Course 18.327 and 1.

130

Wavelets and Filter Banks

Modulation and Polyphase

Representations:

Noble Identities;

Block Toeplitz Matrices

and Block z-transforms;

Polyphase Examples

Modulation Matrix
Matrix form of PR conditions:
[F0 (z) F1 (z)] H0(z) H0(-z)
= [ 2z ?? 0 ]
H1(z) H1(-z)

&'(

Modulation matrix, Hm(z)

So
[ F0(z) F1(z)] = [2z ?? 0] H
Hm1(z) = 1
H1(-z) -H0(-z)
-H1(z) H0(z)

1(z)

H0(z) H1(-z) - H0 (-z) H1 (z) (must be non-zero)

2

2z-?? H1(-z)
2z-??

F1(z) =

H0(-z)

Suppose we choose
Then

F1(z) = -H0(-z)

Require these
to be FIR

= 2z - ?

*+,

F0(z) = H1(-z)

*+,

F0(z) =

Synthesis modulation matrix:

Complete the second row of matrix PR conditions
by replacing z with z:
F0(z) F1(z)

H0(z)

z-??

H0(-z)

= 2

F (-z) F (-z)
&'(
0

H1(z)

H1(-z)

(-z) - ?

Synthesis
modulation
matrix, Fm(z)
Note the transpose convention in Fm(z).
4

Noble Identities
1. Consider
x [n]

u[n]

H(z2)

U(z) = H(z2)X(z)
Y(z) = {U(z ) + U(-z )}

y[n]

(downsampling)

= { H (z) X (z ) + H(z) X (-z )}

= H(z) {X (z ) + X (-z )} can downsample
first
First Noble identity:
x [n]
y[n] x[n]
y[n]

H(z)
2
H(z2)
2
5

2. Consider
x[n]

H(z)

u[n]

U(z) = H(z) X(z)

Y(z) = U(z2)
= H(z2) X(z2)

y[n]

(upsampling)
can upsample first

Second Noble Identity:

x[n]
y[n]
x[n]

2
H(z2)

H(z)

y[n]

Derivation of Polyphase Form

1. Filtering and downsampling:
x[n]
y[n]
H(z)
2

H(z) = Heven(z2) + z-1 Hodd(z2); heven[n] = h[2n]

hodd[n] = h[2n+1]
Heven(z2)

x[n]

z-1

y[n]

Hodd(z2)

Heven(z2)

x[n]

y[n]
+

z-1

Hodd(z2)

xeven[n]

x[n]

Heven(z)
+

z-1

xodd[n-1]

Hodd(z)

y[n]
Polyphase
Form
8

2. Upsampling and filtering

x[n]

y[n]

F(z)

F(z) = Feven(z2) + z-1 Fodd (z2)

x[n]

Feven(z2)

Fodd(z2)

y[n]

z-1

2
x[n]

Feven(z2)
+

x[n]

Feven(z)

Fodd(z2)

yeven[n]

z-1

2
+

Fodd(z)

yodd[n]

y[n]

y[n]

z-1
Polyphase
Form
10

Polyphase Matrix
Consider the matrix corresponding to the analysis
filter bank in interleaved form. This is a block
Toeplitz matrix:
4

Hb=

3 h0[3] h0[2] h0[1] h0[0]

3 h1[3] h1[2] h1[1] h1[0]
0
0

0
0

h0[3] h0[2]
h1[3] h1[2]

0
0

h0[1] h0[0]
h1[1] h1[0]

3
3
3
3

3
3

0
0

4-tap Example
11

Taking block z-transform we get:

Hp(z) =

h0[0] h0[1]
h1[0] h1[1]

+ z-1

h0[0] + z-1 h0[2]

h1[0] + z-1 h1[2]

H0,even (z)
H1,even (z)

h0[2] h0[3]
h1[2] h1[3]

h0[1] + z-1 h0[3]

h1[1] + z-1 h1[3]

H0,odd (z)
H1,odd (z)

This is the polyphase matrix for a 2-channel filter bank.

12

Similarly, for the synthesis filter bank:

Fb =

f0[0] f1[0]
f0[1] f1[1]
3 f0[2] f1[2]
f0[3] f1[3]

0
0

0
0
3

f0[0] f1[0]
f0[1] f1[1]

0
0

0 f0[2] f1[2]
0 f0[3] f1[3]

4
13

Fp(z) =

f0[0] f1[0]
f0[1] f1[1]

+ z-1

F0,even [z] F1,even [z]

F0, odd [z] F1, odd [z]

f0[2] f1[2]
f0[3] f1[3]

Note transpose
convention for
synthesis
polyphase matrix

Perfect reconstruction condition in polyphase domain:

Fp(z) Hp(z) = I (centered form)
This means that Hp(z) must be invertible for all z on the
unit circle, i.e.
det Hp(eiww) 0 for all frequencies w.
14

 Given that the analysis filters are FIR, the

requirement for the synthesis filters to be also
FIR is:
det Hp(z) = z-?? (simple delay)
because Hp-1(z) must be a polynomial.
Condition for orthogonality: Fp(z) is the transpose
of Hp(z), i.e.
HpT(z-1) Hp(z) = I
i.e. Hp(z) should be paraunitary.

15

&)')(

Relationship between Modulation

and Polyphase Matrices
h0,even[n] = h0[2n]

H0(z) = H0,even(z2) + z-1 H0,odd(z2) ;

h0,odd[n] = h0[2n+1]

H1,even(z2)

z-1

H1,odd(z2)

+
H1(z) =
Two more equations by replacing z with -z.
So in matrix form:
H0(z) H0(-z)
H0,even(z2) H0,odd(z2) 1
1
=
2
2
-1
H1,even(z ) H1,odd(z ) z -z-1
H1(z) H1(-z)

&'(

&'(

Hm(z)
Hp(z2)
Modulation matrix Polyphase matrix
16

But
1
1
z-1 -z-1

1
z-1

&'(

1 1
1 -1
&'(

F2
D2(z)
Delay Matrix 2-point DFT Matrix

FN =

-1

FN =

1
1
1
.
.
.
1
1
N

1 1 1
w w2 w N-1
w2 w4 w 2(N-1
; w=e
.
.
.
.
.
.
2
N-1
w w2(N-1) w(N-1)

2p
p
N

N-point DFT
Matrix

FN
Complex conjugate: replace w with w = e-

2p
p
N

17

So, in general

Hm(z) F-1
N

Hp(zN) DN(z)

N = # of channels in filterbank
(N = 2 in our example)

18

Polyphase Matrix
Example: Daubechies 4-tap filter

h0[0] =

1+
3
3 + 3
h0[1] =
4 2
4
2

h0[2] =

3 -
3
1- 3
h0[3] =
4
2
4
2

1
{(1 + 3 ) + (3 + 3 ) z-1 + (3 - 3) z-2 + (1 - 3) z-3}
4
2
1
{(1 - 3) (3 - 3) z-1 + (3 + 3)z-2 (1 + 3)z-3}
H1(z) =
4
2

H0(z) =

19

Time domain:
h0[0]2 + h0[1]2 + h0[2]2 + h0[3]2 =

h0[0] h0[2] + h0[1] h0[3] =

1
32

+ 2
3) + (12 + 6 3) +
(12 6 3) + (4 2 3)}
=1
{(2
3) + (-2
3)}
1 {(4
32

=0
i.e. filter is orthogonal to its double shifts

20

10

Polyphase Domain:
1
{(1 + 3) + (3 - 3) z-1}
H0,even(z) =
4 2

H0,odd(z)

1
{(3 + 3) + (1 - 3) z-1}
4 2

H1,even(z)

1
{(1 - 3) + (3 + 3) z-1}
4 2

H1,odd(z)

1
{ - (3 - 3) (1 + 3) z-1}
4 2

1
Hp(z) =
4
2

1 + 3

3 + 3
+

1 - 3

-(3 - 3)

3 - 3

1
4
2

&'(

1 - 3
z-1

3 + 3

-(1 + 3

&'(
B

21

Hp(z) = A + B z-1
HpT(z-1) Hp(z) = (AT + BT z)(A + Bz-1)
= (ATA + BTB) + ATBz-1 + BTAz
ATA =

1 1 + 3 1 - 3
1
2
4
2 3 + 3 - (3 - 3) 4
1
32

1 + 3
1 - 3

3 + 3
-(3-
3)

(4 + 2
3) + (4 - 2
3) (6 + 4
3) - (6 - 4
3)
(6 + 4
3) - (6 - 4
3) (12 + 6
3) + (12 - 6
3)

3/4

3/4

22

11

BTB =

1
3 - 3 3 + 3 1 3 - 3
1 - 3
1

3
-(1
+

3)
4 2 3 + 3 - (1 + 3)
4 2
1

(12 6
3) + (12 + 6
3) (6 - 4
3) - (6 + 4
3)
(6
4

3)

(6
+
4

3)
(4
2

3)
+
(4
+
2
3)
32

- 3/4

- 3/4

ATA + BTB = I

23

ATB =

1
4 2

1
32

BTA =

1 + 3 1 -
3 1 3 - 3 1 -
3
3 + 3 -(3-
3) 4 2 3 + 3 -(1+
3)
(2 3) + (-2
3)
(6) (6)

(-2) (-2)
(-2 3) + (2 3)

(ATB)T = 0

So
HpT(z-1) Hp(z) = I

i.e. Hp(z) is a Paraunitary Matrix

24

12

Modulation domain:
1

(-z3 + 9z + 16 + 9z-1 z-3)

16
1

H0(-z) H0(-z-1) = P(-z) =

(z3 9z + 16 9z-1 + z-3)

16

H0(z) H0(z-1) = P(z) =

So

H0(z) H0(z-1) + H0(-z) H0(-z-1) = 2

i.e.

|H0(w)|
w ^2 + |H0(w
w + p)|
p ^2 = 2

25

Magnitude Response of Daubechies 4-tap filter.

Ma g nitud e re s po ns e o f Da u be c h ie s 4 -ta p filte r.

2. 5

Fre que nc y re s pons e pha s e

1. 5

0. 5

0
-1

-0. 8

-0. 6

-0.4
-0.2
0
0.2
0.4
Ang ula r fre que nc y (no rm a liz e d b y p)

0 .6

0 .8

26

13

Phase response of Daubechies 4-tap filter.

P ha s e re s pons e of Da ube c hie s 4-ta p filte r.

Fre que ncy re s pons e phas e

-1

-2

-3

-4
-1

-0. 8

-0. 6

-0. 4
-0. 2
0
0. 2
0.4
Angula r fre qu e nc y (norm a liz e d by p)

0.6

0.8

27

14