Statistical Issues in Searches: Photon Science...
Transcript of Statistical Issues in Searches: Photon Science...
![Page 1: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/1.jpg)
Statistical Issues in Searches:Photon Science Response
Rebecca Willett, Duke University
1 / 34
![Page 2: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/2.jpg)
Photon science seen this week
Applications
tomography
ptychography
nanochrystalography
coherent diffraction imaging
Challenges
high dimensional signals
underdetermined problems
limited number of photons
complex forward models
2 / 34
![Page 3: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/3.jpg)
Generic problem statement
f∗ is the signal of interest
Before hitting our detector,the signal is transformed
f∗ 7→ A(f∗)
(tomographic projections,diffraction patterns, codedaperture)
We observe
y ∼ Poisson(A(f∗))
Goal is to infer f∗ from y
3 / 34
![Page 4: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/4.jpg)
Key questions
We infer f∗ via
f = arg minf∈F
− log p(y|A(f)) + pen(f)
where
− log p(y|A(f)) is the negative Poisson log likelihood
pen(f) is a penalty / regularizer / negative log prior
F is a class of candidate estimates
We’d like to understand the following:
How does performance depend on A?
What should we use for pen(f)?
How do we solve the optimization problem?
4 / 34
![Page 5: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/5.jpg)
First a case study
Linear operatorConvex objective
Nonlinear operatorConvex objective
Nonlinear operatorNonconvex objective
Linear operatorNonconvex objective
For now, consider the case where
A(f) = Af is a (possibly underdetermined) linear operator
pen(f) is convex and measures sparsity
5 / 34
![Page 6: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/6.jpg)
An oracle inequality1
Theorem
If y ∼ Poisson(f∗), where ‖f∗‖1 = I, and
f , arg minf∈F
{− log p(y|f) + `(f)}
where `(f) , prefix codelength(f), then (omitting constants)
EH2
(f
I,f∗
I
)︸ ︷︷ ︸
risk
� minf∈F
∥∥∥∥fI − f∗
I
∥∥∥∥22︸ ︷︷ ︸
approximation error
+`(f)
I︸︷︷︸estimation error
(Similar oracle inequalities exist without the codelength perspective.)
Codelengths correspond to generalized notions of sparsity.1Li & Barron 2000, Massart 2003, Willett & Nowak 2005, Baraud & Birge 2006-2011, Bunea Tsybakov &
Wegkamp 2007
6 / 34
![Page 7: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/7.jpg)
A myriad of tradeoffs
We want to choose F and `(·) to make this bound as small aspossible for large families of possible f∗
Choosing a small model class F gives short codelengths butlarge approximation errors (bias)
Choosing a rich model class makes the choice of codingscheme challenging
we want `(·) convex so we can perform search quicklywe want shorter codes for estimators which conform better toprior information
7 / 34
![Page 8: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/8.jpg)
Codes and sparsity
Encode which pixels contain stars and the brightness of those stars
8 / 34
![Page 9: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/9.jpg)
Codes and sparsity
Encode which wavelet coefficients are non-zero and theirmagnitudes
8 / 34
![Page 10: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/10.jpg)
Codes and sparsity
Encode image patches as linear combination of a small number ofrepresentative patches – encode which representatives used and
their weights
8 / 34
![Page 11: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/11.jpg)
Codes and sparsity
Encode location of object along a low-dimensional manifold2
2Fu et al. 2007
8 / 34
![Page 12: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/12.jpg)
An oracle inequality
Theorem
If y ∼ Poisson(f∗), where ‖f∗‖1 = I, and
f , arg minf∈F
{− log p(y|f) + `(f)}
where `(f) , prefix codelength(f), then (omitting constants)
EH2
(f
I,f∗
I
)︸ ︷︷ ︸
risk
� minf∈F
∥∥∥∥fI − f∗
I
∥∥∥∥22︸ ︷︷ ︸
approximation error
+`(f)
I︸︷︷︸estimation error
(Similar oracle inequalities exist without the codelength perspective.)
This framework leads to guarantees on estimation performance forall f∗ with sparse approximations.
9 / 34
![Page 13: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/13.jpg)
Scheffe’s identityWe focus on the squared Hellinger distance:
H2(f1, f2) ,∫ (√
f1(x)−√f2(x)
)2dx
The squared Hellinger distance helps bound the L1 error:
H2(f1, f2) ≤ ‖f1 − f2‖1 ≤ 2H(f1, f2)
which is important due to
Sheffe’s identity
supB∈B
∣∣∣∣∫Bf1 −
∫Bf2
∣∣∣∣ =1
2‖f1 − f2‖1
The L1 distance gives a bound on the absolute different ofprobability measures of any Borel-measurable subset of thedensity’s domain.
10 / 34
![Page 14: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/14.jpg)
Poisson inverse problemsAj,i corresponds to the probability of a photon originating at location i
hitting a detector at location j.
Hardware systems can only aggregate events / photons, not measuredifferences:
Aj,i ≥ 0
The total number of observed events / photons cannot exceed the totalnumber of events:
n∑j=1
(Af)j ≤m∑i=1
fi
11 / 34
![Page 15: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/15.jpg)
Oracle inequality revisited
Theorem
Assume
c‖f1 − f2‖22 ≤ ‖Af1 −Af2‖22 ≤ C‖f1 − f2‖22
for all f1, f2 ∈ F⋃f∗. If y ∼ Poisson(Af∗) and
f , arg minf∈F
{− log p(y|Af) + `(f)}
where `(f) , prefix codelength(f), then (omitting constants)
E
∥∥∥∥∥f∗I − f
I
∥∥∥∥∥2
2︸ ︷︷ ︸risk
� minf∈F
C
c
∥∥∥∥fI − f∗
I
∥∥∥∥22︸ ︷︷ ︸
approximation error
+`(f)
cI︸︷︷︸estimation error
Similar results hold for `1 norms.
12 / 34
![Page 16: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/16.jpg)
Big questions
1. If we have complete control over A, can we design ourmeasurement system to get accurate high-resolution estimatesfrom relatively few photodetectors? (compressed sensing)
2. When we have limited control over A, what does the theorytell us about various design tradeoffs?
13 / 34
![Page 17: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/17.jpg)
Random compressed sensing matrices
Let A be an n×m Bernoulli ensemble matrix, so each Ai,j is −1√n
or 1√n
with equal probability. A then satisfies the “restricted
isometry property” (RIP, so C ≈ c ≈ 1) with high probability for nsufficiently large.
We would like to observe Af∗.
14 / 34
![Page 18: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/18.jpg)
Incorporating physical constraints
Elements of A must benonnegative =⇒ we shift A so allelements are nonnegative
Total measured intensity ‖Af∗‖1must not exceed the totalincident intensity ‖f∗‖1 =⇒ wemust scale A to ensure fluxpreservation.
15 / 34
![Page 19: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/19.jpg)
Shifted and scaled measurements
“Original” sensing matrix
A ∈{−1√n,
1√n
}n×mOur shifted and scaled version is
A =1
2√n
(A+
1√n
)So we measure
Af∗ =Af∗
2√n
+I
2n,
where Af∗ is the ideal signal and I , ‖f∗‖1 is the total signalintensity
16 / 34
![Page 20: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/20.jpg)
Compressible signalsAssume f∗ is compressible in some basis Φ; that is,
f∗ = Φθ∗ and‖θ∗ − θ(k)‖2
I= O(k−α)
where θ(k) is the best k-term approximation to θ∗.
Theorem
There exist a finite set of candidate estimators FI and c ∈ (0, I/n) withthe property
Af � c ∀f ∈ FIand a prefix code
`(f) ∝ ‖ΦT f‖0 log(m)
such that, for n sufficiently large
E
∥∥∥∥∥ fI − f∗
I
∥∥∥∥∥2
2
= O
[n
(log(m)
I
) 2α2α+1
+log(m/n)
n
].
(Recall m is length of f , n is number of detectors, and I is the totalintensity.)
17 / 34
![Page 21: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/21.jpg)
Simulation results, m = 105
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2x 104
0
0.2
0.4
0.6
0.8
1
1.2
Number of Measurements
Rel
ativ
e
Erro
r
Relative Error vs Measurements (10 Trial Average)
I = 100,000, k = 100I = 100,000, k = 500I = 100,000, k = 1,000 I = 10,000, k = 1,000
I = 10,000, k = 500I = 10,000, k = 100
18 / 34
![Page 22: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/22.jpg)
Exercising limited control over A
These bounds can also provide insight when we have a fixedsignal intensity or photon budget, and want to choose themeasurement system A to optimize the tradeoff between
measurement diversity and photon scarcity.
19 / 34
![Page 23: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/23.jpg)
Measurement diversity and photon scarcity
In limited- or sparse-angle tomography, how many and whichdifferent angles should be used?3
3Cederlund et al. 2009
20 / 34
![Page 24: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/24.jpg)
Measurement diversity and photon scarcity
100 200 300 400 500200
300
400
500
600
700
800
900
1000
Representative Observation, 2 Excitation Frequencies
Inte
nsity
Spectral band index100 200 300 400 500
40
60
80
100
120
140
160
180
200
220Representative Observation, 10 Excitation Frequencies
Inte
nsity
Spectral band index100 200 300 400 500
10
20
30
40
50
60
Representative Observation, 40 Excitation Frequencies
Inte
nsity
Spectral band index
100 200 300 400 5000
50
100
150
200
250
300
350
400
450Raman Spectrum, 2 Excitation Frequencies
Spectral band index
Inte
nsity
True Ramen SpectrumEstimate of Ramen Spectrum
100 200 300 400 5000
10
20
30
40
50
60
70
80
Raman Spectrum, 10 Excitation Frequencies
Spectral band index
Inte
nsity
True Ramen SpectrumEstimate of Ramen Spectrum
100 200 300 400 5000
5
10
15
20
Raman Spectrum, 40 Excitation Frequencies
Spectral band index
Inte
nsity
True Ramen SpectrumEstimate of Ramen Spectrum
In Shifted Excitation Raman Spectroscopy, how many and whichdifferent excitation frequencies should be used to separate Ramanspectrum from fluorescent background?
20 / 34
![Page 25: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/25.jpg)
Measurement diversity and photon scarcity
In coded aperture or structured illumination imaging, which maskpatterns best facilitate high-resolution image reconstruction?
20 / 34
![Page 26: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/26.jpg)
Linear operatorConvex objective
Nonlinear operatorConvex objective
Nonlinear operatorNonconvex objective
Linear operatorNonconvex objective
We know a lot about linear systems and convex objectives. A goodstrategy for the other regimes is to find a problem relaxation suchthat
(a) the relaxed problem is linear and convex and
(b) the solution to the relaxed problem closely corresponds to thesolution to the original problem
21 / 34
![Page 27: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/27.jpg)
The following slides are derived from slides made by ThomasStrohmer, UC Davis
22 / 34
![Page 28: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/28.jpg)
Phase retrieval
X-ray diffraction of DNA(Photo 51 by RosalindFranklin)
X-ray crystallography
Optical alignment(e.g., James Web SpaceTelescope)
23 / 34
![Page 29: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/29.jpg)
At the core of phase retrieval lies the problem:
We want to recover a function f(t) from intensity measurementsof its Fourier transform, |f(ω)|2.
Without further information about f , the phase retrievalproblem is ill-posed. We can either impose additionalproperties of f or take more measurements (or both)
We want an efficient phase retrieval algorithm based on arigorous mathematical framework, for which we canguarantee exact recovery, and which is also provably stable inpresence of noise.
Want flexible framework that does not require any priorinformation about the function (signal, image,...), yet canincorporate additional information if available.
24 / 34
![Page 30: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/30.jpg)
At the core of phase retrieval lies the problem:
We want to recover a function f(t) from intensity measurementsof its Fourier transform, |f(ω)|2.
Without further information about f , the phase retrievalproblem is ill-posed. We can either impose additionalproperties of f or take more measurements (or both)
We want an efficient phase retrieval algorithm based on arigorous mathematical framework, for which we canguarantee exact recovery, and which is also provably stable inpresence of noise.
Want flexible framework that does not require any priorinformation about the function (signal, image,...), yet canincorporate additional information if available.
24 / 34
![Page 31: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/31.jpg)
General phase retrieval problem
Suppose we have f∗ ∈ Cm or Cm1×m2 about which we havequadratic measurements of the form
y = A(f∗) = {|〈ak, f∗〉|2 : k = 1, 2, . . . ,m}.
Phase retrieval:
find fobeying A(f) = A(f∗) := y.
Goals:
Find measurement vectors {ak}k∈I such that f∗ is uniquelydetermined by {|〈ak, f∗〉|}k∈I .
Find an algorithm that reconstructs f∗ from {|〈ak, f∗〉|}k∈I .
25 / 34
![Page 32: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/32.jpg)
Standard approach: Oversampling
Sufficient oversampling of the diffraction pattern in theFourier domain gives uniqueness for generic two- orhigher-dim. signals in combination with support constraints
Alternating projections [Gerchberg-Saxton, Fienup]Alternatingly enforce object constraint in spatial domain andmeasurement constraint in Fourier domain
Seems to work in certain cases but has many limitations andproblems.
No proof of convergence to actual solution!
26 / 34
![Page 33: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/33.jpg)
Lifting
Following [Balan, Bodman, Casazza, Edidin, 2007], we willinterpret quadratic measurements of f as linear measurements ofthe rank-one matrix F := ffH :
|〈ak, f〉|2 = Tr(fHakaHk f) = Tr(AkF )
where Ak is the rank-one matrix akaHk . Define linear operator A
that maps pos.sem.def. matrix F into {Tr(AkF )}nk=1.
Now, the phase retrieval problem is equivalent to
find Fsubject to A(F ) = y
F � 0rank(F ) = 1
(RANKMIN)
Having found F , we factorize F as ffH to obtain the phaseretrieval solution (up to global phase factor).
27 / 34
![Page 34: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/34.jpg)
Lifting
Following [Balan, Bodman, Casazza, Edidin, 2007], we willinterpret quadratic measurements of f as linear measurements ofthe rank-one matrix F := ffH :
|〈ak, f〉|2 = Tr(fHakaHk f) = Tr(AkF )
where Ak is the rank-one matrix akaHk . Define linear operator A
that maps pos.sem.def. matrix F into {Tr(AkF )}nk=1.Now, the phase retrieval problem is equivalent to
find Fsubject to A(F ) = y
F � 0rank(F ) = 1
(RANKMIN)
Having found F , we factorize F as ffH to obtain the phaseretrieval solution (up to global phase factor).
27 / 34
![Page 35: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/35.jpg)
Phase retrieval as convex problem?
We need to solve:
minimize rank(F )subject to A(F ) = y
F � 0.(RANKMIN)
Note that A(F ) is highly underdetermined, thus cannot just invertA to get F .Rank minimization problems are typically NP-hard.
Use trace norm as convex surrogate for the rank functional [Beck’98, Mesbahi ’97], giving the semidefinite program:
minimize trace(F )subject to A(F ) = y
F � 0.(TRACEMIN)
28 / 34
![Page 36: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/36.jpg)
Phase retrieval as convex problem?
We need to solve:
minimize rank(F )subject to A(F ) = y
F � 0.(RANKMIN)
Note that A(F ) is highly underdetermined, thus cannot just invertA to get F .Rank minimization problems are typically NP-hard.Use trace norm as convex surrogate for the rank functional [Beck’98, Mesbahi ’97], giving the semidefinite program:
minimize trace(F )subject to A(F ) = y
F � 0.(TRACEMIN)
28 / 34
![Page 37: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/37.jpg)
A new methodology for phase retrieval
Lift up the problem of recovering a vector from quadraticconstraints into that of recovering a rank-one matrix from affineconstraints, and relax the combinatorial problem into a convenientconvex program.
PhaseLift
But when (if ever) is the trace minimization problemequivalent to the rank minimization problem?
29 / 34
![Page 38: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/38.jpg)
A new methodology for phase retrieval
Lift up the problem of recovering a vector from quadraticconstraints into that of recovering a rank-one matrix from affineconstraints, and relax the combinatorial problem into a convenientconvex program.
PhaseLift
But when (if ever) is the trace minimization problemequivalent to the rank minimization problem?
29 / 34
![Page 39: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/39.jpg)
A new methodology for phase retrieval
Lift up the problem of recovering a vector from quadraticconstraints into that of recovering a rank-one matrix from affineconstraints, and relax the combinatorial problem into a convenientconvex program.
PhaseLift
But when (if ever) is the trace minimization problemequivalent to the rank minimization problem?
29 / 34
![Page 40: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/40.jpg)
When is phase retrieval a convex problem?
Theorem: [Candes-Strohmer-Voroninski ’11]
Let f∗ in Rm or Cm and suppose we choose the measurementvectors {ak}nk=1 independently and uniformly at random on theunit sphere of Cm or Rm. If n ≥ cm logm, where c is a constant,then PhaseLift recovers f∗ exactly from {〈ak, f∗〉|2}nk=1 withprobability at least 1− 3e−γ
nm , where γ is an absolute constant.
Note that the “oversampling factor” logm is rather minimal!
This is the first result of its kind: phase retrieval can provably beaccomplished via convex optimization.
30 / 34
![Page 41: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/41.jpg)
Stability in presence of noise
Assume we observe
yi = |〈f, ai〉|2 + νi,
where νi is a noise term with ‖ν‖2 ≤ ε.Consider the solution to
minimize trace(F )subject to ‖A(F )− y‖2 ≤ ε
f � 0.
Theorem: [Candes-Strohmer-Voroninski ’11]
Under the same assumptions as in the other theorem, the solutionto the noisy, the solution f computed via PhaseLift obeys
‖f − f∗‖2 ≤ C0min (‖f∗‖2, ε/‖f∗‖2)
31 / 34
![Page 42: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/42.jpg)
Noise-aware framework
Suppose we observe
y ∼ Poisson(A(f∗))
Our convex formulation suggests to solve
minimize − log p(y|A(F )) + λ trace(F )subject to F � 0.
We can easily include additional constraints frequently used inphase retrieval, such as support constraints or positivity.
32 / 34
![Page 43: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/43.jpg)
Limitations of our theory
Theorems are not yet completely practical, since most phaseretrieval problems involve diffraction, i.e., Fourier transforms, andnot unstructured random measurements.
33 / 34
![Page 44: Statistical Issues in Searches: Photon Science Responseweb.ipac.caltech.edu/.../WillettResponse.pdf · Statistical Issues in Searches: Photon Science Response Rebecca Willett, Duke](https://reader033.fdocuments.in/reader033/viewer/2022042802/5f3aed59af8a11477e26a014/html5/thumbnails/44.jpg)
Conclusions
Oracle inequalities provide valuable insight into tradeoffsrelated to
sensing system design and measurement diversity,photon limitations,generalized sparsity, andoptimization methods
Nonlinear, nonconvex problems notoriously difficult to solve(hard to find global optimum)
Convex relaxations may help alleviate this challenge
34 / 34