16 March 2020
POLITECNICO DI TORINORepository ISTITUZIONALE
Compressed sensing: basics and beyond (tutorial) / Fosson, Sophie; Magli, Enrico. - (2015). ((Intervento presentato alconvegno Fifteenth International Conference on Computer Aided Systems Theory (EUROCAST 2015) tenutosi a LasPalmas de Gran Canaria, Spain nel Feb 8-13, 2015.
Original
Compressed sensing: basics and beyond (tutorial)
Publisher:
PublishedDOI:
Terms of use:openAccess
Publisher copyright
(Article begins on next page)
This article is made available under terms and conditions as specified in the corresponding bibliographic description inthe repository
Availability:This version is available at: 11583/2624990 since: 2015-12-05T15:07:46Z
Compressed SensingBasics and Beyond
. EUROCAST 2015..
. Feb 12, 2015
S.M. Fosson E. Magli
www.crisp-erc.eu
Outline
..1 Mathematical problem
..2 Applications
..3 Recovery
..4 Distributed compressed sensing
Outline
..1 Mathematical problem
..2 Applications
..3 Recovery
..4 Distributed compressed sensing
Mathematical problem
Compressed sensing (compressed sampling, compressive sensing... CS)deals with.Underdetermined linear systems .....
.
Ax = yx ∈ Rn (unknown), y ∈ Rm (measurements), A ∈ Rm×n, m < n
Within the infinite set of solutions, CS looks for the sparsest one.... with sparsity assumptions...x is k-sparse, i.e., it has k non-zero components, where k ≪ n
S.M. Fosson COMPRESSED SENSING 4/33
Mathematical problem
S.M. Fosson COMPRESSED SENSING 5/33
Questions
Ax = y, x ∈ Rn(sparse), y ∈ Rm,m < n
..1 Is the problem well-posed (= is the solution unique)?
..2 Are there feasible algorithms to find the solution?
..3 Which applications motivate this study?
Answers..1 Yes, under some conditions..2 A number of recovery algorithms have been proposed..3 ▶ Sparsity is ubiquitous: many signals are sparse in some basis
(y = Aϕx where ϕ is the sparsifying basis, e.g., DCT, wavelets,Fourier... )
▶ Applications where data acquisition is difficult/expensive, andone aims to move the computational load to the receiver
S.M. Fosson COMPRESSED SENSING 6/33
Outline
..1 Mathematical problem
..2 Applications
..3 Recovery
..4 Distributed compressed sensing
Medical Imaging
Magnetic Resonance Imaging (MRI): acquisition is slow[Lustig (2012)]
→ sense the compressed information directly
S.M. Fosson COMPRESSED SENSING 8/33
Compression and sampling
Ax = y, x ∈ Rn(sparse), y ∈ Rm,m < n
• Sampling: Nyquist-Shannon sampling theorem states given asignal bandlimited in (B,B), to represent it over a timeinterval T, we need at least 2BT samples
• CS indicates a way to merge compression and sampling, andsample at a sub-Nyquist rate [Tropp et al. (2009)]
S.M. Fosson COMPRESSED SENSING 9/33
Compression and sampling
sensing
ADCcompression
?
S.M. Fosson COMPRESSED SENSING 10/33
Wideband spectrum sensing
Modulated wideband converter (MWC) [Mishali and Eldar (2010)]
• Sub-Nyquist sampling for signals sparse in the frequency domain• Realized in hardware (with commercial devices)
S.M. Fosson COMPRESSED SENSING 11/33
Single-pixel camera
Boufonos et al., ICASSP 2008Key ingredient: a microarray consisting of a large number of smallmirrors that can be turned on or off individuallyLight from the image is reflected on this microarray and a lenscombines all the reflected beams in one sensor, the single pixel ofthe camera
S.M. Fosson COMPRESSED SENSING 12/33
Outline
..1 Mathematical problem
..2 Applications
..3 Recovery
..4 Distributed compressed sensing
Mathematical formulation
.ℓ0-norm...∥x∥0 := number of nonzeros entries of x ∈ Rn
Natural formulation of the CS problem:.
.
P0 : minx∈Rn
∥x∥0 subject to Ax = y
• Is the solution unique?• P0 is NP-hard!
S.M. Fosson COMPRESSED SENSING 14/33
Uniqueness of the solution
.Spark..
.spark(A) := minimum number of columns of A that are linearlydependent.Theorem [D. Donoho, M. Elad (2003)]..
.For any vector y ∈ Rm, there exists at most one k-sparse signalx ∈ Rn such that y = Ax if and only if spark(A) > 2k.
S.M. Fosson COMPRESSED SENSING 15/33
Uniqueness of the solution
.Coherence..
.µ(A) := maxi ̸=j|AT
i Aj|∥Ai∥2∥Aj∥2
(Ai = ith column of A)
.Theorem [D. Donoho, M. Elad (2003)]..
.
Ifk <
12
(1 +
1µ(A)
)y ∈ Rm, there exists at most one k-sparse signal x ∈ Rn such thaty = Ax.
S.M. Fosson COMPRESSED SENSING 16/33
Basis Pursuit (BP)
Possible solution: convex relation.Basis Pursuit..
.
P1 : minx∈Rn
∥x∥1 subject to Ax = y
• P1 is convex; can be solved by linear programming• Are P0 and P1 equivalent?
S.M. Fosson COMPRESSED SENSING 17/33
Coherence
.Coherence..
.µ(A) := maxi ̸=j|AT
i Aj|∥Ai∥2∥Aj∥2
(Ai = ith column of A)
.Theorem [Elad and Bruckstein (2002)]..
.
If for the sparset solution x⋆ we have
∥x⋆∥0 <
√2 − 1
2µ(A)
then the solution of P1 is equal to the solution of P0.
S.M. Fosson COMPRESSED SENSING 18/33
Restricted Isometry Property (RIP)
.RIP..
.
Matrix A satisties the RIP of order k if there exists δk ∈ (0, 1) suchthat the following relation holds for any k-sparse x:
(1 − δk) ∥x∥22 ≤ ∥Ax∥2
2 ≤ (1 + δk) ∥x∥22
.Theorem [Candès (2008)]..
.If δk <
√2 − 1, then for all k-sparse x ∈ Rn such that Ax = y, the
solution of P1 is equal to the solution of P0.
S.M. Fosson COMPRESSED SENSING 19/33
Which matrices?
• Coherence, spark, RIP: not easy to compute• Random matrices A with i.i.d. entries drawn from continuous
distributions have spark(A) = m + 1 with probability one.• Gaussian, Bernoulli matrices: given δ ∈ (0, 1) there exist c1, c2
depending on δ such that G. and B. matrices satisfy the RIPwith constant δ and any m ≥ c1k log(n/k) with probability≥ 1 − 2e−c2m [Baraniuk (2008)]
• Structured matrices: circulant matrices, partial Fouriermatrices
S.M. Fosson COMPRESSED SENSING 20/33
Orthogonal Matching Pursuit (OMP)
• “When we talk about BP, we often say that the linearprogram can be solved in polynomial time with standardscientific software, and we cite books on convex programming[...]. This line of talk is misleading because it may take a longtime to solve the linear program, even for signals of moderatelength” [Tropp and Gilbert (2007)]
• Possible solution: greedy algorithm, fast, easy to implement→ OMP
S.M. Fosson COMPRESSED SENSING 21/33
Orthogonal Matching Pursuit (OMP)
..1 Initialize r0 = y, Λ0 = ∅
..2 For t = 1, . . . ,Tmax
..3 λt = argmaxj=1,...,n
|ATj rt−1|
..4 Λt = Λt−1 ∪ {λt}
..5 x̂t = argminx∈Rn
∥y − AΛtx∥2
..6 rt = y − AΛt x̂t
• Tmax ≈ k• OMP requires the knowledge of k!
S.M. Fosson COMPRESSED SENSING 22/33
Variants of BP
.Basis Pursuit Denoise (BPDN)..
.
P1 : minx∈Rn
∥x∥1 subject to ∥Ax = y∥2 ≤ ε
Unconstrained version of BPDN.Lasso...minx∈Rn
(∥Ax − y∥2
2 + λ ∥x∥1)
For some λ > 0, Lasso and BPDN have the same solution (thechoice of λ is tricky!)
S.M. Fosson COMPRESSED SENSING 23/33
Iterative soft thresholding (IST)
..1 x̂0 = 0
..2 For t = 1, . . . ,Tmax
..3 x̂t = Sλ(x̂t−1 + τAT(y − A ∗ x̂t−1))
where the operator Sλ is defined entry by entry asSλ(x) = sgn(|x| − λ) if |x| > λ, 0 otherwise
• IST achieves a minimum of the Lasso [Fornasier (2010)], andin many common situations such minimum is unique[Tibshirani (2012)]
• Faster method to get a minimum of the Lasso: alternatingdirection method of multipliers (ADMM)
S.M. Fosson COMPRESSED SENSING 24/33
Iterative hard thresholding
..1 x̂0 = 0
..2 For t = 1, . . . ,Tmax
..3 x̂t = Hk(x̂t−1 + AT(y − Ax̂t−1))
where the operator Hk(x) is the non-linear operator that sets allbut the largest (in magnitude) k elements of x to zero [Blumensath(2008)]
S.M. Fosson COMPRESSED SENSING 25/33
Outline
..1 Mathematical problem
..2 Applications
..3 Recovery
..4 Distributed compressed sensing
Distributed compressed sensing (DCS)
• Data acquisition is perfomed by a network of sensors
yv = Avxv v ∈ V = { sensors }
• First works: recovery is performed by a fusion center thatgathers information from the network (sensing matrices,measurements)
• New: in-network recovery, exploiting local communication andconsensus procedures
• We need iterative algorithms
S.M. Fosson COMPRESSED SENSING 27/33
Distributed Compressed Sensing (DCS)
S.M. Fosson COMPRESSED SENSING 28/33
Distributed Compressed Sensing (DCS)
S.M. Fosson COMPRESSED SENSING 29/33
Distributed Compressed Sensing (DCS)
S.M. Fosson COMPRESSED SENSING 30/33
Distributed Compressed Sensing (DCS)
S.M. Fosson COMPRESSED SENSING 31/33
DCS: in-network recovery
..1 C. Ravazzi, S.M. Fosson, E. Magli, E., Energy-saving GossipAlgorithm for Compressed Sensing in Multi-agent Systems,ICASSP, 2014
..2 S.M. Fosson, J. Matamoros, C. Antón-Haro , E. Magli,Distributed Support Detection of Jointly Sparse Signals,ICASSP, 2014
..3 J. Matamoros, S.M. Fosson, E. Magli, C. Antón-Haro,Distributed ADMM for in-network reconstruction of sparsesignals with innovations, IEEE GlobalSIP, 2014
..4 C. Ravazzi, S.M. Fosson, E. Magli, Distributed iterativethresholding for ℓ0/ℓ1-regularized linear inverse problems,IEEE Trans. Inf. Theory, 2015.
S.M. Fosson COMPRESSED SENSING 32/33
CS references
..1 http://dsp.rice.edu/cs
..2 E. Candès, J. Romberg, and T. Tao, Robust uncertaintyprinciples: Exact signal reconstruction from highly incompletefrequency information, IEEE Trans. Inf. Theory, Feb. 2006
..3 D. Donoho, Compressed sensing, IEEE Trans. Inf. Theory,Apr. 2006
..4 A mathematical Introduction to Compressive Sensing, editedby S. Foucart and H. Rauhut, 2013
..5 Compressed Sensing: Theory and Applications, edited by Y.C. Eldar and G. Kutyniok, 2012
..6 Theoretical Foundations and Numerical Methods for SparseRecovery, edited by M. Fornasier, 2010
S.M. Fosson COMPRESSED SENSING 33/33