COMP760, LECTURES 2,3: FOURIER ANALYSIS OF FINITE ABELIAN

GROUPS
HAMED HATAMI
In this lecture we develop the basic Fourier analysis of nite Abelian groups. Recall that the
cyclic group Z
N
is the Abelian group with elements 0, 1, . . . , N 1, where the group product is
dened as a + b := a + b (mod N). Finite Abelian groups can be characterized as the products of
cyclic groups:
Theorem 0.1. Every nite Abelian group G is isomorphic to the group Z
N
1
. . . Z
N
k
for some
positive integers N
1
, . . . , N
k
.
In this course, we will be mostly interested in the group Z
n
2
as it can be identied with the set
0, 1
n
. Hence boolean functions f : 0, 1
n
0, 1 can be identied with functions f : Z
n
2

0, 1, and this shall allow us to use Fourier analysis of Z
n
2
to study boolean functions.
1. Basic Fourier Theory
Let G be a nite Abelian group. A function : G C 0 mapping the group to the
non-zero complex numbers is called a character of G if it is a group homomorphism. That is,
(a + b) = (a)(b) for all a, b G, and (0) = 1, where 0 is the identity of G. The constant
function 1 is called the principal character of G.
Let be a character of G, and consider an element a G. Since G is a nite group, a is of some
nite order n (that is na = 0). Hence 1 = (0) = (na) = (a)
n
which shows that (a) is an n-th
root of unity. In particular, every character

G satises
(1) : G T,
where T is the unit complex circle.
Theorem 1.1. If G is a nite Abelian group, then the characters of G together with the usual
pointwise product of complex valued functions form a group

G.
Proof. The principal character 1 is the identity of

G. Note that if and are characters of G,
then is also a character. Indeed (ab)(ab) = (a)(a)(b)(b), and (0)(0) = 1 1 = 1. To
check the existence of the inverse element, note that if is a character, then
1
=
1

= is also
a character.
The group

G is called the Pontryagin dual of G. Fourier analysis is based on expressing functions
f : G C as linear combinations of characters. It will be convenient to treat the set of these
functions as a Hilbert space: Let L
2
(G) denote the set of functions f : G C, where here G is
endowed with the uniform probability measure. Recall (see Lecture 1, Section 3.1) that L
2
(G) is a
Hilbert space with the inner product
f, g = E
xG
f(x)g(x) =
1
[G[

xG
f(x)g(x).
1
2 HAMED HATAMI
In the sequel we will usually consider G as a probability space, and E
xG
shall always mean that
x is a random variable that takes values in G uniformly at random. To simplify the notation we
usually abbreviate E
xG
to simply E. Hence for a function f : G C, the notation E[f] means
E
xG
[f(x)] (which is equal to
1
|G|

xG
f(x)).
Our next goal will be to prove that the characters form an orthonormal basis for this space. First
let us prove a simple lemma.
Lemma 1.2. Let G be a nite Abelian group, and be a non-principal character of G. Then

xG
(x) = 0.
Proof. Suppose to the contrary that

xG
(x) ,= 0. Consider an arbitrary y G. Then We have
(y)

xG
(x) =

xG
(y + x) =

xG
(x)
which shows that (y) = 1. Since y was arbitrary, we conclude that must be the principal
character which is a contradiction.
Now we can prove the orthogonality properties of the characters.
Lemma 1.3. The characters of a nite Abelian group G are orthonormal functions in L
2
(G).
Proof. It follows from (1) that every

G satises
||
2
2
= E
_
[(x)[
2

= E[1] = 1.
So characters are unit vectors in L
2
(G). It remains to verify the orthogonality. Let ,= be two
dierent characters. Then =
1
is a non-principal character of G. Hence by Lemma 1.2, we
have
, = E
_
(x)(x)

= E
_
(x)

= 0.

So far we have discussed the Pontryagin dual of G in an abstract manner. Since nite Abelian
groups have simple structures (Theorem 0.1), it is quite easy to describe the characters of G. We
start with the basic case of G = Z
N
. For every a Z
N
, dene
a
L
2
(G) as

a
: x e
2i
N
ax
.
Let us verify that
a
is actually a character. Indeed
a
(0) = e
2i
N
0
= e
0
= 1, and since e
2i
= 1, we
have

a
(x)
a
(y) = e
2i
N
ax
e
2i
N
ay
= e
2i
N
a(x+y (mod N))
=
a
(x + y).
Note that L
2
(G) is [G[-dimensional, and hence by Lemma 1.3, G has at most [G[ characters. It
follows that
a
: a G are all the characters of G. The principal character is
0
= 1. Also

b
=
a+b
which shows that the dual group

G is isomorphic to G. As we shall see below this is
in general true for all nite Abelian groups.
Now let us consider the general case of G = Z
N
1
. . . Z
N
k
for some positive integers N
1
, . . . , N
k
.
For every a = (a
1
, . . . , a
k
) G, dene
a
L
2
(G) as

a
: x
k

i=1
e
2i
N
i
a
i
x
i
.
As in the case of Z
N
, it is straightforward to verify that
a
is a character by showing that
a
(0) = 1,
and
a
(x + y) =
a
(x)
a
(y). Again Lemma 1.3 shows that
a
: a G are all the characters of
G. We also have the identify
a

b
=
a+b
which proves Theorem 1.4.
COMP760, LECTURES 2,3: FOURIER ANALYSIS OF FINITE ABELIAN GROUPS 3
Theorem 1.4. If G is a nite Abelian group, then the characters of G form an orthonormal basis
for L
2
(G). Furthermore we have

G

= G.
Before continuing further, let us look at the characters of our favorite group Z
n
2
. Consider an
element a Z
n
2
. Then

a
(x) = e
2i
2

n
i=1
a
i
x
i
= (1)

n
i=1
a
i
x
i
.
Note that in this case the characters are actually real valued (they only take values 1 and 1).
Since the coordinates of a are 0 or 1, we will sometimes identify a with the set S = j 1, . . . , n :
a
j
= 1, and denote the characters as
S
for S 1, . . . , n. This notation is sometimes more
intuitive as

a
(x) = e
2i
2

n
i=1
a
i
x
i
= (1)

n
i=1
a
i
x
i
= (1)

iS
x
i
.
Later when we take a probabilistic approach to decomposing functions, this notation becomes more
natural as it extends to general product spaces (where there is no group structure).
Denition 1.5. The Fourier transform of a function f : G C is the unique function

f :

G C
dened as

f() = f, = Ef(x)(x).
Theorem 1.4 shows that G is isomorphic to its dual

G, and so it shall be convenient to identify
the two groups in the sequel. Let us restate Denition 1.5 in this new notation.
Denition 1.6 (Denition 1.5 Restated). The Fourier transform of a function f : G C is the
unique function

f : G C dened as

f(a) = f,
a
= Ef(x)
a
(x).
Let us state a simple example of the Fourier transform of a function on Z
n
2
.
Example 1.7. Let f : Z
n
2
C be the parity function f : x

n
i=1
x
i
(mod 2). Then

f(0) = Ef(x)
0
= Ef(x) =
1
2
.
We also have

f(1, . . . , 1) = Ef(x)(1)

n
j=1
x
j
=
1
2
,
as f(x) = 1 if and only if

n
j=1
x
j
= 1(mod 2). Next consider a Z
n
2
with a ,= (1, . . . , 1) and
a ,= 0. Let j
0
, j
1
be such that a
j
0
= 0 and a
j
1
= 1. We have (why?)

f(a) = Ef(x)
a
(x) =
1
2
E[f(x)
a
(x) + f(x + e
j
0
+ e
j
1
)
a
(x + e
j
0
+ e
j
1
)] ,
where e
j
denotes the vector in Z
n
2
which has 1 at its jth coordinate and 0 everywhere else. Note
that f(x) = f(x+e
j
0
+e
j
1
) and furthermore
a
(x) =
a
(x+e
j
0
+e
j
1
). We conclude that

f(a) = 0
for every a Z
n
2
satisfying a ,= (1, . . . , 1) and a ,= 0.
4 HAMED HATAMI
Lecture 3
The Fourier transform is a linear operator:

f + g =

f + g, and we have the following easy

G) is norm
decreasing:
|

f|

|f|
1
.
Proof. By (1) for every a G, we have
[

f(a)[ =

Ef(x)
a
(x)

E[f(x)[[
a
(x)[ = E[f(x)[ = |f|
1
.

The Fourier coecient

f(0) = E[f(x)].
So if 1
S
is the indicator function of a subset 1
S
G, then

1
S
(0) =
|S|
|G|
corresponds to the density
of S.
It follows from the fact that the characters from an orthonormal basis for L
2
(G) that
f =

aG

f(a)
a
,
and that this expansion of f as a linear combination of characters is unique. This formula is called
the Fourier inversion formula as it shows how the functions f can be reconstructed from its Fourier
transform.
If S G, then the orthogonal complement of S is dened as
S

= a G :
a
(x) = 1 x S.
It follows from the identities
0
= 1 and
a

b
=
a+b
that S

is a subgroup of G. The Fourier

1
H
(a) =
_
[H[/[G[ a H

0 a , H

Proof. If a H

, then

1
H
(a) = 1
H
,
a
= E1
H
(x)
a
(x) = E1
H
(x) = [H[/[G[.
On the other hand if a , H

, then there exists y H such that

zH

a
(z) =
a
(y)

zH

a
(z y) =
a
(y)

zH

a
(z),
which shows that

zH

a
(z) = 0. Hence

1
H
(a) = E1
H
(x)
a
(x) =
1
[G[

zH
E1
H
(z) = 0.

COMP760, LECTURES 2,3: FOURIER ANALYSIS OF FINITE ABELIAN GROUPS 5

1
S
(a) = E1
S
(x)
a
(x) = E1
H
(x y)
a
(x) = E1
H
(x)
a
(x + y) = (y)

1
H
(a)
=
_
(y)[H[/[G[ a H

0 a , H

Example 1.11. Let us revisit Example 1.7 in light of Remark 1.10. Note that H = x Z
n
2
:

n
i=1
x
i
= 0 (mod 2) is a subgroup of Z
n
2
. Now the function f dened in Example 1.7 is the
indicator function of S = e
1
+ H. Note that
H

= a : (1)

n
i=1
x
i
a
i
= 1 x H = (0, . . . , 0), (1, . . . , 1).
Hence

f(a) =

1
S
(a) =
_

a
(e
1
)[H[/[G[ a H

0 a , H

We conclude that

f(0) = 1/2 and

f(1, . . . , 1) = 1/2, and

f(a) = 0 for every a Z
n
2
satisfying
a ,= (1, . . . , 1) and a ,= 0.
Theorem 1.12 (Parseval). For every f L
2
(G),
|f|
2
2
=

aG
[

f(a)[
2
.
Proof. We have
|f|
2
2
= f, f =
_

aG

f(a)
a
,

bG

f(b)
b
_
=

a,bG

f(a)

f(b)
a
,
b
.
The identify now follows from orthonormality of characters:

a
,
b
=
_
0 a ,= b;
1 a = b.

The proof of the Parseval identity, when applied to two dierent functions f, g L
2
(G), implies
the Plancherel theorem:
f, g =

aG

f(a) g(a).
As the rst example of an application of the Parseval identity, let us show that for every subgroup
H of G, we have
(2) [H[[H

[ = [G[.
Indeed by Lemma 1.9, we have
[H[
[G[
= E1
H
= E1
2
H
= 1
H
, 1
H
= |1
H
|
2
2
=

aG
[

1
H
(a)[
2
=

aH

([H[/[G[)
2
=
[H[
2
[H

[
[G[
2
which simplies to (2).
Next we introduce the important notion of convolution.
6 HAMED HATAMI
Denition 1.13. Let G be a nite Abelian group. For two functions f, g : G C, we dene their
convolution f g : G C as
f g(x) = E
yG
[f(x y)g(y)].
Note that f g(x) is the average of f(a)f(b) over all pairs a, b with a + b = x. This gives a
combinatorial nature to convolution which makes it very useful in dealing with certain discrete
problems. Consider a set S G. Then f 1
S
(x) is the average of f over the set x S := x y :
y S. For example if S is the Hamming ball
1
of radius r around 0 in Z
n
2
, then f 1
S
(x) is
the average of f over the Hamming ball of radius r around x. These types of averaging operators
usually smooth f, and makes it more similar to a constant functions. This smoothing property
of the convolution is one of the main tools in harmonic analysis and this course.
Next let us list some basic facts about the convolution. We dene the support of f : G C,
denoted by Supp(f), to be the set of the points x G with f(x) ,= 0.
Lemma 1.14. Consider three functions f, g, h : G C.
(a) We have
f g = g f.
(b) We have
(f g) h = f (g h).
(c) We have
f (h + g) = f h + f g.
(d) We have
Supp(f + g) Supp(f) + Supp(g).
(e) We have
|f g|

|f|
1
|g|

.
(f ) More generally, if p and q are conjugate exponents, then
|f g|

|f|
p
|g|
q
.
(g) We have
|f g|
1
|f|
1
|g|
1
.
Proof. (a) For every x G, we have
f g(x) = E
y
[f(x y)g(y)] = E
y
[f(x y)g(x (x y))] = E
z
[f(z)g(x z)] = g f(x).
(b) By Part (a),
(f g) h(x) = (g f) h(x) = E
z
E
y
[g(x z y)f(y)]h(z) =
= E
y,z
g(x z y)f(y)h(z) = (h g) f(x) = f (g h)(x).
(c) is trivial.
(d) follows from the fact that f(x) is the average of f(a)g(b) over all pairs of points a, b G
with a + b = x.
(e) is a special case of (f ).
(f ) Note that for every x G, by Holders inequality we have
[f g(x)[ E
yG
[f(x y)[[g(y)[ (E[f(x y)[
p
)
1/p
(E[g(y)[
q
)
1/q
= (E[f(y)[
p
)
1/p
|g|
q
= |f|
p
|g|
q
.
(g) We have
|f g|
1
= E
x
[f g(x)[ E
x,y
[f(x y)[[g(y)[ = E
z,y
[f(z)[[g(y)[ = E
z
[f(z)[E
y
[g(y)[ = |f|
1
|g|
1
.
1
The Hamming ball of radius r around 0 is dened as {x Z
n
2
:

n
i=1
xi r} Z
n
2
.
COMP760, LECTURES 2,3: FOURIER ANALYSIS OF FINITE ABELIAN GROUPS 7

The relevance of the Fourier transform to convolution lies in the following lemma.
Lemma 1.15. If f, g : G C, then

f g =

f g.
Proof. We have

f g(a) = E
x
f g(x)
a
(x) = E
x
(E
y
f(x y)g(y))
a
(x) = E
x,y
f(x y)g(y)
a
(x y)
a
(y)
= E
z,y
f(z)g(y)
a
(z)
a
(y) = E
z
f(z)
a
(z)E
y
g(y)
a
(y) =

f(a) g(a).

Note that Lemma 1.15 in particular shows that

yG

f(x y) g(y),
which converts pointwise product back to convolution.
Exercise 1.16. Prove the Identity (3).
For a function h : G C, dene

h : G C as h : x h(x). Note that

h =

aG

h(a)
a
.
Hence it follows from the Parseval identity and Lemma 1.15 that for f, g, h : G C, we have
f h, g = f, g

h =

aG

f(a)

h(a) g(a).
School of Computer Science, McGill University, Montr eal, Canada
E-mail address: [email protected]