Efficient Algorithms for Sparse Recovery Problems Sidharth Jaggi The Chinese University of Hong Kong Sheng Cai Mayank Bakshi Minghua Chen 1 Sparse Recovery Compressive Sensing Network Tomography ID: 926907
Download Presentation The PPT/PDF document "A Framework for Fast and" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
A Framework for Fast and Efficient Algorithms for Sparse Recovery Problems
Sidharth JaggiThe Chinese University of Hong Kong
Sheng
Cai
Mayank Bakshi
Minghua Chen
Slide21Sparse Recovery
Compressive Sensing
Network Tomography
Compressive Phase Retrieval
Group Testing
Slide3a
i
2
x
2
ai3x
3ai1x1
yi=Σj
aijxj
2Problem 1: Compressive Sensing - Linear Measurement
n
Slide4a
i
1
x
1
2
Problem 1: Compressive Sensing - Linear Measurement
a
i
2
x
2
a
i
3
x
3
y
i
=
Σ
j
a
ij
x
j
+
e
j
n
k
a
i
1
z
1
Slide5Problem 1: Compressive Sensing - Linear Measurement?
k
n
?
A
x = y
x
A
y
m
3
Slide6Problem 1: Compressive Sensing - Linear Measurement
?
k
n
?
A
x = y
x
A
y
m
n > m > k
3
Slide7Problem 1: Compressive Sensing - Linear Measurement
Z
A
(x +
z
) = y +
e
+
x
A
y
e
Approximate sparsity
Measurement noise
4
Interference
Transmission noise
Slide8Sparse Recovery
Z
A
(x +
z
) = y +
e
+
x
A
y
e
Approximate sparsity
Measurement noise
4
Linear or Non-Linear
Slide9Bottleneck linkyi
=Σj ∈
path i
xj
x
2
x
3
x
1
5
Problem 2: Network Tomography - Restricted Linear Measurements
Slide106Problem 3: Compressive Phase Retrieval - Non-Linear (Intensity) Measurement
yi
=|Σj
aij
xj
|
Can we recover 3D information from intensity measurements?
Slide117
Problem 4: Group Testing - Non-Linear (OR) Measurement
n
k
y
i
=
˅
j
∈
test
i
x
j
Slide128Our Contribution: Fast and Efficient Algorithms
Compressive Sensing
Network Tomography
Compressive Phase Retrieval
Group Testing
Slide13“Gold
Bar
” Puzzle I
100g
each
Sparse Recovery: Intuition
9
steals unknown
amount from
each gold bar
unknown thief
Which one?
How much?
Slide14“Gold
Bar
” Puzzle I
100g
Sparse Recovery: Intuition
9
Which one?
How much?
Slide15500g
490g
10g
“Gold Bar” Puzzle I
10
Step 1
Sparse Recovery: Intuition
1-sparse
Slide161,500g
1,460g
40g
“Gold Bar” Puzzle I
11
Step 2
40g
10g
=
4
Sparse Recovery: Intuition
Slide171,500g
1,460g
40g
“Gold Bar” Puzzle I
11
Step 2
Sparse Recovery: Intuition
Idea 1: Reconstruction
via
structured measurements
Slide18“Gold
Bar
” Puzzle II
Bad group
12
Sparse Recovery: Intuition
Good group: Puzzle I
Slide19“Gold
Bar
” Puzzle II
Bad group
12
Sparse Recovery: Intuition
110g
30g
∉
Z
structured measurements
Slide20“Gold
Bar
” Puzzle II
12
Good group
Sparse Recovery: Intuition
structured measurements
30g
10g
=
3
Slide21“Gold
Bar
” Puzzle II
12
Sparse Recovery: Intuition
80g
20g
=
4
0
0
0
-80
-20
Good group
Idea 2: Picking & Peeling
Slide22y
1
y
2
x
2
x
5
x4
x3x1
Multiton
Singleton
Measurement Structure: Bipartite Graph
13
Sparse Recovery: Framework
Slide23y
1
y
2
x
2
x
5
x4
x3x1
Multiton
Singleton
Decoding Process: Picking & Peeling
13
Sparse Recovery: Framework
Slide24y
1
y
2
x
2
x
5
x4
x3x1
Multiton
Singleton
Decoding Process: Picking & Peeling
13
Sparse Recovery: Framework
Slide25y1
y
2
x
2
x
5
x
4
x3x1
Decoding Process: Picking & Peeling
13
Sparse Recovery: Framework
Existence of useful nodes
--
Design of bipartite graphs
Decoding: iterative picking & peeling
--
Structured measurements
Multiton
Singleton
Slide2614Our Contribution: Fast and Efficient Algorithms
Compressive Sensing
Network Tomography
Compressive Phase Retrieval
Group Testing
SHO-FA
Decoding Complexity
# Measurements
Slide27Decoding complexity
# of measurements
°
RS’60
°
TG’07
°CM’06°C’08
°IR’08°
SBB’06°GSTV’06
°MV’12,KP’12°DJM’11
SHO-FA
Lower bound
Lower bound
SHO-FA:
Compressive
Sensing
15
O(
k
)
O(
k
)
Pawar, S.; Ramchandran, K., "A hybrid DFT-LDPC framework for fast, efficient and robust compressive sensing,"
50th Annual Allerton Conference on
Communication, Control, and Computing (Allerton)
, pp.1943,1950, 1-5 Oct. 2012
[SK12]
Slide2816?
k ≤
m<
n
?
n
m
k
Compressive Sensing: Problem
Problem 1:
Compressive
Sensing
Slide2917
4
3
4
n
4
3
4
k=2
Compressive Sensing: Intuition
A
(weighted) Biadjacency Matrix
Bipartite Graph → Measurement Matrix
Slide3018
4
3
4
n
Compressive Sensing: Intuition
Multiton
Singleton
Zeroton
k=2
Picking
“Gold Bar” Puzzle Recap
Slide3118
4
3
4
3
4
Multiton
Singleton
Zeroton
Compressive Sensing: Intuition
n
k=2
Peeling
Picking
“Gold Bar” Puzzle Recap
Slide32≥2|
S|
|
S|
L+L’≥
2
|S|3|S|≥L+2L’
(
L+L’)/(
L+2L’) ≥2/319
L/
(
L+L’
)
≥
1/2
Compressive Sensing: Design of Bipartite graphs
n
ck
d=
3
L
: # singletons
Existence of Singletons
Slide33Bipartite Graph → Measurement Matrix
n
ck
d=
3
20
A
Distinct weights
Compressive Sensing: Structured Measurements
Slide34Bipartite Graph → Measurement Matrix 20
Compressive Sensing: Structured Measurements
n
ck
A
Slide3521
Picking
& Peeling
Compressive Sensing: Decoding
Slide3622
Picking
& Peeling
Compressive Sensing: Decoding
Slide3723Our Contribution: Fast and Efficient Algorithms
Network Tomography
Compressive Phase Retrieval
Group Testing
SHO-FA
Slide3823Our Contribution: Fast and Efficient Algorithms
SHO-FA
Network Tomography
Compressive Phase Retrieval
Group Testing
FRANTIC
Slide3924
End-to-end delay
Edge delay
Sparse
Problem 2: Network Tomography
Network Tomography: Problem
Slide40Challenge: Topology constraints
25
Slide41,
Integer entries: “Gold bar” puzzle
Approach: Cancellation+“Loopy” measurements
25
Slide4227Our Contribution: Fast and Efficient Algorithms
SHO-FA
Network Tomography
Compressive Phase Retrieval
Group Testing
FRANTIC
Slide4327
Our Contribution: Fast and Efficient Algorithms
SHO-FA
Compressive Phase Retrieval
Group Testing
FRANTIC
SUPER
Slide44?
n
x
1
x
=
?Problem 3: Compressive Phase Retrieval28
| |
+
a
1
a
5
a
7
x
5
x
7
a
1
x
1
+
a
5
x
5
+
a
7
x
7
y
1
=
|
a
1
x
1
+
a
5
x
5
+
a
7
x
7
|
Intensity
detector
+
+
y
1
y
2
y
m
Decoder
k
x
Compressive Phase Retrieval: Problem
Slide45b
Problem 1: Compressive Sensing
?
n
m
k
29
x
Compressive Phase Retrieval: Problem
Slide46b
30
?
k
2
e
i
π
/
3n
2
2
Intensity
Problem 3: Compressive Phase Retrieval
x
Compressive Phase Retrieval: Problem
Slide47b
Problem 3: Compressive Phase Retrieval
30
?
-
2
eiπ/3
n2
2
Intensity-x
Compressive Phase Retrieval: Problem
Slide48b
Problem 3: Compressive Phase Retrieval
30
?
2
e
i(π/3+ θ)
n2
2
Intensityei
θ
x
Compressive Phase Retrieval: Problem
Slide4931
4
3
4
n
ck
4
3
4
k=2
Multiton
Singleton
Compressive Phase Retrieval: Intuition
Challenges:
“Peeling” is not allowed
Magnitude + Phase
“Gold Bar” Puzzle Recap
Slide5032Phases from Magnitudes?
n,k = 1
y1
=
|a
1 x1|
|x1| =|y
1/a1|
n,k = 2
y1 = |a1 x
1|
y2
=
|
a
2
x
2
|
|
x
1
|
=
|
y
1
/a
1
|
|
x
2
|
=
|
y
2
/a
2
|
y
3
=
|
x
1
+
x
2
|
y
4
=
|
x
1
+i
x
2
|
Δ
|
x
1
+
x
2
|
|
x
1
|
|
x
2
|
OR
Δ
|
x
1
+
x
2
|
|
x
1
|
|
x
2
|
x
1
=
|
y
1
/a
1
|
e
iθ
|
x
1
+
i
x
2
|
|
x
1
+
i
x
2
|
x
=
|
y
1
/
a
1
|
|
y
2
/a
2
|
e
iΔ
e
iθ
y
4
=
x
2
x
1
y
1
y
3,
y
4
y
2
“
Σ
” Structure
Compressive Phase Retrieval: Intuition
Slide51x1x
6x
5
x
4
x3Phases from Magnitudes?
33x2
y
1
y
2
y
3
y
4
n
m
0
0
0
Compressive Phase Retrieval: Intuition
Slide52x2x
6x
4
y
1
y2
y3y4Doubleton
Singleton
Phases from Magnitudes?
33Multiton
Compressive Phase Retrieval: Intuition
Slide53x
2
x
6
x
5
x
4
x
3
y1
y
2
y
3
y
4
x
1
Graph properties
34
n
ck
log
k
with prob.
1/k
The probability that a
right node
is singleton
Compressive Phase Retrieval: Design of Bipartite graphs
Slide54x
2
x
6
x
5
x
4
x
3
y1
y
2
y
3
y
4
x
1
Graph properties
34
with prob.
1/k
A
right node is singleton w.p.
e
-1
A
right node is doubleton w.p.
e
-1
/2
Compressive Phase Retrieval: Design of Bipartite graphs
n
ck
log
k
Slide55x2x
6x
5
x
4
x3
x1
y
1
y
2
y
3
y
4
Bipartite Graph
→
Measurement Matrix
35
x
1
x
2
x
3
x
4
x
5
x
6
y
1
y
2
y
3
y
4
(weighted) Biadjacency Matrix
y
1
y
2
y
3
y
4
x
1
x
2
x
3
x
4
x
5
x
6
Compressive Phase Retrieval: Structured Measurements
n
ck
log
k
Slide56x
2
x
5
y
1
x1
36
x
1x2
x
3
x
4
x
5
x
6
y
1
y
2
y
3
y
4
Compressive Phase Retrieval: Structured Measurements
Bipartite Graph
→
Measurement Matrix
Slide57Identifying Singletons37
α
=
(π/
2)/n
unit phasey1,1
y1,2b1,3
b1,4b
1,5
x1
x2
x
3
x
4
x
5
x
6
y
1
y
2
y
3
y
4
x
2
x
5
y
1
x
1
Compressive Phase Retrieval: Decoding
Slide58Identifying Singletons37
α
=
(π/
2)/n
unit phasey1,1
y1,2b1,3
b1,4b
1,5
x
1
x
2
x
3
x
4
x
5
x
6
y
1
y
2
y
3
y
4
x
2
y
1
0
0
Compressive Phase Retrieval: Decoding
Slide59Identifying Singletons37
α
=
(π/
2)/n
unit phasey1,1
y1,2b1,3
b1,4b
1,5arctan(y1,2
/y1,1)/ α = 2
x2
≠ 0 and |x2|
= y
1,1
/cos2
α
y
1,1
= cos2
α
|x
2
|
y
1,2
= sin2
α
|x
2
|
x
2
y
1
0
0
Singleton!
Compressive Phase Retrieval: Decoding
Slide60Singleton
Identifying Singletons38
Magnitude
recovery
x
6x
4x2
0
0
0
y
1
y
2
y
3
y
4
y
5
Compressive Phase Retrieval: Decoding
Slide61Singleton
Singletons + Doubletons39
Magnitude
recovery
Phase
recovery Resolvable
x6x4x2
y
1
y
2
y
3
y
4
Doubleton
Useless
y
5
Compressive Phase Retrieval: Decoding
Slide62Decoding Graph40
…
n
signal nodes
…
k
non-zero
components
k
nodes
G
H
Compressive Phase Retrieval: Decoding
Slide63Decoding Graph40
n signal nodes
“
Σ
” Structure
x
1
x
2
Hx2
x
1
G
…
…
Compressive Phase Retrieval: Decoding
Transitive Property
Slide64Decoding Graph40
n signal nodes
“
Σ
” Structure
x
1
x
2
Hx2
x
1
G
…
…
Compressive Phase Retrieval: Decoding
ck
log
k
Goal:
O
(
k
) measurements
O
(
k
log
k
) decoding complexity
Slide65Decoding Graph40
n signal nodes
“
Σ
” Structure
x
1
x
2
Hx2
x
1
G
…
…
Compressive Phase Retrieval: Decoding
ck
Goal:
O
(
k
) measurements
O
(
k
log
k
) decoding complexity
Slide66y
1
y
2
y
3
y4
Singleton
Decoding via Resolvable Multiton
41
Doubleton
Multiton
x
2
x
6
x
4
Compressive Phase Retrieval: Decoding
Slide67Multiton
Singleton
Decoding via Resolvable Multiton
41
Resolvable
Solving a quadratic equation
“Peeling” process:
y
1
y2y3
y
4
x
2
x
6
x
4
Doubleton
Compressive Phase Retrieval: Decoding
M
M
P
P
M
P
Slide68Two Phases42
n signal nodes
…
…
…
1
st
Phase
O
(
k
)
measurements
O
(
k
)
complexity
Singletons
&
Doubletons
2
nd
Phase
O
(
k
)
measurements
O
(
k
log
k
)
complexity
Multitons
…
…
Compressive Phase Retrieval: Design of Bipartite graphs
Slide691st Phase43
n
signal nodes
ck
measurement nodes
…
G
Iwith prob. 1/k
H ’
Many SingletonsMany Doubletons
H
…
…
Compressive Phase Retrieval: Design of Bipartite graphs
Slide702nd Phase44
n
signal nodes
ck/
2
measurement
nodes…GII,
1with prob. 2/k
Many
MultitonsH
H ’
…
…
Compressive Phase Retrieval: Design of Bipartite graphs
Slide712nd Phase44
n
signal nodes
ck/
4
measurement nodes
…GII,2
with prob. 4
/k
H
H ’
Many
Multitons
…
…
Compressive Phase Retrieval: Design of Bipartite graphs
Slide7245
Our Contribution: Fast and Efficient Algorithms
SHO-FA
Compressive Phase Retrieval
Group Testing
FRANTIC
SUPER
Slide7345
Our Contribution: Fast and Efficient Algorithms
SHO-FA
Group Testing
FRANTIC
SUPER
GROTESQUE
Slide74Group Testing
IN
OUT
Negative
0
Positive
1
46
Group Testing: Problem
Slide75GROTESQUE47Group Testing: Structured Measurements
Multiplicity
Localization
n
items
k
defectives
GROTESQUE
k
> 1
k
= 1
k
= 0
Slide76Multiplicity
48
Group Testing: Structured Measurements
?
Half of 1’s
Distinct columns
n
m
1
Slide77Localization
49
Group Testing: Structured Measurements
?
n
m
2
Binary
expression of index
Slide7850
Our Contribution: Fast and Efficient Algorithms
SHO-FA
Group Testing
FRANTIC
SUPER
GROTESQUE
Slide79y
1
y
2
x
2
x
5
x4
x3x1
Multiton
Singleton
“Good” Bipartite Graph + Structured Measurements
51
Sparse Recovery: Summary
Picking &
Peeling
Slide80Thank you謝謝
Slide81y1
y
2
x
2
x
5
x
4
x3x1
Decoding Process: Picking & Peeling
15
Sparse Recovery: Framework
y
3
“Useful”
Unknown support
Useful nodes
-- Existence (design of bipartite graphs)
-- Identification (structured measurements)
Decoding
Multiton
Singleton
Slide82Decoding graph43
H
O
(
k
log
k
) right nodes are singletons
O
(
k
log
k
) right nodes are doubletons
(By Chernoff bound)
O
(
k
log
k
) different edges in graph
(By Coupon Collection)
O
(
k
log
k
) measurement nodes
The decoding graph is connected
(By connectivity of random graph)
Goal:
O
(
k
) measurements
O
(
k
log
k
) decoding complexity
March 20, 15
Compressive Phase Retrieval: Decoding
Slide83Decoding graph45
H
O
(
k
) right nodes are singletons
O
(
k
)
right nodes are doubletons
(By Chernoff bound)
O
(
k
) different edges in graph
(By Coupon Collection)
O
(
k
) measurement nodes
Giant connected component
(By percolation results)
Goal:
O
(
k
) measurements
O
(
k
log
k
) decoding complexity
March 20, 15
Compressive Phase Retrieval: Decoding
Slide84Singleton
Identifying doubletons39
Magnitude recovery
x
6
x4
x2
y
1
y
2
y
3
y
4
y
5
Compressive Phase Retrieval: Decoding
Slide85Singleton
Identifying doubletons39
Magnitude recovery
x
6
x4
x2
x
2
x
2
y
1
y
2
y
3
y
4
y
5
Compressive Phase Retrieval: Decoding
Slide86Singleton
Identifying doubletons39
Magnitude recovery
x
6
x4
x2
x
2
x
2
x
4
x
4
y
1
y
2
y
3
y
4
y
5
Compressive Phase Retrieval: Decoding
Slide87Singleton
Identifying doubletons39
Magnitude recovery
x
6
x4
x2
x
2
x
2
x
4
x
4
x
6
y
1
y
2
y
3
y
4
y
5
Compressive Phase Retrieval: Decoding