Date post: | 31-Dec-2015 |
Category: |
Documents |
Upload: | lydia-joseph |
View: | 220 times |
Download: | 1 times |
Multilevel approach for signal restoration problems with Toeplitz matrices
Malena Español, Tufts UniversityMisha Kilmer, Tufts University
10th Copper Mountain Conference on Iterative Methods
2
Outline Background Multilevel Method Algorithm Implementation Numerical Example Conclusion and Future Work
10th Copper Mountain Conference on Iterative Methods
3
Ill-posed Problem
A problem is ill-posed if its solution is not unique, or its solution does not depend continuously
on the data
)()( ,
kindfirst ofequation integral Fredholm :Example
sgdttft)K(s
10th Copper Mountain Conference on Iterative Methods
4
Discrete Ill-Posed problem
holds condition Picard Discrete
yoscillator more become ectorssingular v
noise (white)unknown is
gap without aluessingular v Decaying
:Properties
e
matrix dconditione-ill large, a is where
,
model theand ,given , Find
nm
truetrue
true
RA
ebbAx
bAx
10th Copper Mountain Conference on Iterative Methods
5
Need for regularization
)(
1
)(
1ii
)(
1i
†
)(
1i
†
)(
)(
:are solutions squareleast Then the
. of SVD theLet
Arank
i
Arank
itrue
i
T
i
trueT
Arank
i i
trueT
true
Arank
i i
trueT
truetrue
errorxeubu
ebuebAx
bubAx
AVUA
ii
i
i
10th Copper Mountain Conference on Iterative Methods
6
Regularization
small is if 0
large is if 1
where
solution dRegularize
1i
i
ii
n
i i
T
ireg
bux i
10th Copper Mountain Conference on Iterative Methods
7
Regularization methods
CGLS.or LSQR, eg.
methods, Krylov of iterationsk :Methods Iterative
min
:tionRegulariza Tikhonov
: (TSVD) SVD Truncated
2
2
22
2
1i22
2
1i
LxbAx
bux
bux
x
n
i i
T
i
iTik
k
i i
T
TSVD
i
i
10th Copper Mountain Conference on Iterative Methods
8
Two-Level Method
cii
ic
i
iiii
iiii
iii
iiii
iii
iiii
iii
xxx
rxA
xAbr
xPxx
rxA
PARA
rRr
xAbr
bxA
Solve""
Solve""
Solve""
1
111
1
1
Pre-Smoothing
Post-Smoothing
Coarse-Grid Correction
Restriction
Prolongation
10th Copper Mountain Conference on Iterative Methods
9
Multilevel Method
If End
Solve
,
Solve
Else
Solve
gridcoarsest If
function
1
111
11
cii
ic
i
iiii
iiii
iii
iiiiiii
iiii
iii
iii
iii
xxx
rxA
xAbr
xPxx
),bMGM(Ax
PARArRr
xAbr
bxA
bxA
),bMGM(Ax
10th Copper Mountain Conference on Iterative Methods
10
Restriction and Prolongation Operators
11000000
00110000
00001100
00000011
11000000
00110000
00001100
00000011
2
2TW
Haar wavelet transform
W2W1
10th Copper Mountain Conference on Iterative Methods
11
Wavelet Domain
2
1
2
1
43
21
ˆ
ˆ
ˆ
ˆˆˆ
ˆˆ
.ˆ and ˆ ,ˆ where
ˆˆˆ
becomes domain In wavelet
b
b
x
x
AA
AA
bWbxWxAWWA
bxA
bAx
TTT
10th Copper Mountain Conference on Iterative Methods
12
Coarse-scale equation
x
1x̂111
22111
2
1
2
1
43
21
ˆˆˆ
ˆˆˆˆˆ
ˆ
ˆ
ˆ
ˆˆˆ
ˆˆ
bxA
xAbxA
b
bx
x
AA
AA
10th Copper Mountain Conference on Iterative Methods
13
.ˆˆˆ
Then,
ˆˆ
ˆˆˆ
ˆ
ˆˆ
ˆ
ˆˆ
ˆˆˆˆˆˆˆˆˆ
111
)ˆ(
1i
2211
221122111
1
bxA
xAueubu
xAebxAbxA
Arank
i i
T
i
T
i
trueT
true
iii
Coarse-scale equation
10th Copper Mountain Conference on Iterative Methods
14
p
p
p
xxLbxA
bAx
bxA
1
2
2111ˆ
111
111
ˆˆˆˆmin
3)or 2 ,ˆ,ˆLSQR(ˆ
solve we
ˆˆˆ
ofsolution dregularize aget To
1
Coarse-Grid Correction
Pre-smoothing
Pre-smoothing
Coarse-Grid Correction
10th Copper Mountain Conference on Iterative Methods
15
1
3
1
2
12
4
2 ˆˆ
ˆ
ˆ
ˆˆ
ˆ
ˆx
A
A
b
bx
A
A
Post-Smoothing
p
pprep
newx
xWxWxLrxA
A)ˆˆ(ˆ
ˆ
ˆmin 2211
2
2
2
4
2
ˆ2
newr
10th Copper Mountain Conference on Iterative Methods
16
Multilevel Method
If End
ˆ
),,,(Newtonˆ
ˆ
ˆ
;
If End
,3),(
gridfinest If
Else
)(Newtonor solvedirect
gridcoarsest If
function
122
21
2
111
1111
111
11
iinew
i
inew
inew
Tii
inew
iiinew
iiinew
iii
iTii
Ti
iiii
iii
iii
iii
xWxx
xLWrWWAx
xAbr
xWxx
),bMGM(Ax
WAWArWb
xAbr
bALSQRx
non
,bAx
),bMGM(Ax
10th Copper Mountain Conference on Iterative Methods
17
Toeplitz Matrices
),,,,,(: 1101)1(
0321
)3(012
)2(101
)1(210
mm
mmm
m
m
m
ttttttvectorToeplitz
tttt
tttt
tttt
tttt
A
10th Copper Mountain Conference on Iterative Methods
18
Toeplitz structure inheritance
A
43
21
ˆˆ
ˆˆˆ
AA
AAA
.ˆeach ofvector -Toeplitz theknow weMoreover,
Toeplitz. are ˆ and ˆ,ˆ,ˆ then Toeplitz, is If
:Theorem
4321
iA
AAAAA
10th Copper Mountain Conference on Iterative Methods
19
Pre-smoothing
productvector -matrixFast structure Toeplitz
LSQR. of iterations 3or 2 applyingby
ˆˆˆ
solve We
111
bxA
10th Copper Mountain Conference on Iterative Methods
20
Coarse-Grid Correction
.ˆon depend and where,ˆ
ˆmin
Method sNewton'by ˆˆˆˆmin
or
solvedirect aby ˆˆˆ
solve We
11
2
21
11
1
2
2111ˆ
111
1
xrDxDL
rd
DL
A
xLbxA
bxA
d
p
p
p
x
10th Copper Mountain Conference on Iterative Methods
21
Post-Smoothing
.ˆon depend ~ and ~, where,~
~ˆ
ˆ
min
Method sNewton'by
)ˆˆ(ˆˆ
ˆmin
solve We
221
2
2
2
1
2
4
2
2211
2
2
2
4
2
ˆ2
xrrDrcD
rd
DLW
A
A
xWxWxLrxA
A
d
p
pprep
newx
10th Copper Mountain Conference on Iterative Methods
22
Numerical Example
1.1 operator, derivative
)01.0,0( with
solution edgy :
matrix symmetric Toeplitz, Gaussian, :
pL
bNeeAxb
x
A
true
true
10th Copper Mountain Conference on Iterative Methods
23
Numerical Example
truex b
MGMxLSQRx
10th Copper Mountain Conference on Iterative Methods
24
Conclusions and Future work
General Cases – Non Structured Matrices
Parameter Selection Adaptive p-norm Extension to 2D, 3D