The Abstract Wavelet Transform

Jaime Navarro
Universidad Autonoma Metropolitana
P. O. box 16-306, Mexico City, 02000 Mexico
jnfu@correo.azc.uam.mx
Abstract
Given an abstract locally compact topological group, the continuous wavelet
transform is dened so that the reconstruction formula is used to prove that the
image of the wavelet transform is a reproducing kernel Hilbert space and it is shown
also a generalization of the sampling theorem.
Key words and phrases: wavelet transform, admissible function, inversion formula
2000 AMS Mathematics Subject Classication 42A38, 42C40
1 Introduction
In one dimension, in order to analyze a signal f on the phase space,
we use

f, the Fourier transform of f. In this case,

f(w) =
_
R
e
2iwx
f(x)dx depends only of the frequency w. That
is,

f gives information of the frequency w but it does not give the
changes of w through the time t. To solve this problem, a function
g(x t) depending of t, called a window is introduced to get the
windowed Fourier transform :
_
V
g
f
_
(w, t) =
_
R
e
2iwx
f(x) g(x t) dx (1)
which is a time-frequency localization method. The function g is
chosen so that support of g is compact and g is
1
2
of class C
n
for some n. The windowed Fourier transform is then
used to determine high and low-frequencies of the signal f in a
specic time t.
The continuous wavelet transform of the signal f is also a time-
frequency localization method and it is considered as an alternative
of the windowed Fourier transform.
For a given a in R 0, b in R and a basic function h called
the mother wavelet,
_
L
h
f
_
(a, b) =
_
R
f(x)
1
_
[a[
h
_
x b
a
_
dx (2)
is dened as the continuous wavelet transform of f with
respect to h.
In this case, a represents the frequency , b the time, and h is taken
so that h is in C

0
(R) with

h(0) = 0.
If we introduce the translation, dilation, and modulation
operators dened respectively as :
_
T
t
h
_
(x) = h(x t),
_
J
a
h
_
(x) =
1

|a|
h(
x
a
), and
_
E
w
h
_
(x) = e
2iwx
h(x), then for f, g, h in
L
2
(R), the windowed Fourier transform and the continuous wavelet
transform, both can be written as inner products :
_
V
g
f
_
(w, t) =

f, E
w
T
t
g
_
, and
_
L
h
f
_
(a, b) =

f, T
b
J
a
h
_
. (3)
The dierence between these two transforms is with respect to the
functions E
w
T
t
g and T
b
J
a
h. For the windowed Fourier transform,
the width of E
w
T
t
g is the same as for g translated to the proper
time of localization and lled it out with high-frequencies. The size
3
of the window E
w
T
t
g is the same regardless of the value of the
frequency w. On the other hand, the functions T
b
J
a
h have time
width depending on its frequency a. If a > 1 and b = 0 , T
0
J
a
h
detects low-frequencies, and for 0 < a < 1 , T
0
J
a
h determines high-
frequencies. As a consequence, the wavelet transform determines
better than the windowed Fourier transform the localization of high
and low-frequencies for a specic time t, [1].
2 Notations and denitions
Let us begin by dening a homomorphism for an abelian locally com-
pact topological group. Consider two locally compact topological
groups A and B where A is abelian, and consider a homomorphism
from B into the group of all automorphisms of A such that the
map (a, b)
b
(a) is continuous on A B to A. That is, for each
b in B, the map
b
: A A, a
b
(a) is a homeomorphism.
Denition 1 Dene G as the product of A and B. That is, consider
G = (a, b)[a A, b B, and for (a
1
, b
1
), (a
2
, b
2
) in G dene
(a
1
, b
1
)(a
2
, b
2
) = (a
1

b
1
(a
2
), b
1
b
2
) (4)
With this product, G becomes a group, where e = (e
1
, e
2
) is the
identity (e
1
is de identity in A and e
2
is the identity in B) and where
(a, b)
1
= (
b
1(a
1
), b
1
) is the inverse. Note also that G = A B
is a locally compact topological group. Then we will denote by
d
G
(a, b) the left Haar measure on G, the left Haar measure on A
by d
A
(a) and the left Haar measure in B by d
B
(b).
Denition 2 Given a group G and a set E, an action of G on E is
a map (s, x) sx of GE E such that
4
1) ex = x for any x in E and where e is the identity in G
2) s(tx) = (st)x for any x in E and where s, t are in G.
Then we have the following Lemma.
Lemma 1 The function : G A A given by (a, b) x = a
b
(x)
is an action of G on A where (a, b) G and x A.
Denition 3 Let G be a locally compact topological group. The
support of the function f : G C denoted by supp(f) is dened
as the closure of x G[f(x) ,= 0, and C
0
(G) is dened as the set
of continuous functions f : G C such that supp(f) is compact.
Denition 4 For 1 p < and for a complex valued function
dened on the locally compact topological group A, dene
L
p
(A) =
_
h : A C

_
A
[h(x)[
p
d
A
(x) <
_
(5)
where d
A
(x) is the left Haar measure on A.
Then we have the following result for h L
1
(A) (See [9])
_
A
h
_
(a, b)
1
x
_
d
A
(x) = (a, b)
_
A
h(x)d
A
(x) (6)
Note So, from now on consider : G (0, ) satisfying (6) for
h L
1
(A).
3 Unitary operators
Denition 5 For h L
2
(A) dene the following operators
5
(J
a
h)(x) =
1
_
(a, e
2
)
h((a, e
2
)
1
x) where (a, e
2
) G, x A, a A
(T
b
h)(x) =
1
_
(e
1
, b)
h((e
1
, b)
1
x) where (e
1
, b) G, x A, b B
Lemma 2 For the operators J
a
and T
b
,
1) J
a
1
J
a
2
= J
a
1
a
2
where a
1
, a
2
A
2) T
b
1
T
b
2
= T
b
1
b
2
where b
1
, b
2
B
3) T
b
J
a
= J

b
(a)
T
b
and J
a
T
b
= T
b
J

b
1
(a)
where a A and b B
Lemma 3 The operators J
a
and T
b
are unitary operators
Proof Let h be in L
2
(A). Then by (6),
1)
[[J
a
h[[
2
=
_
A
[(J
a
h)(x)[
2
d
A
(x) =
_
A
(J
a
h)(x)(J
a
h)(x)d
A
(x)
=
_
A
1
_
(a, e
2
)
h((a, e
2
)
1
x)
1
_
(a, e
2
)
h((a, e
2
)
1
x)d
A
(x)
=
1
(a, e
2
)
_
A
(hh)((a, e
2
)
1
x)d
A
(x)
=
1
(a, e
2
)
(a, e
2
)
_
A
(hh)(x)d
A
(x) =
_
A
[h(x)[
2
d
A
(x) = [[h[[
2
6
2)
[[T
b
h[[
2
=
_
A
[(T
b
h)(x)[
2
d
A
(x) =
_
A
(T
b
h)(x)(T
b
h)(x)d
A
(x)
=
_
A
1
_
(e
1
, b)
h((e
1
, b)
1
x)
1
_
(e
1
, b)
h((e
1
, b)
1
x)
=
1
(e
1
, b)
_
A
(hh)((e
1
, b)
1
x)d
A
(x)
=
1
(e
1
, b)
(e
1
, b)
_
A
(hh)(x)d
A
(x)
=
_
A
[h(x)[
2
d
A
(x) = [[h[[
2
Moreover, since J

a
= J
1
a
= J
a
1 and T

b
= T
1
b
= T
b
1, it follows
that both, J
a
and T
b
are unitary.
This proves Lemma 3.
4 Fourier transform
Denition 6 Let G be a locally compact topological abelian group,
and let T = z C[[z[ = 1. We say that the function : G T
is a character on G if is a continuous homomorphism
Denition 7 Given a locally compact topological abelian group G,
we dene the dual group of G as

G = : G T[ is a character
In this case we denote (g) = g, ) where g G and

G
Note that

G is clearly an Abelian group under pointwise multi-
plication (
1

2
)(g) =
1
(g)
2
(g). Its identity element is the constant
function 1 and the inverse element is
1
(g) = (g) = (g
1
).
7
The dual group of a locally compact topological abelian group is
used to dene an abstract version of the Fourier transform.
Denition 8 Given h L
1
(G), the Fourier transform is the
function

h :

G C dened by

h() =
_
G
h(g)(g)d
G
(g) (7)
where the integral is relative to the left Haar measure on G.
Denition 9 For a function h L
1
(

G), the inverse Fourier
transform of h is the function

h : G C dened as

h(g) =
_

G
h()(g)d

G
() (8)
where d

G
() is the left Haar measure on

G
Note. For h L
1
(G) and

h L
1
(

G) we have
h(g) =
_

h()(g)d

G
()
Lemma 4 For h C
0
(A) we have
1)

J
a
h() =
_
(a, e
2
) (a)

h()
2)

T
b
h() =
_
(e
1
, b)

h(
b
)
where

A, a A and b B.
Proof
1) Note that since

(J
a
h)() =
_
A
(J
a
h)(x)(x)d
A
(x)
=
_
A
1
_
(a, e
2
)
h((a, e
2
)
1
x)(x)d
A
(x)
8
and
(x) = (aa
1

e
2
(x)) = (a)(a
1

e
2
(x))
= (a)((a
1
, e
2
) x) = (a)((a, e
2
)
1
x)
it follows from (6) that

J
a
h() =
_
A
1
_
(a, e
2
)
h((a, e
2
)
1
x)(a)((a, e
2
)
1
x)d
A
(x)
=
1
_
(a, e
2
)
(a)
_
A
(h)((a, e
2
)
1
x)d
A
(x)
=
1
_
(a, e
2
)
(a)(a, e
2
)
_
A
(h)(x)d
A
(x)
=
_
(a, e
2
) (a)
_
A
h(x)(x)d
A
(x)
=
_
(a, e
2
) (a)

h()
2) Similarly, since

(T
b
h)() =
_
A
(T
b
h)(x)(x)d
A
(x) =
=
_
A
1
_
(e
1
, b)
h((e
1
, b)
1
x)(x)d
A
(x)
and
(x) = [
bb
1(x)] = [
b
(
b
1(x))] = [
b
(e
1

b
1(x))]
=
_

b
((e
1
, b
1
) x)

=
_

b
((e
1
, b)
1
x)

= (
b
)((e
1
, b)
1
x)
9
It follows from (6) that

T
b
h() =
_
A
1
_
(e
1
, b)
h((e
1
, b)
1
x)(
b
)((e
1
, b)
1
x)d
A
(x)
=
1
_
(e
1
, b)
_
A
(h
b
)((e
1
, b)
1
x)d
A
(x)
=
1
_
(e
1
, b)
(e
1
, b)
_
A
(h
b
)(x)d
A
(x)
=
_
(e
1
, b)
_
A
h(x)(
b
)(x)d
A
(x)
=
_
(e
1
, b)

h(
b
)
This proves Lemma 4.
Corollary 1 Let h be in L
1
(A). Then

J
a
T
b
h() =
_
(a, b) (a)

h(
b
)
Corollary 2 Let h be in L
1
(A). Then
1)

J
a
h() = J
a
1

h()
2)

T
b
h() = T
b
1

h()
3)

J
a
T
b
h() = J
a
1T
b
1

h()
5 Unitary representation
Denition 10 For (a, b) in G = AB, dene the two - parameter
family of operators U(a, b) = J
a
T
b
. Note that U(a, b) acts on the
10
Hilbert space L
2
(A) by:
(U(a, b)h)(x) = (J
a
T
b
h)(x) = (J
a
(T
b
h))(x)
=
1
_
(a, e
2
)
(T
b
h)((a, e
2
)
1
x)
=
1
_
(a, e
2
)
1
_
(e
1
, b)
h((e
1
, b)
1
(a, e
2
)
1
x)
=
1
_
((a, e
2
)(e
1
, b))
h
_
((a, e
2
)(e
1
, b))
1
x

=
1
_
(a, b)
h((a, b)
1
x)
Lemma 5 . U(a, b) = J
a
T
b
is a unitary representation of G acting
on the Hilbert space L
2
(A).
Proof
Note that since the operators J
a
and T
b
are unitary operators
(Lemma 3), it follows that U(a, b) is unitary.
Now, let us prove that U(a, b) is a representation of G acting on
L
2
(A).
On one hand, from Lemma 2,
U((a
1
, b
1
)(a
2
, b
2
)) = U(a
1

b
1
(a
2
), b
1
b
2
) = J
a
1

b
1
(a
2
)
T
b
1
b
2
= J
a
1
J

b
1
(a
2
)
T
b
1
T
b
2
= J
a
1
T
b
1
J
a
2
T
b
2
= U(a
1
, b
1
)U(a
2
, b
2
)
On the other hand, since U(e
1
, e
2
) = J
e
1
T
e
2
= I, where I is the
identity operator, it follows that U(a, b) is a representation of G.
This proves Lemma 5
Lemma 6 The left Haar measure on G = A B is
d(a, b) =
1
(a, b)
d
A
(a)d
B
(b)
11
Proof
Let h be in L
1
(G), then by (6)
_
G
h
_
(a
0
, b
0
)
1
(a, b)

d
A
(a)d
B
(b) = (a
0
, b
0
)
_
G
h(a, b)d
A
(a)d
B
(b)
Now, replacing h by
h

we have
_
G
h
_
(a
0
, b
0
)
1
(a, b)

[(a
0
, b
0
)
1
(a, b)]
d
A
(a)d
B
(b) = (a
0
, b
0
)
_
G
h(a, b)
(a, b)
d
A
(a)d
B
(b)
Then
_
G
h
_
(a
0
, b
0
)
1
(a, b)

1
(a, b)
d
A
(a)d
B
(b) =
_
G
h(a, b)
1
(a, b)
d
A
(a)d
B
(b)
That is
_
G
h
_
(a
0
, b
0
)
1
(a, b)

d(a, b) =
_
G
h(a, b)d(a, b)
This shows that d(a, b) is a left Haar measure on G.
This proves Lemma 6.
6 Admissibility condition
Denition 11 A function h in L
2
(A) is said to be admissible if
_
G
[ h, U(a, b)h) [
2
d(a, b) <
Lemma 7 Let h be in L
1
(A) L
2
(A). If (B) < , then
C
h

_
B
[

h(
b
)[
2
d
B
(b)
is uniformly bounded for

A
12
Proof Note that since

h(
b
) =
_
A
h(x)(
b
)(x)d
A
(x),
it follows that
[

h(
b
)[
_
A
[h(x)[ [(
b
(x))[d
A
(x) = [[h[[
1
Hence, C
h
[[h[[
2
1
(B) < . This proves Lemma 7.
Lemma 8 Let h be in L
1
(A) L
2
(A) if
0 < C
h

_
B
[

h(
b
)[
2
d
B
(b)
is uniformly bounded for

A, then h is admissible.
7 The continuous wavelet transform
Denition 12 Given (a, b) in G = AB and h admissible in L
2
(A),
the continuous wavelet transform with respect to h is dened as the
linear operator
L
h
(a, b) : L
2
(A, d
A
) L
2
(G, d(a, b))
such that for any f in L
2
(A) we have
(L
h
f)(a, b) = f, U(a, b)h)
L
2
(A)
That is,
(L
h
f)(a, b) =
_
A
f(x)[U(a, b)h] (x)d
A
(x)
=
_
A
f(x)
1
_
(a, b)
h((a, b)
1
x)d
A
(x)
13
Note In order to get back the function f from the continuous
wavelet transform (L
h
f)(a, b), we will apply the Grossmann-Morlet-
Paul theorem [2], where the hypotheses for the representation U(a, b)
are: unitary, irreducible and strongly continuous. In our case, our
representation is unitary, so the following lemmas will show that
U(a, b) = J
a
T
b
is irreducible and strongly continuous.
Lemma 9 The representation U(a, b) of the group G = AB acting
on L
2
(A) is irreducible.
Proof Suppose h L
2
(A) 0 and suppose f L
2
(A) is such
that f, U(a, b)h) = 0 for all (a, b) in G. To show the representation
U(a, b) is irreducible we will show that f = 0 in L
2
(A).
Under the assumptions we have
_
G
[ f, U(a, b)h) [
2
d(a, b) = 0,
but ,
0 =
_
G
[ f, U(a, b)h) [
2
d(a, b) = C
h
[[f[[
2
Since h is not identically zero, it follows that [[f[[ = 0. Thus f = 0
This proves Lemma 9.
Denition 13 For a function h : A C, dene the left and right
translations of h by
(I
a
h)(x) = h(a
1
x) and (D
a
h)(x) = h(xa) where a, x A
Denition 14 For a function h : A C, we say that:
a) h is left uniformly continuous if [[I
a
h h[[

0 as a e
1
b) h is right uniformly continuous if [[D
a
hh[[

0 as a e
1
,
where [[ [[

is the uniform norm

0 as a e
1
b) [[T
b
h h[[

0 as b e
2
Lemma 12 Let h be in L
2
(A). Then
1) [[J
a
h h[[
2
0 as a e
1
2) [[T
b
h h[[
2
0 as b e
2
Lemma 13 Let h be in L
2
(A). Then [[U(a, b)h h[[
2
0 as
(a, b) (e
1
, e
2
)
Proof By Lemma 12,
[[U(a, b)h h[[
2
= [[J
a
T
b
h h[[
2
= [[J
a
T
b
h J
a
h +J
a
h h[[
2
[[J
a
(T
b
h h)[[
2
+[[J
a
h h[[
2
= [[T
b
h h[[
2
+[[J
a
h h[[
2
0
as a e
1
and b e
2
.
This proves Lemma 13.
Lemma 14 The representation U(a, b) is strongly continuous.
Proof Le us prove that [[U(a, b)h U(a
1
, b
1
)h[[
2
0 as
(a, b) (a
1
, b
1
) for any h in L
2
(A).
Consider a, a
1
in A. Then for b, b
1
in B,
J
a
T
b
h = J
a
T
b
_
T
b
1
1
(J
a
1
1
J
a
1
)T
b
1
_
h
= (J
a
T
b
T
b
1
1
J
a
1
1
)(J
a
1
T
b
1
h) = (J
a
T
bb
1
1
J
a
1
1
)(J
a
1
T
b
1
h)
= J
a
(J

bb
1
1
(a
1
1
)
T
bb
1
1
)(J
a
1
T
b
1
h) = (J
a
bb
1
1
(a
1
1
)
T
bb
1
1
)(J
a
1
T
b
1
h)
15
Then for u = J
a
1
T
b
1
h,
J
a
T
b
h J
a
1
T
b
1
h = J
a
bb
1
1
(a
1
1
)
T
bb
1
1
u u
= J
a
bb
1
1
(a
1
1
)
T
bb
1
1
u J
a
bb
1
1
(a
1
1
)
u +J
a
bb
1
1
(a
1
1
)
u u
= J
a
bb
1
1
(a
1
1
)
_
T
bb
1
1
u u
_
+
_
J
a
bb
1
1
(a
1
1
)
u u
_
Then by Lemma 12,
[[J
a
T
b
h J
a
1
T
b
1
h[[
2
[[J
a
bb
1
1
(a
1
1
)
_
T
bb
1
1
u u
_
[[
2
+[[J
a
bb
1
1
(a
1
1
)
u u[[
2
= [[T
bb
1
1
u u[[
2
+[[J
a
bb
1
1
(a
1
1
)
u u[[
2
0 as (aa
1
1
, bb
1
1
) (e
1
, e
2
)
Thus, [[J
a
T
b
h J
a
1
T
b
1
h[[
2
0 as (a, b) (a
1
, b
1
).
This proves Lemma 14.
8 Reconstruction formula
Lemma 15 . For any f, g in L
2
(A) and an admissible non-zero
function h in L
2
(A), we have the following identity in the weak sense
f =
1
C
h
_
G
(L
h
f)(a, b) U(a, b)h d(a, b) (9)
Proof The representation U(a, b) is a strongly continuous irre-
ducible unitary representation of the locally compact topological
group G = A B acting on the Hilbert space L
2
(A). So, if there
is a non-zero admissible vector h in L
2
(A), then by the Grossmann-
Morlet-Paul theorem [2], for f, g in L
2
(A),
_
G
f, U(a, b)h) g, U(a, b)h)d(a, b) = C
h
f, g) (10)
16
Hence,
f =
1
C
h
_
G
(L
h
f)(a, b) U(a, b)h d(a, b)
in the weak sense.
This proves Lemma 15.
9 Plancherels Theorem
Lemma 16 Parsevals formula. For f, g in L
2
(A),
f, g)
L
2
(A)
=
1
C
h
L
h
f, L
h
g)
L
2
(G)
(11)
Proof It comes from (10)
Lemma 17 Palncherels Theorem. For f, g in L
2
(A),
[[f[[
2
L
2
(A)
=
1
C
h
[[L
h
f[[
2
L
2
(G)
(12)
Proof If f = g, then from Lemma 16
f, f)
L
2
(A)
=
1
C
h
L
h
f, L
h
f)
L
2
(G)
That is,
[[f[[
2
L
2
(A)
=
1
C
h
[[L
h
f[[
2
L
2
(G)
which means
_
A
[f(a)[
2
d
A
(a) =
1
C
h
_
G
[(L
h
f)(a, b)[
2
d(a, b)
This proves Lemma 17
17
10 Reproducing kernel Hilbert space
We now give a characterization of the image of the wavelet transform
by a reproducing kernel. Note that not every function F(a, b) in
L
2
(G) is the wavelet transform of some function f in L
2
(A). That
is,
Im(L
h
(a, b)) = F(a, b)[(L
h
f)(a, b) = F(a, b) for some f L
2
(A)
is a proper subspace of L
2
(G).
To see this, note that for F(a, b) = (L
h
f)(a, b), we have F(a, b) is
bounded since
[F(a, b)[ = [(L
h
f)(a, b)[ = [ f, U(a, b)h) [ [[f[[
2
[[h[[
2
Hence, any unbounded and square integrable function F(a, b) is not
in Im(L
h
(a, b)). Then we have the following Lemma.
Lemma 18 The image of the wavelet transform with respect to an
admissible function h in L
2
(A) is the closed subspace of functions
F(a, b) in L
2
(G) that satisfy
F(a, b) =
1
C
h
_
G
F(a
0
, b
0
)K(a, b; a
0
, b
0
)d(a
0
, b
0
)
where
K(a, b; a
0
, b
0
) = (L
h
U(a, b)h)(a
0
, b
0
)
is the reproducing kernel associated with h.
Proof If F is in Im(L
h
(a, b)), there is f L
2
(A) such that
(L
h
f)(a, b) = F(a, b). Then by (11) with g = U(a, b)h,
F(a, b) = (L
h
f)(a, b) = f, U(a, b)h) =
1
C
h
L
h
f, L
h
g)
=
1
C
h
_
G
(L
h
f)(a
0
, b
0
)(L
h
g)(a
0
, b
0
)d(a
0
, b
0
)
18
Now by taking
K(a, b; a
0
, b
0
) = (L
h
g)(a
0
, b
0
)
we have
F(a, b) =
1
C
h
_
G
F(a
0
, b
0
)K(a, b; a
0
, b
0
)d(a
0
, b
0
) (13)
This shows that the image of L
h
(a, b) is a reproducing kernel Hilbert
space embedded as a close subspace of L
2
(G,
1
C
h
d(a, b)), where
K(a, b; a
0
, b
0
) = [L
h
J
a
T
b
h] (a
0
, b
0
) = (L
h
h)(
b
1(a
1
a
0
), b
1
b
0
)
is the reproducing kernel
This proves Lemma 18.
References
[1] I. Daubechies, Ten Lectures on Wavelets, Siam, Philadelphia,
1992.
[2] A. Grossmann, J. Morlet, and T. Paul, Transforms associated to
square integrable group representations, I. General Results, J.
Math. Phys., 26(1985), pp.2473 2479.
[3] G. Folland, A course in Abstract Harmonic Analysis, CRC
Press, 2000.
[4] J. Navarro, Singularities of Distributions Via the Wavelet Trans-
form, Siam J. Math. Anal., Vol. 30 , No 2(1999), pp.454467.
[5] J. Navarro, The Wavelet Transform of Distributions, Memo-
rias del XXVII Congreso Nacional de la Sociedad Matem atica
Mexicana, Mexico (1995).
19
[6] J. Navarro, Use of the Wavelet Transform in 1
n
to nd sin-
gularities of functions in L
2
(1
n
) , Revista Colombiana de
Matem aticas, Vol. 32(1998), No.2, pp. 93-99.
[7] J. Navarro, The wavelet transform with rotations , Sampling
Theory in Signal and Image Processing, Vol. 2, (2005), pp. 101-
106.
[8] R. Murenzi, Wavelet Transforms Associated to the n-
Dimensional Euclidean Group with Dilations :Signal in More
Than One Dimension , J.M. Combes, A. Grossmann, and Ph,
Tchamitchian, eds., Springer-Verlag , Berlin, (1989), pp.239246.
[9] H. Reiter, Classical Harmonic Analysis and Locally Compact
Groups, Oxford University Press 2000