Documente Academic
Documente Profesional
Documente Cultură
Michael Elad
The Computer Science Department
The Technion – Israel Institute of technology
Haifa 32000, Israel
Defining Our Objective
and Directions
We Return to (P0)
(P0 ) min x 0
s.t. A x b
x
(P0 ) min x 0
s.t. A x b
x
(P0 ) min x 0
s.t. A x b
x
As opposed to the above, one could
consider the whole vector x as an
unknown rather than focusing on its
support
We have massive knowledge in
continuous optimization… but …
Main difficulty: (P0) is highly non-
smooth due to the L0 penalty
Solution: Smooth (P0) somehow,
which leads to “Relaxation Methods”
(P0 ) min x 0
s.t. A x b
x
Michael Elad
The Computer Science Department
The Technion – Israel Institute of technology
Haifa 32000, Israel
Greedy Algorithms:
The Orthogonal Matching
Pursuit Algorithm
Lets Go Greedy
Core idea:
exploit the
best support
from the last
round
Start: find the atom that best matches
Ax to b
Next: given the previously chosen atoms,
find the next one to best fit Ax to b
The solution grows the support one item at a time
The algorithm should stop when the error Ax-b is
getting close enough to zero
Michael Elad | The Computer-Science Department | The Technion
The Relation to the Pruned Tree
x 0
0 S={}
x 0
1 S={1} S={2} S={m}
x 0
2 S={2,1} S={2,3} S={2,m}
x 0
3 S={2,3,1} S={2,3,4} S={2,3,7} S={2,3,m}
x 0
4 S={2,3,7,1} S={2,3,7,4} S={2,3,7,8} S={2,3,7,m}
Greedy
A
algorithms
such as OMP +
build the
solution
sequentially
by adding one non-zero at a time b r 41320
x0, x1, x2, … , xk, …
In this path, the found solution xk may x 20143
not satisfy the equation: Axkb
We shall refer to this error as the current residual:
rk=b-Axk
OMP Strategy: choose the next non-zero such as to
reduce the “energy” in the residual as best as possible
Michael Elad | The Computer-Science Department | The Technion
OMP: The Details
5. Update Residual: r k b A x k
No Yes
rk 2
Stop
T
a i r k 1
z a i r k 1 0
T T
ai z opt T
a i r k 1
ai ai
2 2 2
T T
E(i) a i r k 1 a i r k 1 r k 1 2
a i r k 1
2
T
A r k 1
T
A r k 1 m
2
x k m in A x b
x
s.t. sup x S k
2
A
2
b
AS x b
2
min Sk x
2 x
1
T T †
xk AS AS AS b ASb
0 A S A S x k b A ST r k
T
2 1
T T
m in A S x b A AS Sk
AS b
x k 2 k k
2
Could we exploit the fact that we already
have the inversion from the previous
A ?
1
T
iteration S k 1
AS
k 1
2 2
The answer is positive –
There is a recursive method to
update the solution (which will
not be discussed here)
2 2
A S , ak
k 1
Michael Elad | The Computer-Science Department | The Technion
OMP: Complexity
Michael Elad
The Computer Science Department
The Technion – Israel Institute of technology
Haifa 32000, Israel
Variations over the
Orthonormal Matching
Pursuit
Other Greedy Algorithms
Weak Matching
OMP More
Pursuit (WMP)
accurate
Faster
Thresholding Matching Pursuit Least-Squares
(MP) OMP
5. Update Residual: r k b A x k
and this part will be
replaced in order to No
rk
Yes
Stop
obtain the LS-OMP 2
3. Update S k : S k S k 1 i0
2
4. LS : x k min A x b 2
s.t. sup x S k
x
5. Update Residual: r k b A x k
Comments:
o Step 1 can be done faster by
No Yes
exploiting the recursive LS alg. rk 2
Stop
o Step 4 is not needed since its
result is already given in step 1
o The residual is computed here
only for the stopping criterion
Claim 1: Complexity(OMP)<Complexity(LS-OMP)
Weak Matching
OMP More
Pursuit (WMP)
accurate
Faster
Thresholding Matching Pursuit Least-Squares
(MP) OMP
Main Iteration
Initialization
T
k 0, x 0 0 1. Compute p(i) a i r k 1 for 1 i m
k k 1
r0 b A x0 b
2. Choose i0 s.t. 1 i m, p(i0 ) p(i)
and S 0
3. Update S k : S k S k 1 i0
4. LS : x k min A x b s.t. sup x S k
x
5. Update Residual: r k b A x k
No Yes
rk 2
Stop
MP approach: x k i0 ai0 r k 1
T The new
coefficient
Keep xk-1 and simply
update it by
adding the new atom with its coefficient
Michael Elad | The Computer-Science Department | The Technion
MP: Details
Main Iteration
Initialization
1. Compute p(i) a iT r k 1 for 1 i m
k 0, x 0 0
k k 1 2. Choose i0 s.t. 1 i m, p(i0 ) p(i)
r0 b A x0 b
and S 0 3. Update S k : S k S k 1 i0
T
4. Update : x k Axx
LS : x k x kmin k 1b &s.t. (i0 ) x x k(iS
x ksup ) a i r k 1
0 k 0
x
Main Iteration
Initialization
1. Compute p(i) a iT r k 1 for 1 i m
k 0, x 0 0
k k 1 2. Choose i0 as soon as p i t r k 1
r0 b A x0 b 2
5. Update Residual: r k b A x k
Comments:
o WMP uses the same
‘trick’ as the MP for No Yes
rk Stop
avoiding the LS
2
computation
o t is a parameter – the
larger it is, the faster
(and less accurate) the
algorithm becomes
Michael Elad
The Computer Science Department
The Technion – Israel Institute of technology
Haifa 32000, Israel
The Thresholding Algorithm
What Have we Seen So Far?
Faster
Matching Pursuit Least-Squares
Thresholding (MP) OMP
2. Set : k 0, x 0 0 , S 0
No Yes
rk Stop
Here as well one could 2
i i i i 3. Update Residual: r k b A x k
1 2 3 k
2. Set : k 0, x 0 0 , S 0
No Yes
In high dim. problems, the LS step becomes
r prohibitive
Stop k 2
We have
a several Draw A nm
A
algorithms for somehow
(m n)
approximating
the solution of b n
A Solver of x̂
min x s.t. A x b Multiply min x
x 0
x 0 b=Ax0
s.t. A x b
Let’s test these
algorithms
Draw an s-sparse
This is the x0 at Random Compare
structure of the x 0 m
, x0 0
s n
experiment
We shall use
a random A Draw A A nm
of size somehow
(m n)
50×100
with normal b n
A Solver of x̂
entries Multiply min x
x 0
b=Ax0
We shall s.t. A x b
L2-normalize
the columns Draw an s-sparse
of A x0 at Random Compare
m
x0 , x0 0
s n
-2 -1 ? 1 2
x0 m
, x0 0
s n
We will compute
the relative error Draw A nm
2 A
x̂ x 0 somehow
(m n)
ErrorL2 2
2
x0 b n
2 A Solver of x̂
Multiply min x
We will also b=Ax0
x 0
2
x̂ x 0 2
2
x0 2
Ŝ S0
1
ˆ,S
max S 0
Michael Elad
The Computer Science Department
The Technion – Israel Institute of technology
Haifa 32000, Israel
Relaxation of the
L0-Norm: The Core Idea
Back to (P0)
(P0 ) min x 0
s.t. A x b
x
m
(P0 ) min x 0
s.t. A x b
x
x2 x
2
x 1 exp x x x
2
x
1.2
0.8
0.6
0.4
0.2
0
-4 -3 -2 -1 0 1 2 3 4
For example: 1
1=5.0
x
2 0.8 2=1.5
x 1 e
Decrease 3=0.5
0.6
4=0.1
0.4
0.2
0
-4 -3 -2 -1 0 1 2 3 4
m This problem
P2 { W } m in
T
x W
w kxx s.t.
2
s.t. AAx xbb
k has a closed
x
k 1
form solution
IRLS iterates between a solution of the L2
problem and an update of the weights
set Update the
Solve
diagonal
(P2{W})
x0 1 ? matrix W
x
P0 m in x
x 0
s.t. A x b
P1 m in x
x 1
s.t. A x b
Michael Elad
The Computer Science Department
The Technion – Israel Institute of technology
Haifa 32000, Israel
A Test Case:
Demonstrating and Testing
Relaxation Algorithms
Proposed Experiment: As Before
Draw A A
nm
somehow
(m n)
n
b An
Multiply Approximate
x̂
b=Ax0 Solver
of (P0)
m
x0
Draw an
s-sparse x0 Compare
at Random
x0 0
s n
Michael Elad | The Computer-Science Department | The Technion
Proposed Experiment: As Before …
WeDraw
use A
a random A of size
A 50×100 with iid
nm
b)Draw
A support
an recovery score
s-sparse x0 Compare
We
at average
Random the results over 200 experiments
x0 0
s n
Michael Elad | The Computer-Science Department | The Technion
Proposed Experiment: Algorithms
Draw A A
nm
Wesomehow
compare three algorithms:
(m n)
1. The OMP, applied exactly as in the previous
experiment
n
b An
2. The IRLS used for approximating the
Multiply Approximate
x̂
solution of (P1/2) (where 1 /2 x x )
0.5
b=Ax0 Solver
o 100 iterations of (P0)
xk 1
x0
m o The weights are given by 2
1.5
xk xk
Draw an
3.s-sparse
A direct
x0
solution of (P1) using the
Compare
Random instruction in Matlab
at linprog
x0 0
s n
Michael Elad | The Computer-Science Department | The Technion
Proposed Experiment: Results
2
x̂ x 0
2
2
x0 2
Ŝ S 0
1
ˆ ,S
m ax S 0
Michael Elad
The Computer Science Department
The Technion – Israel Institute of technology
Haifa 32000, Israel
Our Goal: Theoretical
Justification for the Proposed
Algorithms
Back to (P0)
Choose a A
nm
specific A
(m n)
n
b x̂
Multiply A Pursuit
b=Ax Algorithm
x
m
We shall prove that if s is small
enough (i.e. x0 is sufficiently
Draw an
s-sparse x0 sparse), then OMP,THR, and BP
are all guaranteed to give x̂ x
x 0
s n
Michael Elad | The Computer-Science Department | The Technion
Implications
(P0 ) min x 0
s.t. A x b
x
(P0 ) min x 0
s.t. A x b
x
The guarantees are going to develop adopt a
worst-case point of view
This means that for all {A,b} satisfying the
conditions, success is perfectly guaranteed
There exists a more sophisticated approach
that adopts a probabilistic point of view,
claiming the success of the pursuit with
probability →1
These offer more “generous” bounds but their
analysis is typically more complicated
Michael Elad | The Computer-Science Department | The Technion
Sparse & Redundant Representations
and Their Applications in
Signal and Image Processing
Guarantees of Pursuit Algorithms
Michael Elad
The Computer Science Department
The Technion – Israel Institute of technology
Haifa 32000, Israel
Equivalence:
Analyzing the OMP
Algorithm
Recall the OMP
Main Iteration
Initialization T
1. Com pute p(i) a i r k 1 for 1 i m
k 0, x 0 0
k k 1 2. Choose i0 s.t. 1 i m, p(i0 ) p(i)
r0 b A x0 b
3. Update S k : S k S k 1 i0
and S 0
s.t. sup x S k
2
4. LS : x k m in A x b 2
x
5. Update Residual: r k b A x k
No Yes
rk 2
Stop
The true n
A
Support S
m b
s x
b xa i i where
i 1
x1 x 2 xs 0
T T
RHS m ax b a j m ax x i ai a j
j s j s
i 1
s
b xa i i
i 1
s s
x i a i a j m ax x i a i a j
T T
m ax
j s j s
i 1 i 1
x1 s A
m ax a i a j A
T
i j
xa xa
T T T
LHS b a1 i i a1 x 1 i i a1
i 1 i 2
s
b xa i i
i 1
T
x1 x i a i a1
i 2
s
ab a b T
x1 x i a i a1
i 2
x 1 1 s 1 A
m ax a i a j A ; x 1 x i i 2
T
i j
b a1 x 1 1 s 1 A
T
>
x 1 s A m ax b a j
T
j s
1 s 1 A s A
1 1
1 A 2s A s 1
2 A
Michael Elad
The Computer Science Department
The Technion – Israel Institute of technology
Haifa 32000, Israel
Equivalence:
Analyzing the THR
Algorithm
THR: Terms for Success
T
Lower Upper T
m in b a i bound for bound for m ax b a j
1 i s j s
the LHS the RHS
T T
RHS m ax b a j m ax x i ai a j
j s j s
i 1
s
b xa i i
i 1
s s
x i a i a j m ax x i a i a j
T T
m ax
j s j s
i 1 i 1
x m ax s A
m ax a i a j A
T
i j
x
T T
LHS m in b a i m in t
a t ai
1 i s 1 i s
s t 1
b x t
at
t 1
s
T
m in x i x t a t ai
1 i s
t 1 ,t i
s
x t a t ai
T
m in x i
1 i s
t 1 , t i
ab a b s
T
m in x i m ax x t a t ai
1 i s 1 i s
t 1 , t i
x
T T
LHS m in b a i m in t
a t ai
1 i s 1 i s
s t 1
b x t
at
t 1
s
T
min x i max x t a t ai
1 i s 1 i s
t 1, t i
s
T
x m in m ax x t a t ai
1 i s
t 1 , t i
x m in x m ax s 1 A
m ax a i a j A ; x 1
T
xs
i j
T
Lower Upper T
m in b a i bound for bound for m ax b a j
1 i s j s
the LHS the RHS
x m in x m ax s 1 A >x m ax
s A
x m in
s 1 A s A
x m ax
1 x m in 1
s 1
2 x m ax A
Michael Elad | The Computer-Science Department | The Technion
THR Equivalence
0.2c 0.2c
2
x̂ x 0
2
2
x0 2
2
x̂ x 0
2
2
x0 2
Michael Elad
The Computer Science Department
The Technion – Israel Institute of technology
Haifa 32000, Israel
Equivalence: Analyzing the
Basis-Pursuit Algorithm
--- Part 1 ----
Basis Pursuit (BP) Rationale
P0 ˆ
x Arg m in x
x 0
s.t. b A x
by solving instead
P1 ˆ
x Arg m in x
x 1
s.t. b A x
C z b A x A z, z x , and z 1
x 1
This set represents the possible solutions of
BP that would be considered as erroneous
Our strategy will be to inflate this set (and
simplify it as a consequence), while showing
that it is actually empty under sparsity
conditions on x
Michael Elad | The Computer-Science Department | The Technion
An Error-Driven Set C
C z b A x A z, z x , and z 1
x 1
Rather than defining the set C w.r.t. candidate
alternative solutions, lets redefine it w.r.t. the
solution’s error: z=x+e
b Ax Az Ae 0
C e e 0 A e, e 0, and x e 1
x 1
Michael Elad | The Computer-Science Department | The Technion
Simplifying Ce – Part 1 (1)
T
0 Ae 0 A Ae
Multiply by A
e A A I e
T
Subtract e from
both sides
e A T
A I e
Apply abs on
both sides
T
e A A I e
ax by a x b y
T T
e A A I e e (11 I ) e
T A - I
A
Thus, |ATA-I|(11T-I)
[11T is a square matrix filled with 1-es]
Michael Elad | The Computer-Science Department | The Technion
Simplifying Ce – Part 1 (3)
T
e (11 I ) e T
e 1 e 1 e
e
1
(1 ) e e 1 1 e e 11
1
Observe that any e satisfying Ae=0 satisfies the last
inequality but not vice-versa. This means that
e 0 A e e e 1 e 1 1
We should be pleased with this replacement because
o A is replaced by its property
o The condition is posed w.r.t. |e|
C e e 0 A e, e 0, and x e 1
x 1
and inflated it to
e e 1 1, e 0
C e e 1
& xe1 x 1
We now target the term ||x+e ||1 ||x ||1 and aim
to simplify it to be stated in terms of |e| as well
Michael Elad | The Computer-Science Department | The Technion
Simplifying Ce – Part 2 (1)
xe 1
x 1
x i
ei x i 0
i 1
x i
ei x i ei 0
i 1 i s
xe x
T
e 2 1s e 0
1 1 1
e xe 1
x 1
e e 1
T
2 1s e 0
i.e. every e satisfying the RHS, satisfies the LHS
but not vice versa
We are pleased with this replacement because
o The condition is posed w.r.t. |e|
o The dependency on x is replaced by a simpler
dependency on the support S
Michael Elad | The Computer-Science Department | The Technion
Sparse & Redundant Representations
and Their Applications in
Signal and Image Processing
Guarantees of Pursuit Algorithms
Michael Elad
The Computer Science Department
The Technion – Israel Institute of technology
Haifa 32000, Israel
Equivalence: Analyzing the
Basis-Pursuit Algorithm
--- Part 2 ----
Simplifying Ce - Summary
Michael Elad | The Computer-Science Department | The Technion
Scale-Invariance
e e 1, e 0
C S e 1 1
& e 1 2 1S e 0
T
Observe that if eCS then eCS for all 0
Since our aim is to investigate whether CS is
empty or not, it is sufficient to consider the
intersection of this set with the unit L1 sphere
Thus, we impose || e ||1=1, getting
T
CN e e 1, e 1, 1 2 1 S e 0
1
1
x0
2
2
obtained bound
2
1 1
Perfect x s 1
0
2 A
Recovery for
s<8
?
Answer: In this experiment, 0.48, implying that
this bounds predict success for s<1.5
Too Pessimistic
Michael Elad | The Computer-Science Department | The Technion