at z = η(xh t) are the following
w =Dη
Dt p = gρ0η +pa ν
partqpart z
=1ρ0
τ κpartCpart z
= T
Lead to (at z = 0)
wc = nabla middotMminuswb pc = ηc +pa0minusP
ν
(partvc
part z+S)
= τminusνpartbc
part z κ
partCpart z
= T
where the wave-induced adjustments (at z = 0) are
Mequivlang
uwη
wrang Pequiv
langpw
z ηwrang Sequiv
langpart 2uw
part z2 ηwrang
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 73 88
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 86: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/86.jpg)
Determining b
How to determine b(XT)
Thus far we have answered the question
if breaking occurs how do waves and currents get modified at theselarge spatio-temporal scales
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 74 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 87: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/87.jpg)
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 88: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/88.jpg)
Determining b
How to determine b(XT)
How is the breaking velocity b determined
it can be determined by field dataor
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 75 88
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 89: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/89.jpg)
Determining b
Steps to determine b(XT)
Ingredients in determining b using averaging and breaking kinematics
individual breaker b velocity are obtained parametrically from data
breaking events need to be upscaled to yield b
kinematics of their occurrence can be tied to wave-group dynamics
breaking events should be energetically consistent
breaking events should yield a Poisson process in space-time
Details in JM Restrepo J M Ramırez JC McWilliams amp M Banner Wave Breaking Dissipation and Diffusion in Waves and Currents JPhys Oceanogr 2010B Weir amp J M Restrepo Stability of the Langmuir Circulation in the Presence of Wave Breaking in preparation
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 76 88
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 90: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/90.jpg)
Determining b
Individual breaker velocity b are obtained parametrically
Individual breaker velocity b given by
0 02 04 06 08 1minus1
minus09
minus08
minus07
minus06
minus05
minus04
minus03
minus02
minus01
0
50
100
150
200
250
300
parttb+ b middotnablab =1
Re∆b+A
nabla middot b = 0
where
A = kbX (x)Y (y)Z (z)T (t)a
cf Sullivan McWilliams Melville JFM 507 (2004)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 77 88
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 91: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/91.jpg)
Determining b
Breaking events need to be upscaled to yield b
Model b = B+bprime as the random sum
b(xhzT) = sum(Xhτ)isinΦ
bE(Xhτ)(xhminusXhz)δ (τminusT)
wherebE(xhminusXhz) =
1τE
intτE
0b(xhminusXhz t)dt
The ensemble average at (xhzT) of some functional F of the field b is
〈F (b)〉(xh zT)dT =
langsum
(Xh τ)isinΦ
F (bE(Xh τ)(xhminusXh z))δ (Tminus τ)
rang
=intR
intxhminusΩE
F (bE(xhminusXh z))Λ(dXhdT)p(E)dE
B(z) =intR
[intΩE
bE(XhzT)dXhdT]
λ p(E)dE
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 78 88
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 92: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/92.jpg)
Determining b
Kinematics of their occurrence can be tied to wave-groupdynamics
Waves expressed in terms of a carrier and an envelope
η(Xh t) = Re
ei(khmiddotxh+σ t)ρ(xh t)eiθ(xht)
where
P(ρ(xh t) isin dρ) =ρ
2π〈η2〉expminus ρ2
2〈η2〉
a Rayleigh distribution
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 79 88
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 93: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/93.jpg)
Determining b
Mean growth rate of wave group energy
If δ = 1〈σ〉
Dmicro
Dt gt δ lowast asymp 14times10minus3 a breaking wave group
where the material derivative is taken following the wave group
Local wave energylowast micro(t) = η2(xmax t)k2(xmax t)
xmax(t) is maximum of crest point in the group
(k〈σ〉) wavenumber mean frequency
lowast see Song and Banner J Phys Oceanogr 2002
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 80 88
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 94: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/94.jpg)
Determining b
2D Example
(Loading breakmovie)
minus100 minus80 minus60 minus40 minus20 0 20 40 60 80 100minus100
minus80
minus60
minus40
minus20
0
20
40
60
80
100Trajectories of group maxima and breaking
y (m
)
x (m)
Space-time evolution of breakingevents shown in movie
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 81 88
![Page 95: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/95.jpg)
Determining b
Breaking events should be energetically consistent
Recall
partt b+ b middotnablab =1
Re∆b+ kbX (x)Y (y)Z (z)T (t)a
nabla middot b = 0
kb found by equating the energy E(Tw) from the break velocity field with thetotal energy change of the surface ρ0g∆η2
E(Tw) =ρ0c
TwΩb
intΩb[0Tw]
Adxds =055gρ0kb
k2
If ∆η2 is total change in η2(xmax middot) during the breaking event then Thus
kb = 182k2∆η2
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 82 88
Determining b
Breaking events should be Poisson process
0 02 04 06 08 1 12 140
05
1
15
2
25
3
kb
002 004 006 008 01 0120
10
20
30
40
50
60
70
k (m1)
Histograms of wave group meanwavenumber klowast and breaking strength
kb
0 500 1000 1500 2000 25000
500
1000
1500
2000
2500
r (m)
l(r)r
Poissonfit theoretical indicator function vs
computational l(r)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 83 88
Results
X (km)
Y(k
m)
0 10 20 30 40 500
5
10
15
20
25
30
35
40
45
50
002
003
004
005
006
007
008
Initial Current with top speed 01 msec
Current after 64 hours (01 ms contours) with waves and breaking due to 15 ms winds
(a)0 10 20 30 40 50
0
10
20
30
40
50
X (km)
Y(k
m)
(b)0 10 20 30 40 50
0
10
20
30
40
50
X (km)
Y(k
m)
(a) No Stokes drift (b) With Stokes ustokes = (0260) ms
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 84 88
Results
0 05 1 15 2 25 30
005
01
015
02
025
03
035
04
045
05
(o) Current Stokes and No Breaking(minus) Current Stokes and Breaking(o) Current No Stokes and No Breaking(minus) Current No Stokes and Breaking
(o) No Current Stokes and No Breaking(minus) No Current No Stokes and Breaking
time (hrs)
rate
(m
s)
AVERAGE DISPERSION OF PASSIVE TRACERS
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 85 88
Results
0 05 1 15 2 25 302
04
06
08
1
12
14
16
x 10minus3
(minus) No Current Stokes and Breaking
(minus) No Current No Stokes and Breaking
(minus) Current No Stokes and Breaking
(minus) Current Stokes and Breaking
(o) Current Stokes and No Breaking
TRACER DISPERSION
time (hrs)
dis
per
sio
n |nabla
θ| (
mminus
1)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 86 88
Closing
Closing Comments
Data assimilation with sparse data forces us to put emphasis on climatemodel development
The rdquodimensional-curserdquo is unavoidable In the long term convergentmethods allow one to know how much a computational gain is possibleat the expense of optimality
Closure methods and stochastic parameterizations can lead to robust andefficient models and hence dimension reductionit exploits deepknowledge of the system dynamics
Taking advantage of different degrees of uncertainty among the degreesof freedom can be a productive alternative to reduced representations ofbackground error fields
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 87 88
Further Information
Further Information
Juan M Restrepohttpwwwphysicsarizonaedusimrestrepo
Uncertainty Quantification Grouphttpwwwphysicsarizonaedusimtolwinski
UQGQG
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 88 88
- UQG Group
- Time Dependent Estimation
-
- LinearGaussian Estimation
-
- Nonlinear Non-Gaussian Estimation
-
- The Path Integral Monte Carlo (PIMC)
- Samplers
-
- The DKF
- Application Lagrangian Data Assimilation
-
- Computational Cost EKF vs Particle Filter vs DKF
- Computational Cost Clustered DKF
-
- New Projects
- State-Space Reduction
- Stochastic Parameterization
-
- ProjectionFiltering Strategy
- Model
-
- Determining b
- Results
- Closing
- Further Information
-
![Page 96: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/96.jpg)
Determining b
Breaking events should be Poisson process
0 02 04 06 08 1 12 140
05
1
15
2
25
3
kb
002 004 006 008 01 0120
10
20
30
40
50
60
70
k (m1)
Histograms of wave group meanwavenumber klowast and breaking strength
kb
0 500 1000 1500 2000 25000
500
1000
1500
2000
2500
r (m)
l(r)r
Poissonfit theoretical indicator function vs
computational l(r)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 83 88
Results
X (km)
Y(k
m)
0 10 20 30 40 500
5
10
15
20
25
30
35
40
45
50
002
003
004
005
006
007
008
Initial Current with top speed 01 msec
Current after 64 hours (01 ms contours) with waves and breaking due to 15 ms winds
(a)0 10 20 30 40 50
0
10
20
30
40
50
X (km)
Y(k
m)
(b)0 10 20 30 40 50
0
10
20
30
40
50
X (km)
Y(k
m)
(a) No Stokes drift (b) With Stokes ustokes = (0260) ms
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 84 88
Results
0 05 1 15 2 25 30
005
01
015
02
025
03
035
04
045
05
(o) Current Stokes and No Breaking(minus) Current Stokes and Breaking(o) Current No Stokes and No Breaking(minus) Current No Stokes and Breaking
(o) No Current Stokes and No Breaking(minus) No Current No Stokes and Breaking
time (hrs)
rate
(m
s)
AVERAGE DISPERSION OF PASSIVE TRACERS
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 85 88
Results
0 05 1 15 2 25 302
04
06
08
1
12
14
16
x 10minus3
(minus) No Current Stokes and Breaking
(minus) No Current No Stokes and Breaking
(minus) Current No Stokes and Breaking
(minus) Current Stokes and Breaking
(o) Current Stokes and No Breaking
TRACER DISPERSION
time (hrs)
dis
per
sio
n |nabla
θ| (
mminus
1)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 86 88
Closing
Closing Comments
Data assimilation with sparse data forces us to put emphasis on climatemodel development
The rdquodimensional-curserdquo is unavoidable In the long term convergentmethods allow one to know how much a computational gain is possibleat the expense of optimality
Closure methods and stochastic parameterizations can lead to robust andefficient models and hence dimension reductionit exploits deepknowledge of the system dynamics
Taking advantage of different degrees of uncertainty among the degreesof freedom can be a productive alternative to reduced representations ofbackground error fields
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 87 88
Further Information
Further Information
Juan M Restrepohttpwwwphysicsarizonaedusimrestrepo
Uncertainty Quantification Grouphttpwwwphysicsarizonaedusimtolwinski
UQGQG
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 88 88
- UQG Group
- Time Dependent Estimation
-
- LinearGaussian Estimation
-
- Nonlinear Non-Gaussian Estimation
-
- The Path Integral Monte Carlo (PIMC)
- Samplers
-
- The DKF
- Application Lagrangian Data Assimilation
-
- Computational Cost EKF vs Particle Filter vs DKF
- Computational Cost Clustered DKF
-
- New Projects
- State-Space Reduction
- Stochastic Parameterization
-
- ProjectionFiltering Strategy
- Model
-
- Determining b
- Results
- Closing
- Further Information
-
![Page 97: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/97.jpg)
Results
X (km)
Y(k
m)
0 10 20 30 40 500
5
10
15
20
25
30
35
40
45
50
002
003
004
005
006
007
008
Initial Current with top speed 01 msec
Current after 64 hours (01 ms contours) with waves and breaking due to 15 ms winds
(a)0 10 20 30 40 50
0
10
20
30
40
50
X (km)
Y(k
m)
(b)0 10 20 30 40 50
0
10
20
30
40
50
X (km)
Y(k
m)
(a) No Stokes drift (b) With Stokes ustokes = (0260) ms
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 84 88
Results
0 05 1 15 2 25 30
005
01
015
02
025
03
035
04
045
05
(o) Current Stokes and No Breaking(minus) Current Stokes and Breaking(o) Current No Stokes and No Breaking(minus) Current No Stokes and Breaking
(o) No Current Stokes and No Breaking(minus) No Current No Stokes and Breaking
time (hrs)
rate
(m
s)
AVERAGE DISPERSION OF PASSIVE TRACERS
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 85 88
Results
0 05 1 15 2 25 302
04
06
08
1
12
14
16
x 10minus3
(minus) No Current Stokes and Breaking
(minus) No Current No Stokes and Breaking
(minus) Current No Stokes and Breaking
(minus) Current Stokes and Breaking
(o) Current Stokes and No Breaking
TRACER DISPERSION
time (hrs)
dis
per
sio
n |nabla
θ| (
mminus
1)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 86 88
Closing
Closing Comments
Data assimilation with sparse data forces us to put emphasis on climatemodel development
The rdquodimensional-curserdquo is unavoidable In the long term convergentmethods allow one to know how much a computational gain is possibleat the expense of optimality
Closure methods and stochastic parameterizations can lead to robust andefficient models and hence dimension reductionit exploits deepknowledge of the system dynamics
Taking advantage of different degrees of uncertainty among the degreesof freedom can be a productive alternative to reduced representations ofbackground error fields
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 87 88
Further Information
Further Information
Juan M Restrepohttpwwwphysicsarizonaedusimrestrepo
Uncertainty Quantification Grouphttpwwwphysicsarizonaedusimtolwinski
UQGQG
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 88 88
- UQG Group
- Time Dependent Estimation
-
- LinearGaussian Estimation
-
- Nonlinear Non-Gaussian Estimation
-
- The Path Integral Monte Carlo (PIMC)
- Samplers
-
- The DKF
- Application Lagrangian Data Assimilation
-
- Computational Cost EKF vs Particle Filter vs DKF
- Computational Cost Clustered DKF
-
- New Projects
- State-Space Reduction
- Stochastic Parameterization
-
- ProjectionFiltering Strategy
- Model
-
- Determining b
- Results
- Closing
- Further Information
-
![Page 98: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/98.jpg)
Results
0 05 1 15 2 25 30
005
01
015
02
025
03
035
04
045
05
(o) Current Stokes and No Breaking(minus) Current Stokes and Breaking(o) Current No Stokes and No Breaking(minus) Current No Stokes and Breaking
(o) No Current Stokes and No Breaking(minus) No Current No Stokes and Breaking
time (hrs)
rate
(m
s)
AVERAGE DISPERSION OF PASSIVE TRACERS
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 85 88
Results
0 05 1 15 2 25 302
04
06
08
1
12
14
16
x 10minus3
(minus) No Current Stokes and Breaking
(minus) No Current No Stokes and Breaking
(minus) Current No Stokes and Breaking
(minus) Current Stokes and Breaking
(o) Current Stokes and No Breaking
TRACER DISPERSION
time (hrs)
dis
per
sio
n |nabla
θ| (
mminus
1)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 86 88
Closing
Closing Comments
Data assimilation with sparse data forces us to put emphasis on climatemodel development
The rdquodimensional-curserdquo is unavoidable In the long term convergentmethods allow one to know how much a computational gain is possibleat the expense of optimality
Closure methods and stochastic parameterizations can lead to robust andefficient models and hence dimension reductionit exploits deepknowledge of the system dynamics
Taking advantage of different degrees of uncertainty among the degreesof freedom can be a productive alternative to reduced representations ofbackground error fields
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 87 88
Further Information
Further Information
Juan M Restrepohttpwwwphysicsarizonaedusimrestrepo
Uncertainty Quantification Grouphttpwwwphysicsarizonaedusimtolwinski
UQGQG
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 88 88
- UQG Group
- Time Dependent Estimation
-
- LinearGaussian Estimation
-
- Nonlinear Non-Gaussian Estimation
-
- The Path Integral Monte Carlo (PIMC)
- Samplers
-
- The DKF
- Application Lagrangian Data Assimilation
-
- Computational Cost EKF vs Particle Filter vs DKF
- Computational Cost Clustered DKF
-
- New Projects
- State-Space Reduction
- Stochastic Parameterization
-
- ProjectionFiltering Strategy
- Model
-
- Determining b
- Results
- Closing
- Further Information
-
![Page 99: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/99.jpg)
Results
0 05 1 15 2 25 302
04
06
08
1
12
14
16
x 10minus3
(minus) No Current Stokes and Breaking
(minus) No Current No Stokes and Breaking
(minus) Current No Stokes and Breaking
(minus) Current Stokes and Breaking
(o) Current Stokes and No Breaking
TRACER DISPERSION
time (hrs)
dis
per
sio
n |nabla
θ| (
mminus
1)
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 86 88
Closing
Closing Comments
Data assimilation with sparse data forces us to put emphasis on climatemodel development
The rdquodimensional-curserdquo is unavoidable In the long term convergentmethods allow one to know how much a computational gain is possibleat the expense of optimality
Closure methods and stochastic parameterizations can lead to robust andefficient models and hence dimension reductionit exploits deepknowledge of the system dynamics
Taking advantage of different degrees of uncertainty among the degreesof freedom can be a productive alternative to reduced representations ofbackground error fields
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 87 88
Further Information
Further Information
Juan M Restrepohttpwwwphysicsarizonaedusimrestrepo
Uncertainty Quantification Grouphttpwwwphysicsarizonaedusimtolwinski
UQGQG
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 88 88
- UQG Group
- Time Dependent Estimation
-
- LinearGaussian Estimation
-
- Nonlinear Non-Gaussian Estimation
-
- The Path Integral Monte Carlo (PIMC)
- Samplers
-
- The DKF
- Application Lagrangian Data Assimilation
-
- Computational Cost EKF vs Particle Filter vs DKF
- Computational Cost Clustered DKF
-
- New Projects
- State-Space Reduction
- Stochastic Parameterization
-
- ProjectionFiltering Strategy
- Model
-
- Determining b
- Results
- Closing
- Further Information
-
![Page 100: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/100.jpg)
Closing
Closing Comments
Data assimilation with sparse data forces us to put emphasis on climatemodel development
The rdquodimensional-curserdquo is unavoidable In the long term convergentmethods allow one to know how much a computational gain is possibleat the expense of optimality
Closure methods and stochastic parameterizations can lead to robust andefficient models and hence dimension reductionit exploits deepknowledge of the system dynamics
Taking advantage of different degrees of uncertainty among the degreesof freedom can be a productive alternative to reduced representations ofbackground error fields
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 87 88
Further Information
Further Information
Juan M Restrepohttpwwwphysicsarizonaedusimrestrepo
Uncertainty Quantification Grouphttpwwwphysicsarizonaedusimtolwinski
UQGQG
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 88 88
- UQG Group
- Time Dependent Estimation
-
- LinearGaussian Estimation
-
- Nonlinear Non-Gaussian Estimation
-
- The Path Integral Monte Carlo (PIMC)
- Samplers
-
- The DKF
- Application Lagrangian Data Assimilation
-
- Computational Cost EKF vs Particle Filter vs DKF
- Computational Cost Clustered DKF
-
- New Projects
- State-Space Reduction
- Stochastic Parameterization
-
- ProjectionFiltering Strategy
- Model
-
- Determining b
- Results
- Closing
- Further Information
-
![Page 101: Climate, Assimilation of Data and Models - When Data Fail Us · 2010-06-29 · UQG Group Focus Problems Climate/Weather Hydrogeology Computer Vision Complex Problems.Problems in which](https://reader031.fdocuments.in/reader031/viewer/2022011817/5e7f87c39fefc97eed4f2d97/html5/thumbnails/101.jpg)
Further Information
Further Information
Juan M Restrepohttpwwwphysicsarizonaedusimrestrepo
Uncertainty Quantification Grouphttpwwwphysicsarizonaedusimtolwinski
UQGQG
JUAN M RESTREPO Group Leader Uncertainty Quantification Group ( University of Arizona )Climate Assimilation of Data and Models June 2010 88 88
- UQG Group
- Time Dependent Estimation
-
- LinearGaussian Estimation
-
- Nonlinear Non-Gaussian Estimation
-
- The Path Integral Monte Carlo (PIMC)
- Samplers
-
- The DKF
- Application Lagrangian Data Assimilation
-
- Computational Cost EKF vs Particle Filter vs DKF
- Computational Cost Clustered DKF
-
- New Projects
- State-Space Reduction
- Stochastic Parameterization
-
- ProjectionFiltering Strategy
- Model
-
- Determining b
- Results
- Closing
- Further Information
-