Convolution in the time domain Multiplication in the frequency … · 2019. 2. 28. · Convolution...

18
Convolution in the time domain Multiplication in the frequency domain

Transcript of Convolution in the time domain Multiplication in the frequency … · 2019. 2. 28. · Convolution...

Page 1: Convolution in the time domain Multiplication in the frequency … · 2019. 2. 28. · Convolution Multiplication. 4 Toeplitz Block Toeplitz Doubly infinite Doubly infinite Banded

Convolution in the time domain Multiplication in the frequency domain

Page 2: Convolution in the time domain Multiplication in the frequency … · 2019. 2. 28. · Convolution Multiplication. 4 Toeplitz Block Toeplitz Doubly infinite Doubly infinite Banded

Matrix-vector multiplicationConvolutionMultiplication

Page 3: Convolution in the time domain Multiplication in the frequency … · 2019. 2. 28. · Convolution Multiplication. 4 Toeplitz Block Toeplitz Doubly infinite Doubly infinite Banded
Page 4: Convolution in the time domain Multiplication in the frequency … · 2019. 2. 28. · Convolution Multiplication. 4 Toeplitz Block Toeplitz Doubly infinite Doubly infinite Banded

4

Toeplitz Block ToeplitzDoubly infinite Doubly infiniteBanded Banded

Polynomial A(θ) Matrix Polynomial A(θ)

A-1 (θ) = inverse of A(θ) (not banded?)BC (θ) = B(θ)C(θ) (banded)

Page 5: Convolution in the time domain Multiplication in the frequency … · 2019. 2. 28. · Convolution Multiplication. 4 Toeplitz Block Toeplitz Doubly infinite Doubly infinite Banded

5

Not banded Singly infinite / finite

Periodic A(θ) Circulants

Function theory

Bottcher-Silbermann Szego-Grenander

Grochenig (Wiener L1 Lemma)

Wiener-Hopf Gohberg-Semencul

Page 6: Convolution in the time domain Multiplication in the frequency … · 2019. 2. 28. · Convolution Multiplication. 4 Toeplitz Block Toeplitz Doubly infinite Doubly infinite Banded

Factorization A(θ) = U(θ) L(θ)

6

A = L U or A = L P U

Plemelj / G.D. Birkhoff / ...Gohberg / Kaashoek / Spitkovsky / ...Can we reach linear factors A = A1 A2 ... Ak?

Banded but not Toeplitz

“Time-varying filter”Symbol A(θ) varies from row to rowFactorization still possible?

Page 7: Convolution in the time domain Multiplication in the frequency … · 2019. 2. 28. · Convolution Multiplication. 4 Toeplitz Block Toeplitz Doubly infinite Doubly infinite Banded

7

A = L1 P1 U1 A = L1 σ2 L2

A = U1 σ1 U2 A = U2 P2 L2

Problem for a banded doubly infinite matrix: a11 is in the middle of A

Need elimination starting from - ∞ !!

A = L P U is still possible, even if A is not Toeplitz

a11 a1n

an1 ann[ ]

Page 8: Convolution in the time domain Multiplication in the frequency … · 2019. 2. 28. · Convolution Multiplication. 4 Toeplitz Block Toeplitz Doubly infinite Doubly infinite Banded
Page 9: Convolution in the time domain Multiplication in the frequency … · 2019. 2. 28. · Convolution Multiplication. 4 Toeplitz Block Toeplitz Doubly infinite Doubly infinite Banded

when

/ offset

Page 10: Convolution in the time domain Multiplication in the frequency … · 2019. 2. 28. · Convolution Multiplication. 4 Toeplitz Block Toeplitz Doubly infinite Doubly infinite Banded

Blocks in C are offset by w from blocks in B

Page 11: Convolution in the time domain Multiplication in the frequency … · 2019. 2. 28. · Convolution Multiplication. 4 Toeplitz Block Toeplitz Doubly infinite Doubly infinite Banded
Page 12: Convolution in the time domain Multiplication in the frequency … · 2019. 2. 28. · Convolution Multiplication. 4 Toeplitz Block Toeplitz Doubly infinite Doubly infinite Banded
Page 13: Convolution in the time domain Multiplication in the frequency … · 2019. 2. 28. · Convolution Multiplication. 4 Toeplitz Block Toeplitz Doubly infinite Doubly infinite Banded

Which is the main diagonal of a doubly infinite matrix?

13

Example: one-way shift S (lower diagonal of 1s)

Inverse = reverse shift ST (upper diagonal of 1s)

Index (S+) = dim (nullspace(S+)) - dim(nullspace((S+)T)) = 0 - 1

Main diagonal of shift matrix S is diagonal -1

Page 14: Convolution in the time domain Multiplication in the frequency … · 2019. 2. 28. · Convolution Multiplication. 4 Toeplitz Block Toeplitz Doubly infinite Doubly infinite Banded

14

Fredholm: Index not changed if an entry changes

Index(BC) = Index(B) + Index(C)

Index (S+) locates the main diagonal

Index (S+) = 0 when A is centered

Then inverse of upper is upper/ finite sections OK

Banded permutation: Index by counting 1’s (Lindner, GS)

Page 15: Convolution in the time domain Multiplication in the frequency … · 2019. 2. 28. · Convolution Multiplication. 4 Toeplitz Block Toeplitz Doubly infinite Doubly infinite Banded

Matrix entries now have 4 indices (not 2)Vector entries now have 2 indices (not 1)

Page 16: Convolution in the time domain Multiplication in the frequency … · 2019. 2. 28. · Convolution Multiplication. 4 Toeplitz Block Toeplitz Doubly infinite Doubly infinite Banded

A is a 2D convolution matrixcontains

The symbol is

for this example?when A is banded?

Page 17: Convolution in the time domain Multiplication in the frequency … · 2019. 2. 28. · Convolution Multiplication. 4 Toeplitz Block Toeplitz Doubly infinite Doubly infinite Banded

17

Page 18: Convolution in the time domain Multiplication in the frequency … · 2019. 2. 28. · Convolution Multiplication. 4 Toeplitz Block Toeplitz Doubly infinite Doubly infinite Banded

18