Admissibility condition on a wavelet

Continuous wavelet transform (CWT) of a function x(t) w.r.t. (t) is given by:

t b
WT {x; b, a} C x(t ) dt
a
where, a and b are dilation and translation parameters and C is the normalizing factor
such that ,
2
t b
C dt has unit energy
a

1
=> C
a
Hence the CWT is expressed as:

1 t b
W x(b, a ) x(t ) dt
a a

Theorem: The wavelet transform obeys a Parseval theorem like relationship.

da
W x(b, a)W y (b, a)db C x(t ) y (t )dt
a2
Proof:
First let us consider the single integral of LHS with w.r.t. b which can be transformed to
w-Domain as:
1 ^ ^
W x(b, a )W y (b, a )db W x(b, a ) W y (b, a )d (1)
2
which is defined as
^
^ 1 t b
W x(b, a) x(t ) dt
a a
ie. Transformation w.r.t b:

1 j b t b
e { x(t ) dt}db
a a
Exchanging db and dt;

1 t b j b
x(t ){ e db}dt (2)
a a
t b
Substitute ;t a b ; db ad in
a
t b j b
e db
a
j (t a )
to get a e

a ej (t a )

ae j t eja
^
j t
ae a (3)

Substituting (3) back in (2)

^
^ 1 j t
W x(b, a ) 12
x(t ){ a e a }dt
a

1 ^
2 j t
a { x (t )e dt} a

1 ^ ^
2
a x( ) a (4)

Substituting (4) back in (1)

^ ^
1 1
2
^ 1
2
^
W x ( b , a )W y ( b , a )db {a x( ) a }{ a y( ) a }d
2

^ 2
1 ^ ^
a x( ) y( ) a d
2
(5)
 ^ 2
da 1 ^ ^ da
LH S W x ( b , a )W y ( b , a )db a x( ) y( ) a d
a2 2 a2

^ 2
1 ^ ^ da
x ( ) y ( ){ a }d
2 a

da d
Substituting aw ; we get and
a
^ 2
da 1 ^ ^ d
W x (b , a )W y (b , a )db 2 { x ( ) y ( )d }{ }
a 2
d
for the RHS of the above integral to converge | ( ) |2 must converge say to a
0

positive finite constant C

1
=> LHS C x( ) y ( )d
2

C x(t ) y (t )dt (RHS) (Using Parseval s Theorem)

Hence Proved

Example: Checking admissibility condition for Haar MRA:

sin 2
Expression for Haar wavelet: | ( )| 4
4

d sin 4
Applying admissibility condition: | ( ) | 2 4 d
2
0 0
4

The integral can be expressed as:

0 0
 2
2
sin 2 d d
4 d sin
For integral 4
0 0 0
16
4
which converges as -> 0.
2
2
sin
4 d 16
Also the integral < 3 ( sin < 1) therefore as ->
4
the integral converges.
Hence Haar wavelet satisfies admissibility condition.

Inverse Continuous Wavelet Transform (ICWT)

From Parseval theorem, to get inverse wavelet transform choose y(t) to be an unit area
narrow pulse around t=t0 (or as impulse)
t0 b dbda
x (t 0 ) K W x (b , a )
a a2
d
where again, K = | ( ) |2 ; a constant

d
on the condition that | ( ) |2 converges ie. ADMISSIBILITY CONDITION.

It is because of admissibility, that the CWT is invertible.

d
If spectrum of (t ) i.e. ( ) does not 0 as 0, then | ( ) |2 diverges.

Therefore ( ) must decay atleast as fast as , which says if . is a band pass

d
function then | ( ) |2 converges.

The transform we have seen so far are redundant in nature. So now we go for
discretization of scale and translation parameter.
1.1 Discretization of scale parameter a:
Consider
d
| ( ) |2

Put e

| ( e )2 d This quantity is Fourier transform of some autocorrelation

sequence.

For 0; e e e 0
i.e. for uniform movement on there is exponential

movement in . This is logarithmic discretization. This means that the no. of points of
dicretization between a = 1 and a =10 should be same as beween a = 10 and a = 100
where a is the scaling parameter.

m
Now choosing a = a0 ( > a 0 > 0)
m 1

dw a0 dw
we get | ( w ) |2 | ( w ) |2
0
w m m w
a0
m dw d
now substituting w a 0
;
w
we get:

a0 d
m
| (a 0
) |2
m 1

taking the summation sign inside the integral we get :

a0 d
m
{ | ( a 0 ) |2 }
1 m

for this integral to converge;

 m
0 A | ( a 0 ) |2 B
m

ie. strictly between 2 positive constants.

The bounded quantity is called the Sum of Dilated Spectra given by:

SDS ( , a0 )( ) | (a0m ) |2
m

and the condition of bounding is referred as Frame Property (it is also a check for
possibility of inversion).

Wavelet Transform on function x(t) is given by

1 t b
W x(b, a ) x(t ) dt
a a

1 t
Define g (t )
a a

1 t b
g (b t )
a a

x(t ) g (b t )dt

This is convolution of x(t) with g(.-t). So we can interpret wavelet transform as a linear
filtering operation.
Frequency response of the LSI filter
jwt
g ( w) g (t )e dt

1 t jwt t
e dt Put
a a a
1 jw( a )
( )e ad
a

j ( aw)
a ( )e d

g ( w) a (aw)
That is we are filtering x(t).
(t ) is band-pass function and a (aw) is a band-pass filter with center frequency
W /a.
In this way the signal x(t) is filtered by different filters with different scale parameter.

Similarly it can be shown that inverse wavelet transform is also a filtering operation.
Freq. domain output of mth filter at analysis side is

x( w) (a0 m w) where a0>1

Output of mth synthesis filter

x( w) (a0 m w) (a 0 m w)

x( w) | (a0 m w) | 2

Output of the synthesis filter bank.

x( w) | (a0 m w) | 2
m

x( w) | (a0 m w) | 2
m
Therefore for perfect reconstruction (i.e. synthesis filter bank output = a multiple of the
analysis filter bank input).
i.e. | (a0 m w) | 2 C1 for all w.
m

The term | (a0 m w) | 2 , sum of dilated spectra SDS should be a constant value. If 0 <
m

A | (a0 m w) | 2 B< is true we can use different analysis and synthesis filter.
m
^
~ ~
Also we can scale ( w ) such that SDS ( , a0 )( w ) = 1 which is given by

^
~ ( w)
( w)
SDS ( , a0 )( w)
to get normalized bank of filters.
This document was created with Win2PDF available at http://www.daneprairie.com.
The unregistered version of Win2PDF is for evaluation or non-commercial use only.