Lecture 3: Dual 4D-Var · 2010. 6. 30. · Innovation vector × H Observation ... finite-amp...
Transcript of Lecture 3: Dual 4D-Var · 2010. 6. 30. · Innovation vector × H Observation ... finite-amp...
-
Lecture 3: Dual 4D-Var
-
Outline
• 4D-Var recap• Dual 4D-Var (4D-PSAS & R4D-Var)• The ROMS 4D-PSAS & R4D-Var algorithms• Weak constraint 4D-Var
-
,y R
Data Assimilation: Recap
bb(t), Bb
fb(t), Bf
xb(0), Bx
Model solutions depends on xb(0), fb(t), bb(t), h(t)
time
x(t)
Obs, yPriorPosterior
ROMS
Prior
-
Notation & Nomenclature: Recap
⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥=⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦
TS
x ζuv
(0)( )( )( )
ttt
⎡ ⎤⎢ ⎥⎢ ⎥=⎢ ⎥⎢ ⎥⎣ ⎦
xf
zbη
1
N
y
y
⎡ ⎤⎢ ⎥⎢ ⎥=⎢ ⎥⎢ ⎥⎣ ⎦
y ( )( )H= − bd y x
Statevector
Controlvector
Observationvector
Innovationvector
×
HObservation
operator
Prior
-
( (0), ( ), ( ), ( ))T T T Tt t tδ δ= b fz x ε ε ηinitial
conditionincrement
boundaryconditionincrement
forcingincrement
corrections for model
error
bb(t),Bb
fb(t), Bf
xb(0), Bx
( ) ( )1 11 12 2
TTJ δ δ δ δ− −= + − −z D z G z d R G z d
diag( , , , )= x b fD B B B Q
Prior (background) error covariance
TangentLinear Modelsampled atobs points
ObsErrorCov.
Innovation
= − bd y Hx
Prior
Incremental Formulation: Recap
(Courtier et al., 1994)
-
⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢
•••••••
⎥⎢ ⎥⎢•
⎥⎢ ⎥⎢ ⎥⎣
•
⎦•
••
⎡ ⎤⎢ ⎥⎣ ⎦
⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢
•••••••
⎥⎢ ⎥⎢•
⎥⎢ ⎥⎢ ⎥⎣
•
⎦•
zPrimalSpace
yObservation
vector
zDual
Space
Primal vs Dual Formulation: RecapVector of
increments
-
The Solution: Recap
= +a bz z Kd
T T -1( )= +K DG GDG R
Analysis:
Gain (dual form):
Gain (primal form):
( )T T− − − −= +1 1 1 1K D G R G G R
-
Two Spaces: Recap
T T -1( )= +K DG GDG RGain (dual):
Gain (primal):
( )T T− − − −= +1 1 1 1K D G R G G R
obs obsN N×
model modelN N×obs modelN N
-
Two Spaces: Recap
G maps from model spaceto observation space
GT maps from observation spaceto model space
-
Primal Formulation: Recap
= +a bz z KdAnalysis:Goal of 4D-Var is to identify:
11 T 1 T 1( )δ−− − −= = +z Kd D G R G G R d
Solve the equivalent linear system:1 T 1 T 1( )δ− − −+ =D G R G z G R d
by minimizing:1 T 1 T 1 11 1( )
2 2T T TJ δ δ δ− − − −= + − +z D G R G z z G R d d R d
( ) ( )1 11 12 2
TTδ δ δ δ− −= + − −z D z G z d R G z d
-
Dual Formulation
= +a bz z KdAnalysis:Goal of 4D-Var is to identify:
1T T( )δ−
= = +z Kd DG GDG R dSolve the equivalent linear system:
by minimizing:
T1( ) ( )2
T TI = + −w w GDG R w w d
T T( ) ; δ+ = =GDG R w d z DG w
then compute: Tδ =z DG w
There is no physicalsignificance attached to w
-
Contours of J
Conjugate Gradient (CG) Methods
-
Matrix-less Operations
TGDG δThere are no matrix multiplications!
Zonal shear flow
-
Matrix-less Operations
There are no matrix multiplications!
Adjoint Model
TGDG δ
Zonal shear flow
-
Matrix-less Operations
There are no matrix multiplications!
Adjoint Model
TGDG δ
Zonal shear flow
-
Matrix-less Operations
There are no matrix multiplications!
Adjoint Model
TGDG δ
Zonal shear flow
-
Matrix-less Operations
There are no matrix multiplications!
Covariance
TGDG δ
Zonal shear flow
-
Matrix-less Operations
There are no matrix multiplications!
Covariance
TGDG δ
Zonal shear flow
-
Matrix-less Operations
There are no matrix multiplications!
Tangent LinearModel
TGDG δ
Zonal shear flow
-
Matrix-less Operations
There are no matrix multiplications!
Tangent LinearModel
TGDG δ
Zonal shear flow
-
TGDG δ
A covarianceZonal shear flow
Matrix-less Operations
There are no matrix multiplications!
-
Matrix-less Operations
There are no matrix multiplications!
Tangent LinearModel
TGDG δ
Zonal shear flow
-
Physical-space Statistical Analysis System(PSAS) – Da Silva et al. (1995)
-
Dual 4D-PSAS Algorithm
(define W4DPSAS, w4dpsas_ocean.h)
TG w
T( )I∂ ∂ = + −w GDG R w d
Xb(t), d
Choose w
Run ADROMS
D
Run TLROMS
ConjugateGradient
Algorithm
w
NLROMS, zb
NLROMS, za Xa(t)
TDG w
obs space
ADROMS forced by w
TGDG w
Run ADROMS
D
TaG w
δ az
-
Dual 4D-PSAS Algorithm
(define W4DPSAS, w4dpsas_ocean.h)
Choose w
Run ADROMS
D
Run TLROMS
ConjugateGradient
Algorithm
NLROMS, zb
NLROMS, za
Run ADROMS
D
Outer-loopInner-loop
-
The method of representers (R4D-Var)Bennett (2002)
-
The Dual of State-Space
ROMS state-vector increments:
( )( )
( ) ( )( )( )
tt
t ttt
δδ
δ δδδ
⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥=⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦
TS
x ζuv
The set of all continuous, linear functionals of δx(t) iscalled the dual of δx
For example, ym=Gδz belongs to the dual of δx
-
Zonal shear flow
The Dual of State-SpaceConsider the assimilation window t=[0,T] for thezonal shear flow…
δ z
-
Zonal shear flow
The Dual of State-Space
×××
× Observations
Consider the assimilation window t=[0,T] for thezonal shear flow with 3 observations.
δ z
-
Tangent LinearModel
δG z
Zonal shear flow
The Dual of State-Space
×××
× Observations
Consider the assimilation window t=[0,T] for thezonal shear flow with 3 observations.
-
Tangent LinearModel
δG z
Zonal shear flow
The Dual of State-Space
×××
× Observations
-
Tangent LinearModel
δG z
Zonal shear flow
The Dual of State-Space
× Observations
-
Tangent LinearModel
δG z
Zonal shear flow
The Dual of State-Space
× Observations
-
Tangent LinearModel
δG z
Zonal shear flow
The Dual of State-Space
× Observations
-
Tangent LinearModel
δG z
Zonal shear flow
The Dual of State-Space
× Observations
ym=
-
Tangent LinearModel
δG z
Zonal shear flow
The Dual of State-Space
× Observations
ym=1my2my3my
-
Tangent LinearModel
δG z
Zonal shear flow
The Dual of State-Space
× Observations
d=1 1m oy y−2 2m oy y−3 3m oy y−
×××
Innovationvector
-
The Dual of State-Space
The innovation vector belongs to the dual of δx(t).
According to Riesz representation theorem:
Let1
2
(0)( )( )
(T)
tt
δδδ
δ
⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥=⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦
xx
u x
x
for t=[0,T]
T T T1 1 1 2 2 2 3 3 3; ; ;m o m o m oy y y y y y− = − = − =ρ u ρ u ρ u
where ρi are referred to as “representer functions.”
-
The Dual of State-Space
Let1
2
(0)( )( )
(T)
i
i
i i
i
tt
⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥=⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦
rr
ρ r
r
for t=[0,T]
and ( )(t) (t)i= rR
-
Representer Functions
δ
Zonal shear flow
Observationlocation
-
Representer Functions
TGDG δ
Zonal shear flow
Observationlocation
-
TGDG δ
A covariance
= A representer
Green’s Function
Zonal shear flow
Representer Functions
-
Zonal shear flow
3
1(t)= (t) (t)= (t)+ (t)i i
iw
=
+ ∑a b bx x r x wR
The analysis increments can be written as theweighted sum of the representers
( )(t) (t)i= rR
Representer Functions
-
Zonal shear flow
The analysis increments can be written as theweighted sum of the representers
3
1(t)= (t) (t)= (t)+ (t)i i
iw
=
+ ∑a b bx x r x wR( )(t) (t)i= rR
Representer Functions
1(T)r
2 (T)r
3 (T)r
-
= +a bz z KdAnalysis:Goal of 4D-Var is to identify:
1T T( )δ−
= = +z Kd DG GDG R dSolve the equivalent linear system:
by minimizing:
T1( ) ( )2
T TI = + −w w GDG R w w d
T T( ) ; (0)δ+ = = ≡GDG R w d z DG w wR
then compute:
The elements of w are theweighting coefs for the ri(t)
Indirect Representer Algorithm(Egbert et al, 1994)
T (0) ; (t)= (t)δ δ δ= ≡ =z DG w w x z wR M RTLROMS
-
Indirect R4D-Var algorithm for a
single outer-loop TG w
T( )I∂ ∂ = + −w GDG R w d
Xb(t)
Choose w
Run ADROMS
D
Run TLROMS
ConjugateGradient
Algorithm
w
NLROMS, zb
TLROMS, δza δXa(t)
TDG w
obs space
ADROMS forced by w
TGDG w
Run ADROMS
D
TaG w
δ az
-
Indirect R4D-Var Algorithm
(define W4DVAR, w4dvar_ocean.h)
TG w
T( )I∂ ∂ = + −w GDG R w d
Xb(t)
w
Xa(t)
TDG w
obs space
ADROMS forced by w
TGDG w
TaG w
δ az
Choose w
Run ADROMS
D
Run TLROMS
ConjugateGradient
Algorithm
RPROMS
RPROMS, za
Run ADROMS
D
NLROMS, zb
dRPROMS=finite-ampTLROMS
-
Choose w
Run ADROMS
D
Run TLROMS
ConjugateGradient
Algorithm
RPROMS
RPROMS, za
Run ADROMS
NLROMS, zb
D
Indirect R4D-Var Algorithm
(define W4DVAR, w4dvar_ocean.h)
Outer-loopInner-loop
-
Weak Constraint 4D-Var
1 1( ) ( , )( ( ), ( ), ( ))i i i i i it M t t t t t− −=x x f bNonlinear ROMS (NLROMS):
Nonlinear ROMS (NLROMS) with model error:
1 1( ) ( , )( ( ), ( ), ( ), ( ))i i i i i i it M t t t t t t− −=x x f b εModel error prior: 0Model error prior covariance:
4D-Var control vector:
(0)( )( )( )
ttt
⎡ ⎤⎢ ⎥⎢ ⎥=⎢ ⎥⎢ ⎥⎣ ⎦
xf
zbη Correction for modelerror
Q(no explicit timecorrelation in Q, but thereis some in practice)
-
Weak Constraint 4D-Var
1 1( ) ( , ) ( )i i i it t t tδ δ− −=x M uTangent linear ROMS (TLROMS):
( )( )
( )( )( )
i
ii
i
i
tt
ttt
δδ
δδδ
⎡ ⎤⎢ ⎥⎢ ⎥=⎢ ⎥⎢ ⎥⎣ ⎦
xf
ubη 4D forcing for TLROMS
( )itδ =η 0Strong constraint:
-
Two Spaces
T T -1( )= +K DG GDG RGain (dual):
Gain (primal):
( )T T− − − −= +1 1 1 1K D G R G G R
obs obsN N×
model modelN N×obs modelN N
-
Two Spaces
( )model x times f b xN N N N N N= + + +Weak constraint:
( )model x times f bN N N N N= + +Strong constraint:
Weak constraint is only practical in dual formulationof 4D-Var since Nobs is unaffected:
T T -1( )= +K DG GDG R
obs obsN N×
-
Mechanics of Dual 4D-Var: Preconditioning
= +a bz z KdAnalysis:Goal of 4D-Var is to identify:
1T T( )δ−
= = +z Kd DG GDG R dSolve the equivalent linear system:
by minimizing:
T1( ) ( )2
T TI = + −w w GDG R w w d
T T( ) ; δ+ = =GDG R w d z DG w
Preconditioning via the change of variable 1 2−=v R w
-
Mechanics of Dual 4D-Var: Lanczos vectors
Lanczos formulation of conjugate gradient algorithmin observation space is used (congrad.F).
1T T( )−
= +K DG GDG R
Dual formulation of gain matrix:
Dual formulation of practical gain matrix:
T 1 2 1 T 1 2k k k k
− − −=K DG R V T V R
Many practical diagnostic applications using thisformulation (Lectures 4 & 5).
-
An Example: ROMS CCS
30km, 10 km & 3 km grids, 30- 42 levelsVeneziani et al (2009)Broquet et al (2009)Moore et al (2010)
COAMPSforcing
ECCO openboundaryconditions
fb(t), Bf
bb(t), Bb
xb(0), Bx
Previous assimilationcycle
-
Observations (y)
CalCOFI &GLOBEC
SST &SSH
Argo
TOPP Elephant SealsIngleby andHuddleston (2007)
Data from Dan Costa
-
4D-Var Configuration
• Case studies for a representative case3-10 March, 2003.
• 1 outer-loop, 100 inner-loops• 7 day assimilation window• Prior D: x Lh=50 km, Lv=30m, σ from clim
f Lτ=300km, LQ=100km, σ from COAMPS b Lh=100 km, Lv=30m, σ from clim
• Super observations formed• Obs error R (diagonal):
SSH 2 cmSST 0.4 Chydrographic 0.1 C, 0.01psu
-
Primal, strongDual, strongDual, weakJmin
4D-Var Performance
3-10 March, 2003(10km, 42 levels)
-
Issues, Things to do, & Coming Soon
• Slow convergence of dual 4D-Var compared to primalformulation:
- w has no physical significance, so need not be physically realizable
- minimum residual method may be the answer(El Akkraoui and Gauthier, 2010)
Tδ =z DG w
-
Summary• Strong and weak constraint 4D-Var, dual
formulation: define W4DPSASDrivers/w4dpsas_ocean.hdefine W4DVARDrivers/w4dvar_ocean.h
• Matrix-less iterations to identify cost functionminimum using TLROMS and ADROMS
-
References• Bennett, A.F., 2002: Inverse Modeling of the Ocean and
Atmosphere. Cambridge University Press, 234pp.• Courtier, P., 1997: Dual formulation of four-dimensional variational
assimilation. Q. J. R. Meteorol. Soc.,123, 2449-2461.• Da Silva, A., J. Pfaendtner, J. Guo, M. Sienkiewicz and S. Cohn,
1995: Assessing the effects of data selection with DAO's physical-space statistical analysis system. Proceedings of the second international WMO symposium on assimilation of observations in meteorology and oceanography, Tokyo 13-17 March, 1995. WMO.TD 651, 273-278.
• Di Lorenzo, E., A.M. Moore, H.G. Arango, B.D. Cornuelle, A.J. Miller, B. Powell, B.S. Chua and A.F. Bennett, 2007: Weak and strong constraint data assimilation in the inverse Regional Ocean Modeling System (ROMS): development and application for baroclinic coastal upwelling system. Ocean Modeling, 16, 160-187.
• Egbert, G.D., A.F. Bennett and M.C.G. Foreman, 1994: TOPEX/POSEIDON tides estimated using a global inverse method. J. Geophys. Res., 99, 24,821-24,852.
• El Akkraoui, A. and P. Gauthier, 2010: Convergence properties of the primal and dual forms of variational data assimilation. Q. J. R. Meteorol. Soc., 136, 107-115.
-
References• Moore, A.M., H.G. Arango, G. Broquet, B.S. Powell, J. Zavala-Garay
and A.T. Weaver, 2010: The Regional Ocean Modeling System (ROMS) 4-dimensional data assimilation systems: Part I. Ocean Modelling, Submitted.