Download presentation
Presentation is loading. Please wait.
1
CIS 700: βalgorithms for Big Dataβ
Lecture 9: Compressed Sensing Slides at Grigory Yaroslavtsev
2
Compressed Sensing Given a sparse signal π₯β β π can we recover it from a small number of measurements? Goal: design π΄β β πΓπ which allows to recover any π -sparse π₯β β π from π΄π₯. π΄ = matrix of i.i.d. Gaussians π(0,1) Application: signals are usually sparse in some Fourier domain
3
Reconstruction Reconstruction: min π₯ 0 , subject to: π΄π₯=π
Uniqueness: If there are two π -sparse solutions π₯ 1 , π₯ 2 : π΄ π₯ 1 β π₯ 2 =0 then π΄ has 2π linearly dependent columns If π=Ξ©( π 2 ) and π΄ is Gaussian then unlikely to have linearly dependent columns π₯ 0 not convex, NP-hard to reconstruct π₯ 0 β π₯ 1 : min π₯ 1 , subject to: π΄π₯=π When does this give sparse solutions?
4
Subgradient min π₯ 1 , subject to: π΄π₯=π
π₯ 1 is convex but not differentiable Subgradient π»f: equal to gradient where π is differentiable any linear lower bound where π is not differentiable β π₯ 0 ,Ξπ₯: π π₯ 0 +Ξπ₯ β₯π π₯ π»π π Ξπ₯ Subgradient for π₯ 1 : π» π₯ π =π πππ( π₯ π ) if π₯ π β 0 π» π₯ π β β1,1 if π₯ π =0 For all Ξπ₯ such that π΄Ξπ₯=0 satisfies π» π Ξπ₯β₯0 Sufficient: βπ€ such that π»= π΄ π π€ so π» π Ξπ₯= π€π΄Ξπ₯=0
5
Exact Reconstruction Property
Subgradient Thm. If π΄ π₯ 0 =π and there exists a subgradient π» for π₯ 1 such that π»= π΄ π π€ and columns of π΄ corresponding to π₯ 0 are linearly independent then π₯ 0 minimizes π₯ 1 and is unique. (Minimum): Assume π΄π¦=π. Will show π¦ 1 β₯ π₯ π§=π¦β π₯ 0 βπ΄π§=π΄π¦βπ΄ π₯ 0 =0 π» π π§=0β π¦ 1 = π₯ 0 +π§ β₯ π₯ π» π π§= π₯
6
Exact Reconstruction Property
(Uniqueness): assume π₯ 0 is another minimum π» at π₯ 0 is also a subgradient at π₯ 0 βΞπ₯:π΄Ξπ₯=0: π₯ 0 +Ξπ₯ 1 = π₯ 0 + π₯ 0 β π₯ 0 +Ξπ₯ β₯ π₯ π» π ( π₯ 0 β π₯ 0 +Ξπ₯) = π₯ π» π π₯ 0 β π₯ 0 + π» T Ξπ₯ π» π π₯ 0 β π₯ 0 = π€ π π΄ π₯ 0 β π₯ 0 = π€ π πβπ =0 π₯ 0 +Ξπ₯ 1 β₯ π₯ π» T Ξπ₯ π» i =sign (x 0 i )=sign ( π₯ 0 i ) if either is non-zero, otherwise equal to 0 β π₯ 0 and π₯ 0 have same sparsity pattern By linear independence of columns of π΄: π₯ 0 = π₯ 0
7
Restricted Isometry Property
Matrix π΄ satisfies restricted isometry property (RIP), if for any π -sparse π₯ there exists πΏ π : 1 β πΏ π π₯ β€ π΄π₯ β€ 1+ πΏ π π₯ 2 2 Exact isometry: all eigenvalues are Β±1 for orthogonal π₯,π¦: π₯ π π΄ π π΄π¦=0 Let π΄ π be the set of columns of π΄ in set π Lem: If π΄ satisfies RIP and πΏ π 1 + π 2 β€ πΏ π 1 + πΏ π 2 : For π of size π singular values of π΄ π in [1β πΏ π ,1+ πΏ π ] For any orthogonal π₯,π¦ with supports of size π 1 , π 2 : π₯ π π΄ π π΄π¦ β€ π₯ | π¦ |( πΏ π 1 + πΏ π 2 )
8
Restricted Isometry Property
Lem: If π΄ satisfies RIP and πΏ π 1 + π 2 β€ πΏ π 1 + πΏ π 2 : For π of size π singular values of π΄ π in [1β πΏ π ,1+ πΏ π ] For any orthogonal π₯,π¦ with supports of size π 1 , π 2 : π₯ π π΄ π π΄π¦ β€3/2 π₯ π¦ ( πΏ π 1 + πΏ π 2 ) W.l.o.g π₯ = π¦ =1 so π₯+π¦ 2 =2 2 1β πΏ π 1 + π 2 β€ π΄ π₯+π¦ 2 β€ 2 1+ πΏ π 1 + π 2 2 1β πΏ π 1 + πΏ π 2 β€ π΄ π₯+π¦ 2 β€ πΏ π 1 + πΏ π 2 1 β πΏ π 1 β€ π΄π₯ 2 β€(1+ πΏ π 1 ) 1 β πΏ π 2 β€ π΄π¦ 2 β€(1+ πΏ π 2 )
9
Restricted Isometry Property
2 π₯ π π΄ π π΄π¦ = π₯+π¦ π π΄ π π΄ π₯+π¦ β π₯ π π΄ π π΄π₯β π¦ π π΄ π π΄π¦ = π΄ π₯+π¦ 2 β π΄π₯ 2 β π΄π¦ 2 2 π₯ π π΄ π π΄π¦β€2 1+ πΏ π 1 + πΏ π 2 β 1 β πΏ π 1 β 1 β πΏ π 2 =3( πΏ π 1 + πΏ π 2 ) π₯ π π΄ π π΄π¦β€ π₯ β
π¦ β
( πΏ π 1 + πΏ π 2 )
10
Reconstruction from RIP
Thm. If π΄ satisfies RIP with πΏ π +1 β€ π and π₯ 0 is π -sparse and satisfies π΄ π₯ 0 =π. Then a π»( β
1 ) exists at π₯ 0 which satisfies conditions of the βsubgradient theoremβ. Implies that π₯ 0 is the unique minimum 1-norm solution to π΄π₯=π. π= π π₯ 0 π β 0 , π = π π₯ 0 π =0 Find subgradient π’ search for π€: π’= π΄ π π€ for πβπ: π’ π =π πππ π₯ 0 2-norm of the coordinates in π is minimized
11
Reconstruction from RIP
Let π§ be a vector with support π: π§ π =sign (π₯ 0 π ) Let π€= π΄ π π΄ π π π΄ π β1 π§ π΄ π has independent columns by RIP For coordinates in π: π΄ π π€ π = π΄ π π π΄ π π΄ π π π΄ π β1 π§=π§ For coordinates in π : π΄ π π€ π = π΄ π π π΄ π π΄ π π π΄ π β1 π§ Eigenvalues of π΄ π π π΄ π are in 1β πΏ π 2 , 1+ πΏ π 2 || π΄ π π π΄ π β1 ||β€ 1 1β πΏ π 2 , let π= π΄ π π π΄ π β1 π§, π β€ π 1β πΏ π 2 π΄ π π=π΄π where π has all coordinates in π equal 0 For πβ π : π΄ π π€ π = π π π π΄ π π΄π so | π΄ π π€ π |β€ πΏ π + πΏ 1 π 1β πΏ π 2 β€ πΏ π π 1β πΏ π 2 β€ 1 2
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.