Space-Frequency Atoms

FREQUENCY
SPACE SPACE
FREQUENCY
SPACE
FREQUENCY FREQUENCY
SPACE
Figure 1: Space-frequency atoms.
Windowed Fourier Transform
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
0 100 200 300 400 500 600 700 800 900 1000
line 1
Figure 2: A Gabor function.
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
0 100 200 300 400 500 600 700 800 900 1000
line 1
Figure 3: A second Gabor function.
Windowed Fourier Transform (contd.)
Analysis
F(u, b) = f , w(x b)e
j2ux

=
Z

f (x)w(x b)e
j2ux
dx
Synthesis
f (x) =
Z

F(u, b)w(x b)e

j2ux
du db
What is a Wavelet?
All basis functions (daughter wavelets) are gen-
erated by translation and dilation of a mother
wavelet:

a,b
(x) =
1

x b
a

when a < 1 it shrinks the wavelet. The

a fac-
tor keeps the norm constant:

x b
a

x b
a

2
dx
=

a|| f (x)||.
What is a Wavelet? (contd.)
The mother wavelet, , must satisfy the admis-
sibility criterion:
C

=
Z

(s)|
2
|s|
ds <
where

is the Fourier transform of . This
means that:
|

(s)|
2
decays faster than 1/|s|

(0) = 0.
What is a Wavelet? (contd.)
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
0 100 200 300 400 500 600 700 800 900 1000
line 1
Figure 4: A Morlet wavelet.
-0.4
-0.2
0
0.2
0.4
0.6
0.8
0 100 200 300 400 500 600 700 800 900 1000
line 1
Figure 5: A second Morlet wavelet.
Vanishing Moments
The n-th moment of is dened to be:
M
n
{} =
Z

t
n
(t)dt.
If M
0
{} = 0 then has one vanishing mo-
ment.
Because
M
0
{} =
Z

(x)dx =

(0) = 0
all wavelets have at least one vanishing mo-
ment.
If M
0
{} = M
1
{} = 0, then has two
vanishing moments, etc.
Vanishing Moments (contd.)
If has one vanishing moment, then
a
0
,
a,b
= 0.
If has two vanishing moments, then
a
1
x +a
0
,
a,b
= 0.
If has n vanishing moments, then
a
n1
x
n1
+ +a
1
x +a
0
,
a,b
= 0,
i.e., the daughter wavelets are orthogonal to
any polynomial of degree less than n.
Vanishing moments are the reason why smooth
signals have sparse representations in wavelet
bases.
Three Kinds of Wavelet Transform:
Continuous wavelet transform
analysis synthesis input output
discrete
continuous
Wavelet series transform
analysis synthesis input output
discrete
continuous
Discrete wavelet transform
analysis synthesis input output
discrete
continuous
Continuous Wavelet Transform
Analysis
F(a, b) = f ,
a,b

=
Z

f (x)
a,b
(x) dx
Synthesis
f (x) =
1
C

F(a, b)
a,b
(x) db
da
a
2
where
C

=
Z

(s)|
2
|s|
ds
Two Dimensional Continuous Wavelet Transform
Analysis
F(a, b
x
, b
y
) = f ,
a,b
x
,b
y

=
Z

f (x, y)
a,b
x
,b
y
(x, y) dx dy
Synthesis
f (x, y) =
1
C

F(a, b
x
, b
y
)
a,b
x
,b
y
(x, y) db
x
db
y
da
a
3
where

a,b
x
,b
y
(x, y) =
1
|a|

x b
x
a
,
y b
y
a

and
C

=
Z

(u, v)|
2

|u|
2
+|v|
2
dudv
Wavelet Transform as Convolution
Recall that the relationship between daughter
wavelet
a,b
and mother wavelet involves both
translation and dilation:

a,b
(x) =
1

x b
a

.
Lets dene a function
a
to represent a daugh-
ter which is dilated by a factor a but is not trans-
lated:

a
(x b) =
a,b
(x) =
1

x b
a

and a function
a
(x) to represent a reected
and conjugated instance of
a
:

a
(x) =
a
(x).
Wavelet Transform as Convolution (contd.)
Using
a
and
a
the forward and inverse con-
tinuous wavelet transforms can be expressed as
follows:
Analysis
F(a, b) = f ,
a,b

=
Z

f (x)
a
(x b) dx
=
Z

f (x)
a
(bx) dx
= {f
a
}(b)
Synthesis
f (x) =
1
C

f ,
a,b

a,b
(x) db
da
a
2
=
1
C

{f
a
}(b)
a
(x b) db
da
a
2
=
1
C

{f
a

a
}(x)
da
a
2
Wavelet Series Transform
Is it possible to replace the integrals over a and
b in the synthesis formula with sums? Can we
represent any f in a Hilbert space, H , using a
discrete set, S, of wavelet coefcients? If for all
f H there exist A > 0 and B < such that
A|| f ||
2

(a,b)S
| f ,
a,b
|
2
B|| f ||
2
then
a,b
for (a, b) S form a frame for H .
Furthermore, there exists a set of functions

a,b
for (a, b) S which form a dual frame for H :
1
B
|| f ||
2

(a,b)S
| f ,

a,b
|
2

1
A
|| f ||
2
.
Wavelet Series Transform (contd.)
The wavelets,
a,b
, are used for analysis:
f ,
a,b
=
Z

f (x)
a,b
(x) dx
and the wavelets,

a,b
, are used for synthesis:
f (x) =

(a,b)S
f ,
a,b

a,b
(x).
Self-inverting Wavelet Series
If A = B, then

(a,b)S
| f ,
a,b
|
2
= A|| f ||
2
and the
a,b
for (a, b) S form a tight-frame
for H , in which case
f (x) =
1
A

(a,b)S
f ,
a,b

a,b
(x).
Such frames are said to be self-inverting be-
cause

a,b
(x) =
1
A

a,b
(x).
Redundancy
Recall that for a tight-frame
A =

(a,b)S
| f ,
a,b
|
2
|| f ||
2
.
Assuming that |||| =1, then A provides a mea-
sure of the redundancy of the expansion, i.e.,
the degree of overcompleteness. If A = 1 there
is no redundancy, and the expansion is orthonor-
mal. How can one nd wavelet series trans-
forms with no redundancy?
Dyadic Sampling
A sampling pattern is dyadic if the daughter
wavelets are generated by dilating the mother
wavelet by 2
j
and translating it by k2
j
:

j,k
(x) =
1

2
j

x k2
j
2
j

Dyadic sampling is optimal because the space

variable is sampled at the Nyquist rate for any
given frequency.
Dyadic Sampling (contd.)
FREQUENCY
SPACE
Figure 6: Dyadic sampling pattern.