FOURIER ANALYSIS ON FINITE NON-ABELIAN GROUPS

TERENCE TAO

Abstract. Some notes on the Fourier transform on finite non-abelian groups

G. These notes are aimed at readers who are familiar with Fourier analysis on
abelian groups such as R, Z, and T, but are not familiar with representation
theory of non-abelian groups.

1. Hilbert spaces
In these notes, when we say that V is a Hilbert space, we mean that V is a finite
dimensional complex vector space equipped with an inner product ⟨v, w⟩ which is
Hermitian (linear in v, anti-linear in w) and positive definite.
Given any linear transformation T : V → V from a Hilbert space to itself, we
can define its trace tr(T ) by the formula
∑
tr(T ) := ⟨e, T e⟩
e
where e ranges over some orthonormal basis of V ; it is easy to see that this definition
is independent of the choice of basis. Observe that trace is linear and that tr(ST ) =
tr(T S).
If V, W are finite dimensional complex Hilbert spaces, we define HS(V → W ) to
be the space of all Hilbert-Schmidt operators from V to W ; since V, W are finite-
dimensional, HS(V → W ) is just the same as Hom(V → W ). We observe that
HS(V → W ) is also a Hilbert space, with the Hilbert-Schmidt inner product
⟨S, T ⟩HS(V →W ) := tr(ST ∗ ).
where S ∗ is of course the adjoint of S. We abbreviate HS(V → V ) as HS(V ).
We use IV : V → V to denote the identity operator on V ; note that tr(IV ) =
dim(V ).
We use U (V ) ⊂ HS(V ) to denote the space of unitary operators on V , i.e. those
operators T ∈ HS(V → V ) such that T T ∗ = T ∗ T = IV . This forms a group with
identity IV .

2. Finite groups
Let G be a finite group with group operation ◦, identity id, and cardinality |G|.
We give this group normalized counting measure dx:
∫
1 ∑
f (x) dx := f (x),
G |G|
x∈G

and then define L (G) to be the Hilbert space of functions f : G → C endowed

2

with the inner product ∫

⟨f, g⟩ := f (x)g(x) dx.
G
1
2 TERENCE TAO

Note that this inner product is always finite since G is finite.

We note a special type of function in L2 (G): the Dirac delta functions. Specifi-
cally, for every x ∈ G, we define δx ∈ L2 (G) by duality, setting
⟨f, δx ⟩ = f (x) for all f ∈ L2 (G). (1)
Equivalently, we can define δx (y) := |G| when x = y, and δx (y) := 0 when x ̸= y.
Note that δx is indeed in L2 (G) because G is finite. Indeed, the functions {δx : x ∈
G} form a basis for L2 (G), thanks to the reproducing formula
∫
f= ⟨f, δx ⟩δx dx.
G

Given any y ∈ G, we can define the translation operator τy : L2 (G) → L2 (G) by

the formula
τy δx := δy◦x for all x ∈ G
and then extending τy by linearity. In other words, we have
τy f (x) := f (y −1 ◦ x) for all f ∈ L2 (G).
Observe that τy acts unitarily on L2 (G) and we have the homomorphism property
τx τy = τx◦y
(this is best seen by first checking this on Dirac delta functions). In particular, τid
is the identity, and τx−1 is the inverse of τx .
We define the operation of convolution ∗ : L2 (G) × L2 (G) → L2 (G) by defining
δx ∗ δy := δx◦y and extending by bilinearity. Thus
∫
f ∗ g(x) = f (y)g(y −1 ◦ x) dy.
G
1 2
This operation turns L (G) into an algebra, called the convolution algebra; this
algebra is commutative if and only if the group G is. One of the main purposes of
the Fourier transform is to recast this algebra in a more tractable form.

3. Representations
By a (unitary) representation ρ of G, we mean a Hilbert space Vρ , together with
a homomorphism ρ : G → U (V ). In other words, for each group element x we
assign a unitary operator ρ(x) : Vρ → Vρ such that ρ(x)ρ(y) = ρ(xy).
Examples of representations include the zero representation 0, with V0 = {0}
and 0(x) := 0; the trivial representation ρ0 , with Vρ0 = C and ρ0 (x) = IC ; and the
regular representation τ , with Vτ := L2 (G) and τ (x) := τx .
A morphism ϕ : ρ1 → ρ2 between two representations is any unitary map ϕ :
Vρ1 → Vρ2 such that we have the intertwining relationship
ρ2 (x)ϕ = ϕρ1 (x) for all g ∈ G.
If the morphism has an inverse, we say it is an isomorphism.
We say that two representations ρ1 , ρ2 are isomorphic, and write ρ1 ≡ ρ2 , if
there is an isomorphism from ρ1 to ρ2 . This is clearly an equivalence relation.

1In the case of infinite groups, the relevant algebra would be L1 (G) rather than L2 (G). For
finite groups, of course, these spaces are the same.
 FOURIER THEORY OF FINITE GROUPS 3

A representation ρ gives a map from G to U (Vρ ), which in turn is a subset of

HS(Vρ ). This induces a map from L2 (G) to HS(Vρ ) by defining ρ(δx ) := ρ(x) and
extending by linearity. Thus
∫
ρ(f ) := f (x)ρ(x) dx. (2)
G
For any representation ρ, we observe the identity
ρ(f ∗ g) = ρ(f )ρ(g); (3)
this is easiest to see by checking first on delta functions. (This gives the first hint
as to what the Fourier transform on G should be; but more on this later.)

4. The regular representation

Perhaps the most important representation for Fourier analysis is the regular
representation τ , defined in the previous section. Observe that for any f ∈ L2 (G),
the Hilbert-Schmidt operator τ (f ) is the same as convolution by f :
τ (f )g = f ∗ g. (4)
This is easiest seen by first verifying the case when f, g are both delta functions,
and then using bilinearity.
The map f 7→ τ (f ) is a linear map from L2 (G) to HS(L2 (G)). This map is in
fact an isometry:
Lemma 4.1. Let f ∈ L2 (G). Then we have
∥τ (f )∥HS(L2 (G)) = ∥f ∥L2 (G) .
In particular, this implies the regular representation is faithful (i.e. the map
f 7→ τ (f ) is injective).
Proof Squaring both sides, we reduce to showing that
∫
tr(τ (f )∗ τ (f )) = |f (x)|2 dx.
G
∫
We have the easily verified identity tr(T ) = x
⟨δx , T δx ⟩ dx for any T ∈ HS(L2 (G)).
Thus
∥τ (f )∥2HS(L2 (G)) = tr(τ (f )∗ τ (f ))
∫
= ⟨τ (f )∗ τ (f )δx , δx ⟩ dx
∫G

= ⟨τ (f )δx , τ (f )δx ⟩ dx
∫G
= ∥f ∗ δx ∥2L2 (G) dx
∫G
= ∥f ∥2L2 (G) dx
G
= ∥f ∥2L2 (G)
as desired.
From (3) we have τ (f ∗ g) = τ (f )τ (g) (this is just a fancy way of saying convo-
lution is associative). Thus the map f 7→ τ (f ) is not only a Hilbert space isometry,
4 TERENCE TAO

it is also a homomorphism of algebras, mapping the convolution algebra of L2 (G)

to some operator sub-algebra of HS(L2 (G)).
In some sense, the map f 7→ τ (f ) already captures the essence of the Fourier
transform. However, things are not yet satisfactory because this map f 7→ τ (f ) is
not onto; not every Hilbert-Schmidt operator on L2 (G) is a convolution operator.
Thus we must discover exactly what subspace of HS(L2 (G)) is the image of τ .
In the abelian case, we know that there is a one-to-one correspondence between
convolution operators and Fourier multipliers; to put it another way, the convolution
operators are precisely those operators which are diagonal in the Fourier basis. In
the non-abelian case, we will have a similar characterization, except that we will
replace “diagonal” by “block-diagonal”. To do this we need a notion of direct sum
and irreducible representation, to which we now turn.

5. Irreducibility
If V and W are Hilbert spaces, then V ⊕ W := {(v, w) : v ∈ V, w ∈ W } is also
a Hilbert space, and if T : V → V and S : W → W are Hilbert-Schmidt, we can
define the direct sum T ⊕ S : V ⊕ W → V ⊕ W as (T ⊕ S)(v, w) := (T v, Sw).
Given two representations ρ1 and ρ2 , we can define their direct sum ρ1 ⊕ ρ2 by
defining Vρ1 ⊕ρ2 := Vρ1 ⊕ Vρ2 and (ρ1 ⊕ ρ2 )(x) := ρ1 (x) ⊕ ρ2 (x). By linearity, we
thus have that
(ρ1 ⊕ ρ2 )(f ) = ρ1 (f ) ⊕ ρ2 (f ) (5)
Also observe that ρ1 ⊕ ρ2 ≡ ρ2 ⊕ ρ1 .
We call a representation ρ reducible if it can be written as ρ ≡ ρ1 ⊕ ρ2 for some
non-zero representations ρ1 , ρ2 , and irreducible if it is non-zero and not reducible;
we consider the zero representation to be neither reducible nor irreducible. We will
use Γ to index the irreducible representations of G, up to isomorphism; thus every
ξ ∈ Γ gives rise to an irreducible representation ρξ , and ρξ ≡ ρξ′ if and only if
ξ = ξ ′ . We shall index Γ so that ρ0 is the trivial representation. We shall also
abbreviate Vρξ as Vξ .
Because we are requiring all our representations to be finite dimensional, every
representation ρ can be decomposed as the sum of a finite number of irreducible
representations,
⊕ ⊕c
ρ≡ ρξ ξ
ξ∈Γ

where the cξ are some non-negative integers (called the multiplicity of ρξ in ρ),
⊕c
and ρξ ξ is the direct sum of cξ copies of ρξ . There is an issue as to whether the
coefficients cξ are unique; they are, but we will prove this later when we derive an
explicit formula for them.
In particular, the regular representation can be decomposed as
⊕ ⊕m
τ≡ ρξ ξ (6)
ξ∈Γ

for some non-negative integers mξ . Note that since τ is finite dimensional, at most
finitely many of the mξ are non-zero.
We now fix the mξ (if there is failure of uniqueness, we pick mξ arbitrarily; later
on we will compute mξ explicitly and show that one does have uniqueness, but it
 FOURIER THEORY OF FINITE GROUPS 5

is not needed for this present discussion). Now let f ∈ L2 (G). By (6) and (5) we
see that ∑
∥τ (f )∥2HS(L2 (G)) = mξ ∥ρξ (f )∥2HS(Vξ ) .
ξ∈Γ
We rewrite this by defining a measure dξ on Γ by
∫ ∑
f (ξ) dξ := mξ f (ξ).
Γ ξ∈Γ

We also rewrite ρξ (f ) as fˆ(ξ), thus

∫
fˆ(ξ) = f (x)ρξ (x) dx (7)
G
is a Hilbert-Schmidt operator on Vξ for each ξ. Then, using Lemma 4.1, we have
the Plancherel formula
∫
∥f ∥L2 (G) =
2
∥fˆ(ξ)∥2HS(Vξ ) dξ (8)
Γ
for all f ∈ L2 (ξ).
We have∫ thus defined a Fourier transform
∫ f 7→ fˆ, which is an isometry from
2
L (G) to Γ HS(Vξ ) dξ, where we use Γ HS(Vξ ) dξ to denote the Hilbert space
obtained by taking the direct sum of all the HS(Vξ ) and endowing the resulting
space with the inner product indicated (8). Also, from (3) we observe the familiar
identity
f[
∗ g(ξ) = f (ξ)g(ξ). (9)
Thus the Fourier transform is an algebra homomorphism as well as an isometry.
Next, we consider how to invert the Fourier transform. By depolarizing (8), we
obtain the Parseval identity
∫
⟨f, g⟩ = ⟨fˆ(ξ), ĝ(ξ)⟩HS(Vξ ) dξ
Γ
for all f, g ∈ L2 (G). In particular, from (1) we have
∫
f (x) = ⟨fˆ(ξ), δ̂x (ξ)⟩HS(Vξ ) dξ
Γ
for all x ∈ G. Since
δ̂x (ξ) = ρξ (x) (10)
by definition, we thus have the Fourier inversion formula
∫
f (x) = ⟨fˆ(ξ), ρξ (x)⟩HS(Vξ ) dξ. (11)
Γ
From (9) and (11) we see that translations map to modulations on the Fourier
transform side:
τd \ ˆ
x f (ξ) = δx ∗ f (ξ) = ρξ (x)f (ξ). (12)
We have already derived many of the familiar properties of the Fourier transform.
There are two major tasks left to do to obtain a truly satisfactory theory: first, we
have to show that the Fourier transform is in fact surjective (we already know it
is injective by (11)), and secondly we have to work out what the coefficients mξ .
(There is also the question of how one locates all the irreducible representations Γ,
6 TERENCE TAO

but this leads of course into the field of representation theory, which we will not
detail here.)
In order to address these two questions we need a fundamental tool in represen-
tation theory, namely Schur’s lemma.

6. Schur’s lemma
A basic fact about irreducible representations is that they have no non-trivial
invariant spaces:
Lemma 6.1. Let ξ ∈ Γ, and let W be a subspace of Vξ such that ρξ (x)W ⊆ W for
all x ∈ G (i.e. W is an invariant space of ρξ ). Then W = Vξ or W = {0}.
Proof Suppose W is a proper subspace of Vξ . Then W ⊥ , the orthogonal comple-
ment of W in Vξ , is also a proper subspace. Since ρξ (x) is unitary and leaves W
invariant, it also leaves W ⊥ invariant (here we are using the finite dimensionality).
Thus ρξ ≡ ρξ |W ⊕ ρξ |W ⊥ , contradicting irreducibility.
This has the following important consequence:
Lemma 6.2 (Schur’s lemma). If ξ, ξ ′ ∈ Γ and ϕ : ρξ → ρξ′ is an morphism, then
either ϕ = 0, or ϕ is an isomorphism. If ξ = ξ ′ , then ϕ is a constant multiple of
the identity IVξ .
Proof Observe that ϕ(Vξ ) is an invariant subspace of Vξ′ , and hence is either equal
to Vξ′ or {0}. Similarly, ker(ϕ) is an invariant subspace of Vξ , and is hence equal
to either Vξ or {0}. Considering all the possibilities we see that ϕ is either 0 or an
isomorphism.
Now suppose ξ = ξ ′ . Then by the first part of Schur’s lemma, we know that
ϕ − λIVξ (which is also a morphism) is either an isomorphism or zero for every
λ ∈ C. Since ϕ must have at least one eigenvalue, the claim follows.
As a consequence, we can obtain the following ergodic theorem.
Lemma 6.3. Let ξ, ξ ′ ∈ Γ, and let T∫ ∈ HS(Vξ → Vξ′ ). Let ⟨T ⟩ ∈ HS(Vξ → Vξ′ )
denote the averaged operator ⟨T ⟩ := G ρξ′ (x)T ρξ (x)−1 dx.
• If ξ ̸= ξ ′ , then
⟨T ⟩ = 0. (13)
′
• If ξ = ξ , then
tr(T )
⟨T ⟩ = IV . (14)
dim(Vξ ) ξ
Proof By a change of variables z := y ◦ x we observe the morphism property
∫
ρξ′ (y)⟨T ⟩ = ρξ′ (y ◦ x)T ρξ (x)−1 dx
∫G

= ρξ′ (z)T ρξ (y −1 ◦ z)−1 dz

G
= ⟨T ⟩ρξ (y).
Thus by Schur’s lemma, ⟨T ⟩ is either 0 or an isomorphism. If ξ ̸= ξ ′ then it cannot
be an isomorphism, hence it must be 0. If ξ = ξ ′ then ⟨T ⟩ is a constant multiple of
 FOURIER THEORY OF FINITE GROUPS 7

the identity by Schur’s lemma. But since

∫
tr(⟨T ⟩) = tr(ρξ′ (x)T ρξ (x)−1 ) dx
∫G

= tr(T ) dx
G
= tr(T )
tr(T ) I as claimed.
we must have tr(⟨T ⟩) = dim(V ξ)
Vξ

7. Surjectivity and multiplicity

We now settle the surjectivity and multiplicity questions.
ˆ
∫Proposition 7.1. The Fourier transform f 7→ f is a surjection from L (G) to
2

Γ
HS(Vξ ) dξ, and mξ = dim(Vξ ).
˜ by
Proof We define the modified measure dξ
∫ ∑
˜ :=
f (ξ) dξ f (ξ) dim(Vξ );
Γ ξ∈Γ

˜
our second task is to show ∫that dξ = dξ.
Let F be any element of Γ HS(Vξ ) dξ, so that each F (ξ) is an element of HS(Vξ ),
and let f ∈ L2 (G) be the modified inverse Fourier transform
∫
˜
f (x) := ⟨F (ξ), ρξ (x)⟩ dξ.
Γ

We claim that F = fˆ; this will prove surjectivity, and by comparing the above
˜ (recall that
formula against the Fourier inversion formula (11) we see that dξ = dξ
F is arbitrary).
Now we show F = fˆ. Expanding this, we see that we have to show that
∫ ∫
F (ξ ′ ) = ˜ dx
⟨F (ξ), ρξ (x)⟩ρξ′ (x) dξ
G Γ
′
for all ξ ∈ G. By Fubini’s theorem, it will suffice to show that
∫ { 1
dim(Vξ ) F (ξ) if ξ = ξ ′
⟨F (ξ), ρξ (x)⟩ρξ′ (x) dξ dx = (15)
G 0 if ξ ̸= ξ ′ .
To do this, we use Lemma 6.3, but with T : Vξ → Vξ′ specialized to a rank 1
operator of the form
Ta,a′ v := ⟨v, a⟩a′ for all v ∈ Vξ
for some a ∈ Vξ , a′ ∈ Vξ′ to be chosen later. By definition, we have
∫
⟨Ta,a′ ⟩v = ρξ′ (x)a′ ⟨ρξ (x)−1 v, a⟩ dx.
G

Specializing to the case v := F (ξ)a, and then summing a over an orthonormal basis,
we obtain
∑ ∫
⟨Ta,a′ ⟩F (ξ)a = tr(ρξ (x)−1 F (ξ))ρξ′ (x)a′ dx.
a G
8 TERENCE TAO

When ξ ̸= ξ ′ , we thus see from (13) that

∫
tr(ρξ (x)−1 F (ξ))ρξ′ (x)a′ = 0
G

for all a′ ∈ Vξ′ , which proves the second half of (15). When ξ = ξ ′ , we see from
(14) that
∫ ∑ tr(Ta,a′ )
tr(ρξ (x)−1 F (ξ))ρξ (x)a′ dx = F (ξ)a.
G a
dim(Vξ )
But tr(Ta,a′ ) = ⟨a′ , a⟩, so we can rewrite the right-hand side as
1 ∑
F (ξ) ⟨a′ , a⟩a.
dim(Vξ ) a
∑ ′
Since a ranges over an orthonormal basis, a ⟨a , a⟩a = a′ , and this proves the first
half of (15).
One particular consequence of the above is
Theorem
∫ 7.2 (Peter-Weyl theorem). The convolution algebra L2 (G) is isomorphic
to Γ HS(Vξ ) dξ, as an algebra, as a Hilbert space, and as a representation of G
(this last fact follows from (12); in fact the two spaces are isomorphic for G acting
either on the left or on the right).
As a particular consequence of this isomorphism we see from dimension count
that
∑
|G| = dim(Vξ )2 ; (16)
ξ∈Γ

in particular, we
∫ see that Γ is finite.
The space Γ HS(Vξ ) dξ can be thought of as a space of block-diagonal ma-
trices, with each block having dim(Vξ ) rows and columns (and having weight
mξ = dim(Vξ ). This realizes our earlier stated goal of representing the convolution
operators in HS(L2 (G)) as block-diagonal matrices.

8. Conjugation-invariant functions
The group G acts on L2 (G) by conjugation, f (x) → f (y ◦ x ◦ y −1 ), or in other
words f 7→ τy f τy−1 . The group G similarly
∫ acts on HS(Vξ ) by conjugation, F (ξ) 7→
−1
ρξ (y)F (ξ)ρξ (y) , and hence acts on Γ HS(Vξ ) dξ as well. From the Peter-Weyl
theorem we know that these actions are intertwined by the Fourier transform.
In particular, a function f is invariant under conjugation if and only if its Fourier
transform fˆ(ξ) is invariant under conjugation as well. But by Schur’s lemma (or
(14)) we see that an operator in HS(Vξ ) is invariant under conjugation by the
ρξ (y) if and only if it is a constant multiple of the identity IVξ . Thus the Fourier
transform can be restricted to an isomorphism ∫ between the conjugation invariant
functions L2 (G)G of L2 (G), and the space Γ CIVξ dξ.
Conjugation-invariant functions in L2 (G)G are sometimes called class functions
since they are constant on each conjugacy class. The Fourier analysis of class
functions is simpler than that of general functions, because each block HS(Vξ )
in the matrix representation has been replaced by a scalar. To compute more
effectively we introduce the notion of a character.
 FOURIER THEORY OF FINITE GROUPS 9

For each ξ ∈ Γ, define the Dirac delta mass δξ on Γ by setting δξ (ξ ′ ) := 0 if

′
ξ ̸= ξ and δξ (ξ) := dim(V
1
IVξ . Observe that the Dirac delta masses δξ form an
∫ ξ)

orthonormal basis of Γ CIVξ dξ.

The inverse Fourier transform of δξ is denoted χξ , the conjugate irreducible
character corresponding to ξ; we have
∫
χξ (x) = ⟨δξ , ρξ′ (x)⟩ dξ ′
Γ

and hence
χξ (x) = tr(ρξ (x)).
∫
Since the δξ are an orthonormal basis of Γ CIVξ dξ, the characters χξ form an
orthonormal basis of L2 (G)G ; in particular we have
∫
f = ⟨f, χξ ⟩χξ dξ
Γ

for all f ∈ L (G) ; conjugating, we obtain

2 G
∫
f = ⟨f, χξ ⟩χξ dξ
Γ

Taking Fourier transforms, we obtain

∫
fˆ = ⟨f, χξ ⟩δξ dξ
Γ

and in particular
⟨f, χξ ⟩
fˆ(ξ) = IV .
dim(Vξ ) ξ
We can in fact define a character χρ ∈ L2 (G) for every representation ρ, by the
same formula χρ (x) := tr(ρ(x)). From (5) we have χρ1 ⊕ρ2 = χρ1 + χρ2 , so every
character is a linear combination of the irreducible characters. Also, from the linear
independence of the χξ we thus see that the multiplicity of the irreducible represen-
tation ρξ in any other representation ρ is well-defined; in fact, by orthonormality of
the characters, this multiplicity is equal to ⟨χξ , χρ ⟩. Thus one can identify the space
of representations of G (modulo equivalence) with the positive integer combinations
of the characters χξ .
We now specialize to the case where G is abelian. In this case L2 (G)G = L2 (G),
which implies that HS(Vξ ) = CIVξ for all ξ ∈ Γ; in other words, all irreducible
representations ρξ must be one-dimensional. Thus we may as well set Vξ := C, and
the characters χξ are essentially the same as the representations ρξ . In particular,
since ρξ is unitary, we have |χξ (x)| = 1 for all x ∈ G.
Now observe that if ρξ and ρξ′ are one-dimensional representations, then ρξ ρξ′
is another one-dimensional representation, which must of course be irreducible. We
can thus define an addition operation on Γ by setting ρξ+ξ′ := ρξ ρξ′ . This can
easily be verified to turn Γ into an abelian group, sometimes called the dual group
G∗ of G; from (16) we thus see that |Γ| = |G|. This theory then collapses to the
familiar theory of the Fourier transform on abelian groups.
The trick of multiplying two representations ρξ and ρξ′ to form a third repre-
sentation ρξ ρξ′ is not restricted to one-dimensional representations; it is easy to
take two higher-dimensional representations ρ and ρ′ and form a tensor product
10 TERENCE TAO

ρ ⊗ ρ′ of the two representations. On the character level, this corresponds to point-

wise multiplication of the two characters. The new feature in the non-abelian case,
however, is that the tensor product of two irreducible representations need not be
irreducible. An equivalent formulation of this is that the product of two irreducible
characters is character, but not necessarily irreducible; it may instead split as a non-
negative integer combination of irreducible characters. The structure constants of
this multiplication rule are extremely rich combinatorially, and are important in a
vast number of applications, but will not be discussed here.
Department of Mathematics, UCLA, Los Angeles, CA 90024
E-mail address: [email protected]