Академический Документы
Профессиональный Документы
Культура Документы
Michael Elad
The Computer Science Department
The Technion – Israel Institute of technology
Haifa 32000, Israel
Uniqueness vs. Stability
– Gaining Intuition
The (P0) Problem
(P0 ) min x 0
s.t. A x b
x
2 2
(P ) m in x
0 0
s.t. Ax b 2
x
A x0
1.0 0 0.6 0.8
A
0 1.0 0.8 0.6 0.5
b A x0 v
0
0
x0
2
0
The 4 atoms
Michael Elad | The Computer-Science Department | The Technion
An Illustrative Example
A x0
The Possible
Solutions b A x0 v
A x̂
2 2
Ax b 2
Solutions of
cardinality 1
2 2
(P0 ) m in x 0
s.t. Ax b 2
x
The Possible A x0
Solutions
b A x0 v
A x̂
A x0
The Possible
Solutions b A x0 v
A x̂
2 2
Ax b 2
2 2
(P0 ) m in x 0
s.t. Ax b 2
x
Michael Elad
The Computer Science Department
The Technion – Israel Institute of technology
Haifa 32000, Israel
The Restricted Isometry
Property (RIP)
Recall: The Uniqueness of (P0)
Thus:
A ˆ
x x0 2 & ˆ
x x0 2s
2 0
Define z ˆ
x x0 :
Az 2
2 & z 0
2s
Az 2
2
A As
s atoms
Consider all (m-choose-s) such possible
sub matrices As. Then …
Assume that spark{A}=s. Then, at least one
of these As has linearly dependent columns
Equivalently: m in A s 0
Michael Elad | The Computer-Science Department | The Technion
Generalizing this Definition
1 s
s 2 2
z z 2
Asz 2
such that
x
m
Michael Elad
The Computer Science Department
The Technion – Israel Institute of technology
Haifa 32000, Israel
Key Properties of the
Restricted Isometry Property
(RIP)
Property 1: 1=0
A A s ak
s
Thus 2 T
T 2
Asz z ak ak z z 1 0
2
1
0
2
On the other hand we must have 0 A s z 2
1
s≥Spark{A}
s
1
Michael Elad | The Computer-Science Department | The Technion
Property 4: Relation to (A)
s (s 1)
Michael Elad
The Computer Science Department
The Technion – Israel Institute of technology
Haifa 32000, Israel
Theoretical Study of (P0)
The Stability Problem
An v
2
s-sparse
vector
x0 Multiply
by A
b xv
b0 A x 0
Clearly, x̂ x0 s
0 0 Min x 0
s.t.
x̂ x
b Ax 2
Stability How far is x0 from x̂ ?
Thus: A x̂ x 0 2
A x̂
2
& ˆ
x x0 2s
0
1 2 s
2 2 2
d 2
Ad 2
4
2
2 2 4
d ˆ
x x0
2 2
1 2s
Conclusions:
We have stability – the solution of (P0) is not far off
The sparser x0 and the less coherent the dictionary is,
the stronger the stability
Michael Elad
The Computer Science Department
The Technion – Israel Institute of technology
Haifa 32000, Israel
Performance of Pursuit
Algorithms – General
Stability of Any Pursuit Algorithm
Conclusions:
o Any pursuit algorithm that can create a feasible
and sparse solution is necessarily stable
o However, this does not explain why such
pursuit will succeed in the first place
Michael Elad | The Computer-Science Department | The Technion
What is the Best Possible Performance ?
Suppose that x is s-sparse, and we create
b A x v A s x s v w here v~ 0, I
2
2 1
2 T T
E A s ˆz opt A s x s
b
E As A As s
A b As xs
s
2
2
1
2
E As A As
b
T
s
A
T
s As xs v As xs
2
2
1
T T
E As A As s
As v
2
1
2 T T
E As A As
b s
A v s
2
=I
A A A A A v
T 1 1
T T T T
E v As A As s s s s s s
E v A A A A v
T 1
T T
s s s s
= I
2
tr A E v v A A A tr I s
T 1
T T 2 2
s s s s s
A
2 1
2 T T
E ˆz opt x s
x
E s
As A b xs
s
2
2
A
1
2
E
x
T
s
As A
T
s As xs v xs
2
2
A
1
T T
E s
As As v
2
A
1
2 T T
E
x s
As A v s
2
T 1 1
T T T
E v As As As As As As v
=2I
tr A E v v A tr A A
1
A
T 2
T T 2 T
s s s
As s s
x is an b=Ax+v
s-sparse where v is
vector noise
Solve (P0ε) or
If s is small enough, the approximate it in
solution obtained is stable order to recover x
Michael Elad
The Computer Science Department
The Technion – Israel Institute of technology
Haifa 32000, Israel
Basis-Pursuit Stability
Guarantee
Basis-Pursuit Stability (1)
A x̂ x 0 2
2
Ad 2
A ˆ
x x0
2
2 d x x0
ˆ
1 d
2 2
2
d 1
x̂ x0 1
1
0 d x0 1
x0 1
d xk xk
0 0
dk
k
k s k s
T
dk dk d 1
2 1s d
k s k s
ab b a
T
d 1
2 1s d
4 1 d
2 2 2 T
2
d 1
& d 2 1s d
1
Therefore,
2
4 1 d
2 2
1 d
2 2 T
2
d 1
4 1s d
2
Using v s
v 1
s v 2
for the last term
4 1 d
2
4 s d k 1 4 s d
2 2 2
2 2
k s
Michael Elad
The Computer Science Department
The Technion – Israel Institute of technology
Haifa 32000, Israel
Thresholding Stability
Guarantee: Worst-Case
Proof Strategy
s m x0
b xa i i v where
i 1
x1 x 2 xs 0
Michael Elad | The Computer-Science Department | The Technion
THR: Terms of 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 T
RHS m ax b a j m ax v a j x i ai a j
j s j s
i 1
s
b v xa i i
T
m ax a i a j
i j
i 1
s
m ax v a j m ax x i a i a j
T T
j s j s
i 1
x m ax s
T
v aj v 2
aj
2
Cauchy-Schwartz Inequality
x
T T T
LHS m in b a i m in v a i t
a t ai
1 i s 1 i s
s t 1
b v x t
at
t 1
s
T T
m in v a i x i x t a t ai
1 i s
t 1 ,t i
s
x t a t ai
T T
m in x i v a j
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
LHS
T
x m in x t a t ai
t 1 , t i
x m in x m ax s 1 A
m ax a i a j A
T
i j
x1 xs
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 >x m ax
s
x m in 2
s 1 s
x m ax x m ax
1 x m in 1
s 1
2 x m ax x
m ax
Michael Elad
The Computer Science Department
The Technion – Israel Institute of technology
Haifa 32000, Israel
OMP Stability Guarantee
OMP: First Step Success
T T T
RHS m ax b a j m ax v a j x i ai a j
j s j s
i 1
s
b v 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
T
m ax a i a j
i j
LHS b a1 v a1 i
T T T T
x m axx i a a1 x i a i a1
i 1 i 2
s
b v xa i i
i 1
s
T
x m ax x i a i a1
i 2
s
T
x m ax x i a i a1
i 2
x m ax 1 s 1
m ax a i a j A ; x m ax x i i 2
T
i j
b a1 x m ax 1 s 1
T
T
> x m ax s m ax b a j
j s
2
1 s 1 s
x m ax
1 1
s 1
2 x m ax
Michael Elad | The Computer-Science Department | The Technion
Moving to the Next Step
Observe That
The updated residual is a linear
combination of the same s atoms (as in b)
s
r 1 b c 1 a i0 xa i i v
i 1
we have that x m ax m ax x i x m in
1 i s
Proof:
In rk, only the coefficients of the chosen atoms
change, while the rest remain with their original
values
Thus, the maximal non-zero (in abs value) cannot
be smaller than |xmin|, as claimed
Michael Elad | The Computer-Science Department | The Technion
Worsening the Condition
If x0 is sparse enough
1 1
x s 1
0
2 A x m ax
then the first step of the OMP is successful
Obviously, this implies that the same remains
true with the following worse condition
1 1
x s 1
0
2 A x m in
The idea: With this change we guarantee the
success of every OMP step
Michael Elad | The Computer-Science Department | The Technion
Worsening the Condition
1 1 4
2
BP s 1
4 1 4s 1
1 1
2
4
OMP s 1
2
A x m in
1 s 1
2
1 x m in 1 4
THR s 1
2
x m ax A x m ax 1 s 1
Michael Elad
The Computer Science Department
The Technion – Israel Institute of technology
Haifa 32000, Israel
Rate of Decay of the
Residual in Greedy Methods
Recall the Matching Pursuit Algorithm
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
4. Update x k : x k x k 1 & x k (i0 ) x k (i0 ) aiT r k 1
0
5. Update Residual: r k b A x k
rk b A xk
T
b A x k 1 a i0 r k 1 a i0 No
rk 2
Yes
Stop
T
r k 1 a i0 r k 1 a i0
The question we address: How fast can we expect
this residual to decay?
While different from the rest of the analysis shown
so far, this result is quite interesting and elegant
Michael Elad | The Computer-Science Department | The Technion
The Residual Recursion
T
r k r k 1 a i0 r k 1 a i0
2
2 T
rk 2
r k 1 a i0 r k 1 a i0
2
a
2 2 2
T T
r k 1 2
2 a i0 r k 1 i0 r k 1
2 2
T
r k 1 2
a i0 r k 1
2 2
T
r k 1 2
m ax a i r k 1
1 i m
2 2 2
T
rk 2
r k 1 2
A r k 1
v 2
v T
Could s(A)=0? Since A is
2
A v m
full-rank and mn, this is
impossible because it implies A v
T
2
T
A v
We have defined s A mvin 2
v 2
v
v 2
2
1 s A r 0
k 2
1 s A b
2 k 2
rk 2 2
Implications:
Clearly this rule applies to OMP and LS-OMP
as they are more aggressive in reducing the
residual’s energy
This result provides another justification for
the hope to get a sparse solution from MP
and other greedy methods