Download presentation
Presentation is loading. Please wait.
Published byRosalyn James Modified over 9 years ago
1
SUPER: Sparse signals with Unknown Phases Efficiently Recovered Sheng Cai, Mayank Bakshi, Sidharth Jaggi and Minghua Chen The Chinese University of Hong Kong
2
b Compressive Sensing 2 ? ? n m k I. Introduction
3
b Compressive Phase Retrieval 2 ? ? k I. Introduction -2e iπ/3 Complex number m n 2 x → -x x →e iθ x
4
b Compressive Phase Retrieval ? ? k I. Introduction Applications: X-ray crystallography, Optics, Astronomical imaging… m n 2 Complex number
5
Compressive Phase Retrieval 2 ? ? I. Introduction Our contribution: 1. O(k) number of measurements (best known O(k) [1] ) 2. O(klogk) decoding complexity (best known O(knlogn) [2] ) [1] H. Ohlsson and Y. C. Eldar, “On conditions for uniqueness in sparse phase retrieval,” e-prints, arXiv:1308.5447 [2] K. Jaganathan, S. Oymak, and B. Hassibi, “Sparse phase retrieval: Convex algorithms and limitations,” in 2013 IEEE International Symposium on Information Theory Proceedings (ISIT), 2013, pp. 1022–1026. b m n
6
x2x2 x6x6 x5x5 x4x4 x3x3 b1b1 b2b2 b3b3 b4b4 x1x1 Bipartite graph n signal nodesO(k) measurement nodes k non-zero components II. Overview/High-Level Intuition 3
7
x2x2 x6x6 x5x5 x4x4 x3x3 b1b1 b2b2 b3b3 b4b4 x1x1 Bipartite Graph → Measurement Matrix II. Overview/High-Level Intuition & IV. Measurement Design x1x1 x2x2 x3x3 x4x4 x5x5 x6x6 b1b1 b2b2 b3b3 b4b4 Adjacent Matrix 4
8
x2x2 x6x6 x4x4 b1b1 b2b2 b3b3 b4b4 Useful Measurement Nodes n signal nodesO(k) measurement nodes II. Overview/High-Level Intuition 5
9
b1b1 b2b2 b3b3 b4b4 x2x2 x6x6 x4x4 Doubleton Multiton Singleton Useful Measurement Nodes II. Overview/High-Level Intuition 5
10
b1b1 b2b2 b3b3 b4b4 x2x2 x6x6 x4x4 Doubleton Multiton Singleton Useful Measurement Nodes Magnitude recovery Phase recovery Resolvable Δ |x 2 +x 4 | |x 2 | |x 4 | Solving a quadratic equation “Cancelling out” process: II. Overview/High-Level Intuition & V. Reconstruction Algorithm 5
11
Three Phases … n signal nodes … … Seeding Phase: Singletons and Resolvable Doubletons … Geometric-decay Phase: Resolvable Multitons … Cleaning-up Phase: Resolvable Multitons O(k) measurement nodes II. Overview/High-Level Intuition 6
12
Seeding Phase II. Overview/High-Level Intuition … n signal nodes … GIGI H 7
13
Seeding Phase II. Overview/High-Level Intuition … n signal nodes … GIGI “Sigma” Structure 7 x1x1 x2x2 H x2x2 x1x1
14
Seeding Phase II. Overview/High-Level Intuition … n signal nodes … GIGI H’ 7 H 1/2
15
Geometric-decay phase II. Overview/High-Level Intuition … n signal nodes … G II,l 1/4 H’ 8 H
16
Geometric-decay phase II. Overview/High-Level Intuition … n signal nodes … G II,l 1/8 O(k/logk) O(loglogk) stages H’ 8 H
17
Cleaning-up Phase II. Overview/High-Level Intuition … n signal nodes … G III |V( H ’)|=k H’ 9 H
18
Seeding Phase II. Overview/High-Level Intuition & III. Graph Properties … n signal nodes … ck measurement nodes … GIGI with prob. 1/k H H ’ Many Singletons Many Doubletons 10
19
Geometric-decay phase II. Overview/High-Level Intuition & III. Graph Properties … n signal nodes … ck/2 measurement nodes … G II,l H H ’ with prob. 2/k O(loglogk) Many Multitons 11
20
Geometric-decay phase II. Overview/High-Level Intuition & III. Graph Properties … n signal nodes … ck/4 measurement nodes … G II,l H H ’ with prob. 4/k O(k/logk) O(loglogk) 11 Many Multitons
21
Cleaning-up Phase II. Overview/High-Level Intuition & III. Graph Properties … n signal nodes … c(k/logk)log(k/logk) = O(k) measurement nodes … G III H H ’ |V( H ’)|=k with prob. logk/k Many Multitons 12
22
x2x2 x6x6 x5x5 x4x4 x3x3 b1b1 b2b2 b3b3 b4b4 x1x1 Bipartite Graph → Measurement Matrix II. Overview/High-Level Intuition & IV. Measurement Design Adjacent Matrix 13 x1x1 x2x2 x3x3 x4x4 x5x5 x6x6 b1b1 b2b2 b3b3 b4b4
23
x2x2 x5x5 b1b1 x1x1 Bipartite Graph → Measurement Matrix II. Overview/High-Level Intuition & IV. Measurement Design x1x1 x2x2 x3x3 x4x4 x5x5 x6x6 b1b1 b2b2 b3b3 b4b4 13
24
x2x2 x5x5 b1b1 x1x1 Bipartite Graph → Measurement Matrix II. Overview/High-Level Intuition & IV. Measurement Design x1x1 x2x2 x3x3 x4x4 x5x5 x6x6 b1b1 b2b2 b3b3 b4b4 α = (π/2)/n unit phase b 1,1 b 1,2 b 1,3 b 1,4 b 1,5 13
25
x2x2 x5x5 b1b1 x1x1 Bipartite Graph → Measurement Matrix II. Overview/High-Level Intuition & V. Reconstruction Algorithm α = (π/2)/n unit phase b 1,1 b 1,2 b 1,3 b 1,4 b 1,5 arctan(b 1,2 /ib 1,1 )/ α = 2 Guess: x 2 ≠ 0 and |x 2 | = b 1,1 /cos2α Verify: |x 2 | = b 1,5 ? 13
26
Seeding Phase: Giant Connected Component VI. Parameters Design O(k) different edges in graph H’ (By Coupon Collection) O(k) right nodes Each edge appears with prob. 1/k O(k) right nodes are singletons O(k) right nodes are doubletons Size of H’ is (1-f I )k (By percolation results) H H’ EXPECTATION! 14
27
Geometric-decay Phase VI. Parameters Design O(f II,l-1 k) different nodes appended in graph H’ (By Coupon Collection) O(f II,l-1 k) right nodes Each edge appears with prob. 1/f II,l-1 k O(f II,l-1 k) right nodes are resolvable multitons H H’ EXPECTATION! 15
28
VII. Performance of Algorithm Number of Measurements … n signal nodes … … … … Seeding Phase Geometric-decay Phase Cleaning-up Phase cf II,l-1 k measurement nodes O(k) ck measurement nodes O(k) c(k/logk)log(k/logk) measurement nodes O(k) 16
29
Decoding Complexity and Correctness BFS: O(|V|+|E|) for a graph G(V,E). O(k) in the seeding phase. “Cancelling out”: O(logk) for a right node. Overall decoding complexity is O(klogk). (1-ε II,l-1 )f II,l-1 <g II,l-1 <(1+ε II,l-1 )f lI,l-1 hold for all l. – Generalized/traditional coupon collection – Chernoff bound – Percolation results – Union bound VII. Performance of Algorithm 17
30
THANK YOU 謝謝
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.