Talk given at Sampta 2013.
The corresponding paper is :
Model Selection with Piecewise Regular Gauges (S. Vaiter, M. Golbabaee, J. Fadili, G. Peyré), Technical report, Preprint hal-00842603, 2013.
http://hal.archives-ouvertes.fr/hal-00842603/
6. L2
error stability: ||x(y) x0|| = O(||w||).
Promoted subspace (“model”) stability.
Goal: Performance analysis:
Regularized inversion:
Estimators
x(y) 2 argmin
x2RN
1
2
||y x||2
+ J(x)
! Criteria on (x0, ||w||, ) to ensure
Data fidelity Regularity
Observations: y = x0 + w 2 RP
.
7. Overview
• Inverse Problems
• Gauge Decomposition and Model Selection
• L2 Stability Performances
• Model Stability Performances
8. Coe cients x Image x
Union of Linear Models for Data Processing
Union of models: T 2 T linear spaces.
Synthesis
sparsity:
T
9. Coe cients x Image x
Union of Linear Models for Data Processing
Union of models: T 2 T linear spaces.
Synthesis
sparsity:
T
Structured
sparsity:
10. Coe cients x Image x
Union of Linear Models for Data Processing
D
Image x Gradient D⇤
x
Union of models: T 2 T linear spaces.
Synthesis
sparsity:
T
Structured
sparsity:
Analysis
sparsity:
11. Coe cients x Image x
Multi-spectral imaging:
xi,· =
Pr
j=1 Ai,jSj,·
Union of Linear Models for Data Processing
D
Image x Gradient D⇤
x
Union of models: T 2 T linear spaces.
Synthesis
sparsity:
T
Structured
sparsity:
Analysis
sparsity:
Low-rank:
S1,·
S2,·
S3,·x
12. Gauge: J : RN
! R+
8 ↵ 2 R+
, J(↵x) = ↵J(x)
Gauges for Union of Linear Models
Convex
13. Gauge: J : RN
! R+
8 ↵ 2 R+
, J(↵x) = ↵J(x)
J(x) = C(x) = inf {⇢ > 0 x 2 ⇢C}
C = {x J(x) 6 1} (assuming 0 2 C)
Gauges for Union of Linear Models
J(x)
C
1
Convex
14. Gauge:
, Union of linear models (T)T 2TPiecewise regular ball
J : RN
! R+
8 ↵ 2 R+
, J(↵x) = ↵J(x)
J(x) = C(x) = inf {⇢ > 0 x 2 ⇢C}
C = {x J(x) 6 1} (assuming 0 2 C)
Gauges for Union of Linear Models
J(x)
C
1
Convex
15. x
T
Gauge:
, Union of linear models (T)T 2TPiecewise regular ball
J : RN
! R+
8 ↵ 2 R+
, J(↵x) = ↵J(x)
J(x) = C(x) = inf {⇢ > 0 x 2 ⇢C}
C = {x J(x) 6 1} (assuming 0 2 C)
Gauges for Union of Linear Models
J(x) = ||x||1
T = sparse
vectors
J(x)
C
1
Convex
16. x
T
Gauge:
, Union of linear models (T)T 2TPiecewise regular ball
J : RN
! R+
8 ↵ 2 R+
, J(↵x) = ↵J(x)
J(x) = C(x) = inf {⇢ > 0 x 2 ⇢C}
C = {x J(x) 6 1} (assuming 0 2 C)
Gauges for Union of Linear Models
J(x) = ||x||1
x0
T0
T = sparse
vectors
J(x)
C
1
Convex
17. x
T
Gauge:
, Union of linear models (T)T 2TPiecewise regular ball
J : RN
! R+
8 ↵ 2 R+
, J(↵x) = ↵J(x)
J(x) = C(x) = inf {⇢ > 0 x 2 ⇢C}
C = {x J(x) 6 1} (assuming 0 2 C)
Gauges for Union of Linear Models
J(x) = ||x||1
x0
T0
T = sparse
vectors
|x1|+||x2,3||
x0
xT
T0
T = block
vectors
sparse
J(x)
C
1
Convex
18. x
T
Gauge:
, Union of linear models (T)T 2TPiecewise regular ball
J : RN
! R+
8 ↵ 2 R+
, J(↵x) = ↵J(x)
J(x) = C(x) = inf {⇢ > 0 x 2 ⇢C}
C = {x J(x) 6 1} (assuming 0 2 C)
Gauges for Union of Linear Models
J(x) = ||x||1
T = low-rank
matrices
J(x) = ||x||⇤
x
x0
T0
T = sparse
vectors
|x1|+||x2,3||
x0
xT
T0
T = block
vectors
sparse
J(x)
C
1
Convex
19. x
T
Gauge:
, Union of linear models (T)T 2TPiecewise regular ball
J : RN
! R+
8 ↵ 2 R+
, J(↵x) = ↵J(x)
J(x) = C(x) = inf {⇢ > 0 x 2 ⇢C}
C = {x J(x) 6 1} (assuming 0 2 C)
Gauges for Union of Linear Models
J(x) = ||x||1
T = low-rank
matrices
J(x) = ||x||⇤
x
x0
T0
T = anti-
sparse
vectors
J(x) = ||x||1
x
x0
T0
T = sparse
vectors
|x1|+||x2,3||
x0
xT
T0
T = block
vectors
sparse
J(x)
C
1
Convex
31. Noiseless recovery: min
x= x0
J(x) (P0)
Dual certificates:
x = x0
⌘
Proposition:
D = Im( ⇤
) @J(x0)
9 ⌘ 2 D () x0 solution of (P0)
Dual Certificate and L2 Stability
@J(x0)
x?
32. Noiseless recovery: min
x= x0
J(x) (P0)
Dual certificates:
Tight dual certificates:
x = x0
⌘
Proposition:
D = Im( ⇤
) @J(x0)
¯D = Im( ⇤
) ri(@J(x0))
9 ⌘ 2 D () x0 solution of (P0)
Dual Certificate and L2 Stability
@J(x0)
x?
33. Noiseless recovery: min
x= x0
J(x) (P0)
Dual certificates:
Tight dual certificates:
x = x0
⌘
Proposition:
D = Im( ⇤
) @J(x0)
¯D = Im( ⇤
) ri(@J(x0))
9 ⌘ 2 D () x0 solution of (P0)
Dual Certificate and L2 Stability
@J(x0)
x?
Theorem:
[Fadili et al. 2013] for ⇠ ||w|| one has ||x?
x0|| = O(||w||)
If 9 ⌘ 2 ¯D and ker( ) Tx0 = {0}
34. Noiseless recovery: min
x= x0
J(x) (P0)
Dual certificates:
Tight dual certificates:
x = x0
⌘
Proposition:
! The constants depend on N . . .
D = Im( ⇤
) @J(x0)
¯D = Im( ⇤
) ri(@J(x0))
9 ⌘ 2 D () x0 solution of (P0)
Dual Certificate and L2 Stability
@J(x0)
x?
Theorem:
[Fadili et al. 2013] for ⇠ ||w|| one has ||x?
x0|| = O(||w||)
If 9 ⌘ 2 ¯D and ker( ) Tx0 = {0}
35. Noiseless recovery: min
x= x0
J(x) (P0)
Dual certificates:
Tight dual certificates:
x = x0
⌘
Proposition:
[Grassmair 2012]: J(x?
x0) = O(||w||).
[Grassmair, Haltmeier, Scherzer 2010]: J = || · ||1.
! The constants depend on N . . .
D = Im( ⇤
) @J(x0)
¯D = Im( ⇤
) ri(@J(x0))
9 ⌘ 2 D () x0 solution of (P0)
Dual Certificate and L2 Stability
@J(x0)
x?
Theorem:
[Fadili et al. 2013] for ⇠ ||w|| one has ||x?
x0|| = O(||w||)
If 9 ⌘ 2 ¯D and ker( ) Tx0 = {0}
36. Overview
• Inverse Problems
• Gauge Decomposition and Model Selection
• L2 Stability Performances
• Model Stability Performances
37. ⌘ 2 D () and J (⌘) = 1
Minimal-norm Certificate
⌘ = ⇤
q
⌘T = e
⇢
T = Tx0
e = ex0
38. ⌘ 2 D ()
We assume ker( ) T = {0} and J piecewise regular.
and J (⌘) = 1
Minimal-norm Certificate
⌘ = ⇤
q
⌘T = e
⇢
T = Tx0
e = ex0
39. ⌘0 = argmin
⌘= ⇤q,⌘T =e
||q||
⌘ 2 D ()
We assume ker( ) T = {0} and J piecewise regular.
and J (⌘) = 1
Minimal-norm Certificate
⌘ = ⇤
q
⌘T = e
Minimal-norm pre-certificate:
⇢
T = Tx0
e = ex0
40. ⌘0 = argmin
⌘= ⇤q,⌘T =e
||q||
⌘ 2 D ()
We assume ker( ) T = {0} and J piecewise regular.
and J (⌘) = 1
Minimal-norm Certificate
Proposition: One has
⌘ = ⇤
q
⌘T = e
Minimal-norm pre-certificate:
⇢
T = Tx0
e = ex0
⌘0 = ( +
T )⇤
e
41. ⌘0 = argmin
⌘= ⇤q,⌘T =e
||q||
⌘ 2 D ()
We assume ker( ) T = {0} and J piecewise regular.
and J (⌘) = 1
Minimal-norm Certificate
Proposition:
||w|| = O(⌫x0 ) and ⇠ ||w||,Theorem:
the unique solution x?
of P (y) for y = x0 + w satisfies
Tx? = Tx0
and ||x?
x0|| = O(||w||) [Vaiter et al. 2013]
One has
⌘ = ⇤
q
⌘T = e
Minimal-norm pre-certificate:
⇢
T = Tx0
e = ex0
⌘0 = ( +
T )⇤
e
If ⌘0 2 ¯D,
42. [Fuchs 2004]: J = || · ||1.
[Bach 2008]: J = || · ||1,2 and J = || · ||⇤.
[Vaiter et al. 2011]: J = ||D⇤
· ||1.
⌘0 = argmin
⌘= ⇤q,⌘T =e
||q||
⌘ 2 D ()
We assume ker( ) T = {0} and J piecewise regular.
and J (⌘) = 1
Minimal-norm Certificate
Proposition:
||w|| = O(⌫x0 ) and ⇠ ||w||,Theorem:
the unique solution x?
of P (y) for y = x0 + w satisfies
Tx? = Tx0
and ||x?
x0|| = O(||w||) [Vaiter et al. 2013]
One has
⌘ = ⇤
q
⌘T = e
Minimal-norm pre-certificate:
⇢
T = Tx0
e = ex0
⌘0 = ( +
T )⇤
e
If ⌘0 2 ¯D,
45. J(x) = ||rx||1 (rx)i = xi xi 1
= Id I = {i (rx0)i 6= 0}
8 j /2 I, ( ↵0)j = 0⌘0 = div(↵0) where
Example: 1-D TV Denoising
x0
46. +1
1
I
J
Support stability.
J(x) = ||rx||1 (rx)i = xi xi 1
= Id I = {i (rx0)i 6= 0}
8 j /2 I, ( ↵0)j = 0⌘0 = div(↵0) where
||↵0,Ic || < 1
Example: 1-D TV Denoising
x0
x0
47. +1
1
I
J
`2
stability onlySupport stability.
x0
J(x) = ||rx||1 (rx)i = xi xi 1
= Id I = {i (rx0)i 6= 0}
8 j /2 I, ( ↵0)j = 0⌘0 = div(↵0) where
||↵0,Ic || < 1 ||↵0,Ic || = 1
Example: 1-D TV Denoising
+1
1
J
x0
x0