Haar Measure on Compact Groups

∗ †
Tsz Ho Lee Dateng Lin

1 Haar measure
1.1 Introduction
One of the most useful properties of the Lebesgue integral is invariance under translations and
rotations. That is, if a ∈ Rn , r ∈ Rn×n and f is any Lebesgue integrable function in Rn , then
Z Z
f (x)dx = f (rx + a)dx.
Rn Rn

The notion of Haar measure is a generalization of this example. It turns out that in any compact
group (more generally, locally compact group) G, there exists a measure m such that
Z Z
f (x)dm = f (bx)dm.
G G

for any integrable function f on G and any b ∈ G.

This measure was introduced by Alfréd Haar, a Hungarian mathematician, in 1933. He
proved that there exists an invariant measure on any separable compact group. Later, by using
Haar’s result, von Neumann proved that a compact locally Euclidean group is a Lie group, which
is a special case of the Hilbert’s fifth problem.
Here we will present a proof for the existence and uniqueness of Haar measure on compact
groups based on the Fixed Point Theorem from functional analysis, which is different from the
method using the techniques from measure theory (see [6]).

1.2 Topological group

Definition 1 A topological group is a group G which at the same time is a topological space
such that the mapping (x, y) 7→ xy −1 of G × G into G is continuous.

For each a ∈ G, it is easy to check that the mappings x 7→ ax, x 7→ xa and x 7→ x−1 are
homeomorphisms of G onto G. So the topology of G is therefore completely determined by any
local base at the identity e of G.

Definition 2 A compact group is a topological group which is a compact space.

∗
Department of Mathematics, The Chinese University of Hong Kong, Shatin, Hong Kong, China
([email protected]).
†
Department of Mathematics, The Chinese University of Hong Kong, Shatin, Hong Kong, China
([email protected]).

1
 Here we present several examples of topological groups.

• Euclidean space Rn with addition and the usual topology is a topological group, but it is
not compact.

• T = {z ∈ C : |z| = 1} = {eiθ : 0 ≤ θ < 2π} with multiplication in C and the usual

topology in C is a compact group. It is called the circle group.

• SO(n) = {A : A ∈ Rn×n , det(A) = 1, AA| = I}, with matrix multiplication, and topology
2
given by the Euclidean metric on Rn , is a compact group. It is called the special orthogonal
group.

1.3 Existence and uniqueness of Haar measure

Definition 3 If ∆ is any Borel set of G and x ∈ G, let ∆x = {ax : a ∈ ∆}, x∆ = {xa : a ∈ ∆}
and ∆−1 = {a−1 : a ∈ ∆}.

Definition 4 If f is any function defined on a topological group G, then for every s ∈ G, its
left translates Ls f is defined by (Ls f )(x) = f (sx), x ∈ G; similarly, its right translates Rs f is
defined by (Rs f )(x) = f (xs), x ∈ G. Define f # (x) = f (x−1 ), x ∈ G.

Definition 5 Take a ∈ G, a complex function f on G is said to be continuous at a if to every

 > 0 corresponds a neighborhood Wa of e (the identity of G) such that

|f (x) − f (a)| < ,

if xa−1 ∈ Wa .

As usual, C(G) denotes the Banach space of all complex continuous functions on G, with
the supremum norm.

Definition 6 If K is a convex set, Y is a vector space, and T : K 7→ Y satisfies

T ((1 − λ)x + λy) = (1 − λ)T (x) + λT (y)

whenever x, y ∈ K, 0 < λ < 1. Then T is called an affine map.

Theorem 1 (Kakutani’s Fixed Point Theorem) Let K be a nonempty compact convex set in
a locally convex space X, and G be an equicontinuous group of affine maps taking K into K.
Then G has a common fixed point in K.

A proof of this theorem can be found in [11]. A more general conclusion is the Ryll-
Nardzewski Fixed Point Theorem. For detail see [1].

Proposition 1 Let G be a compact group, suppose f ∈ C(G), and define HL (f ) to be the convex
hull of the set of all left translates of f . Then

(a) s 7→ Ls f is a continuous map from G into C(G);

2
 (b) The closure of HL (f ) is compact in C(G).

Proof: Fix  > 0. Since f is continuous, there corresponds to each a ∈ G a neighborhood Wa

of e such that |f (x) − f (a)| <  if xa−1 ∈ Wa . The continuity of the group operations gives
neighborhoods Va of e which satisfy Va−1 Va ⊂ Wa . Since G is compact, there is a finite set
A ⊂ G such that
[
G= Va a.
a∈A

Put
\
V = Va .
a∈A

Choose x, y ∈ G, so that yx−1 ∈ V , and choose a ∈ A so that ya−1 ∈ Va . Then |f (y) − f (a)| < ,
and since xa−1 = (xy −1 )(ya−1 ) ∈ V −1 Va ⊂ Wa , we also have |f (x) − f (a)| < .
Thus |f (x) − f (y)| < 2 whenever yx−1 ∈ V .
For any s ∈ G, (ys)(xs)−1 = yx−1 . Hence yx−1 ∈ V implies that |f (xs) − f (ys)| < 2. This
is just another way of saying that
kLx f − Ly f k ≤ 2

whenever y lies in the neighborhood V x of x. This proves (a).

As a consequence of (a), {Lx f : x ∈ G} is compact in the Banach space C(G). Since if X is
a Banach space and K ⊂ X is compact, then the closure of the convex hull of K is compact, so
(b) follows.
Note that regarding to the right translates of f , we have similar conclusions.

Theorem 2 If G is a compact group, then there is a unique positive regular Borel measure m
on G such that

(a) m(G) = 1;

(b) If U is a nonempty open subset of G, then m(U ) > 0;

(c) If ∆ is any Borel set of G and s ∈ G, then m(∆) = m(s∆) = m(∆s) = m(∆−1 ).

The measure m is called the Haar measure for G. If G is only locally compact, i.e., it is a
topological group which is a locally compact space, then it is also true that there is a positive
Borel measure m on G satisfying (b) and such that m(∆) = m(x∆) for all x in G and every
Borel subset ∆ of G. However, it is not necessarily true that m(∆) = m(∆x), a counterexample
is given by [10].
By using the Riesz Representation Theorem for representing bounded linear functionals on
C(G), Theorem 2 is equivalent to the following theorem.

Theorem 3 If G is a compact group, then there exists a unique positive linear functional I :
C(G) 7→ F such that

(a) I(id) = 1, where id is the identity mapping;

3
 (b) If f ∈ C(G), f ≥ 0 and f 6= 0, then I(f ) > 0;

(c) If f ∈ C(G) and s ∈ G, then I(f ) = I(Ls f ) = I(Rs f ) = I(f # ).

Proof: The operators Ls satisfy Ls Lt = Lts , and it is easy to verify that (Ls Lt )Lu = Ls (Lt Lu ),
Ls Le = Le Ls = Ls , Ls Ls−1 = Ls−1 Ls = Le . Also it is easy to check that each Ls is a linear
isometry of C(G) onto itself. So {Ls : s ∈ G} is an equicontinuous group of affine maps on C(G).
If f ∈ C(G), let Kf be the closure of HL (f ), by Proposition 1, Kf is compact. It is obvious that
Ls (Kf ) = Kf for every s ∈ G. By Kakutani’s Fixed Point Theorem, Kf contains a function φ
such that Ls φ = φ for every s ∈ G. In particular, φ(s) = φ(e), so that φ is a constant. By the
definition of Kf , this constant can be uniformly approximated by functions in HL (f ).
So far we have proved that to each f ∈ C(G) corresponds at least one constant c which can
be uniformly approximated on G by convex combinations of left translates of f . Likewise, there
is a constant c0 which bears the same relation to the right translates of f . We claim that c0 = c.
To prove this, pick  > 0. There exist finite sets {ai }ni=1
1
and {bj }nj=1
2
in G, and there exist
n
P1 n
P 2
numbers αi > 0, βj > 0, with αi = βj = 1, such that for x ∈ G,
i=1 j=1
 n1

 X 
c − α f (a i  < ,
x) (1)
 
 i
 
i=1

and  
 n2 
 0 X 
c − βj f (xb j ) < . (2)
 
 j=1 

Put x = bj in (1); multiply (1) by βj , and add with respect to j. The result is
 

Xn2 Xn1 
 
c − βj αi f (ai bj ) < . (3)

 j=1 i=1 

Put x = ai in (2), multiply (2) by αi , and add with respect to i, to obtain

 
 n1 Xn2 
 0 X 
c − α i β j f (a i bj ) < . (4)
 
 i=1 j=1 

Now (3) and (4) imply that c0 = c.

So it follows that to each f ∈ C(G) corresponds to a unique number, which we shall write
I(f ), which can be uniformly approximated by convex combinations of left translates of f and
right translates of f at the same time. Then the following properties of I are obvious:

I(f ) ≥ 0, if f ≥ 0,

I(id) = 1,

I(f ) = I(Ls f ) = I(Rs f ), s ∈ G,

I(αf ) = αI(f ), if α is a scalar.

4
 Next we prove that I(f + g) = I(f ) + I(g).
Pick  > 0. Then there exists some finite set {di }ni=1
3
⊂ G and some numbers γi , i =
n3
P
1, 2, . . . , n3 with γi = 1, such that for x ∈ G,
i=1
 n3

 X 
I(f ) − γi f (di x) < . (5)
 
 
i=1

Define
n3
X
h(x) = γi g(di x). (6)
i=1
Then h ∈ Kg , hence Kh ⊂ Kg , and since each of these sets contains a unique constant function,
we have I(h) = I(g). Hence there is a finite set {lj }nj=1
4
⊂ G, and there are numbers δj > 0 with
n
P4
δj = 1, such that for x ∈ G,
j=1
 
 Xn4 
 
I(g) − δ j h(lj x) < ,
 
 j=1 

by (6), this gives  


Xn4 Xn3 
 
I(g) − γi δj g(di lj x) < .

 j=1 i=1 
Replace x by lj x in (5), multiply (5) by δj , and add with respect to j, to obtain
 

Xn4 Xn3 
 
I(f ) − γi δj f (di lj x) < .

 j=1 i=1 

Thus  
 Xn4 Xn3 
 
I(f ) + I(g) − γ δ
i j (f + g)(d l x)
i j  < 2.


 j=1 i=1 
n4 P
P n3
Since γi δj = 1, we obtain I(f + g) = I(f ) + I(g).
j=1 i=1
Now we prove (b).
Suppose f ∈ C(G), f ≥ 0 and f 6= 0. Then there is a λ > 0 such that U = {x ∈ G : f (x) > λ}
S
is nonempty. Since U is open, U x is open for x ∈ G, and G = U x, but G is compact, so
x∈G
n
{xk }nk=1
S
there exists ⊂ G such that G = U xk .
k=1
n
Define gk (x) = f (xx−1
P
k ) and put g = gk , then g ∈ C(G) and
k=1
n
X n
X
I(g) = I(gk ) = I(Rx−1 f ) = nI(f ).
k
k=1 k=1

However, for any x in G there is an xk such that xx−1 −1

k ∈ U ; hence g(x) ≥ gk (x) = f (xxk ) > λ.
Thus
1 λ
I(f ) = I(g) ≥ > 0.
n n

5
This proves (b).
By the Riesz Representation Theorem, there exists a regular Borel probability measure m
which satisfies Z
I(f ) = f dm, for any f ∈ C(G).
G
To prove uniqueness, let f ∈ C(G), J be another linear positive functional on C(G) which
R
satisfies (a), (b) and (c), and J = G f dµ, then
Z Z Z  Z Z  Z
f dµ = f (yx)dµ(x) dm(y) = f (yx)dm(y) dµ(x) = f dm.
G G G G G G

Hence J = I.
Last we ought to prove I(f ) = I(f # ). Since
Z Z  Z Z 
−1 # −1
f (yx )dm(y) dm(x) = f (xy )dm(x) dm(y),
G G G G

and the two inner integrals are independent of x and y, respectively. Hence I(f ) = I(f # ).

2 Haar Distributed Matrices

2.1 Motivation
Let us start with a long-time mathematical problem: What is the number of fixed points after
permuting a series of numbers randomly? For instance, the following permutation

1 2 3 4 5 6 7
↓
3 2 6 1 5 7 4

has 2 fixed numbers 2 and 5. Obviously, the number of fixed points is a random variable with
respect to the random permutation. Therefore we need to find a way to express the randomness
of a permutation. In fact, we notice that the randomness will be revealed if we write the above
example into the following form:
    
3 0 0 1 0 0 0 0 1
2 0 1 0 0 0 0 0 2
    
    
6 0 0 0 0 0 1 0 3
    
1 = 1 0 0 0 0 0 0 4 .
    
    
5 0 0 0 0 1 0 0 5
    
7 0 0 0 0 0 0 1  6
    
  
4 0 0 0 1 0 0 0 7

We now see that the randomness of the permutation actually comes from this 7 × 7 matrix,
which is called the permutation matrix. Thus, the idea of random matrices is required. Simply
put, a random matrix is just a random variable taking values in the space of matrices. However,
we prefer to study random matrices just like the way we study real-valued random variables

6
with probability distribution. Hence we need to define the probability distribution for random
matrices.
In the permutation problem, it is easy to verify that all permutation matrices form a compact
group. More generally, the permutation group belongs to a larger compact group known as the
orthogonal group. From the previous section, we know that there exists a unique Haar measure
defined on such groups. The Haar measure is normalized to one and invariant under group
operation. Thus, it is a natural candidate for a probability measure. This inspires us to connect
the random matrices and the Haar measure.
The reason we study random matrices is that interesting information can be extracted from
their characteristics such as traces, eigenvalues, maxima, etc.. In the permutation problem,
the trace (summation of all diagonal entries) of the permutation matrix gives the number of
fixed points. Previous studies have shown that the trace of a random permutation matrix,
which is in fact a real random variable, follows the Poisson(1) distribution [2]. Other than the
trace, the eigenvalue distribution of a random matrix also appears in different applications like
telecommunications, the zeros of Riemann’s zeta function [3], multivariate statistics [4], and
problems from physics [8].
In the following, we first review the orthogonal group and prove that it is a compact group.
Then we focus on the Haar measure defined on it.

2.2 Orthogonal group

First of all, we introduce some definitions and notations. Let R be the set of real numbers, Rn×n
be the set of n × n real matrices, I be the identity matrix, and X | be the transpose of a real
matrix X ∈ Rn×n . Then

On = {Q = (qij )n×n : Q| Q = I, and qij ∈ R, 1 ≤ i, j ≤ n}

contains all the orthogonal matrices. Our aim here is to show that On is a compact group.
Remark: For simplicity, we only consider the orthogonal group in this context. In general, we
can also generalize the results to the unitary group. Let C be the set of complex numbers, Cn×n
be the set of n × n complex matrices, and X ∗ be the conjugate transpose of a complex matrix
X ∈ Cn×n . The following set

Un = {X = (xij )n×n : X ∗ X = I, and xij ∈ C, 1 ≤ i, j ≤ n}

is known as the unitary group. Furthermore, one can also study the sympletic group [2] which
comes from mechanics and elsewhere. 
We now proceed to show that On is a compact group in three steps.
Claim 1: On is a group.
The group operation is just the matrix multiplication. Given Q1 , Q2 ∈ On , we have

(Q1 Q2 )| (Q1 Q2 ) = Q|2 Q|1 Q1 Q2 = Q|2 Q2 = I.

Thus Q1 Q2 ∈ On . The identity element is the identity matrix I ∈ On . The inverse element of
Q ∈ On is the inverse matrix Q−1 , which happens to be Q| in this case, and Q−1 ∈ On because

(Q−1 )| Q−1 = (Q| )−1 Q| = I.

7
 Claim 2: On is a topological group.
Recall that Rn×n is a real vector space of n×n matrices. Let the Frobenius norm of X ∈ Rn×n
be  1
n 2
X
2
kXk =  |xij | .
i,j=1

It is well known that the Frobenius norm has the following properties:

(a) kX | k = kXk, X ∈ Rn×n ;

(b) kXY k ≤ kXkkY k, X, Y ∈ Rn×n ;

(c) kQXk = kXk = kXQk, X ∈ Rn×n , Q ∈ On ;

√
(d) kQk = kIk = n, Q ∈ On .
2
Also the Frobenius norm is just the usual norm in Rn . Thus, the topology of On is inherited
from Rn ∼
2
= Rn×n .
Using the following estimation

k(X + dX)(Y + dY ) − XY k ≤ kY kkdXk + kXkkdY k + kdXkkdY k,

we conclude that
(X, Y ) 7→ XY, X, Y ∈ Rn×n

is a continuous function from Rn×n × Rn×n to Rn×n .

Note that the inverse is simply the transpose in On . By

k(X + dX)| − X | k = kdX | k = kdXk,

we know that
X 7→ X −1 , X ∈ Rn×n

is a continuous function from Rn×n to Rn×n .

Claim 3: On is a compact group.
First, we have the following proposition.

Proposition 2 The determinant function

X 7→ det(X), X ∈ Rn×n

is a continuous function from Rn×n to R.

Proof: The determinant can be regarded as a polynomial of n2 real variables, hence it is a

continuous function from Rn×n to R.
Now note that
On = {X ∈ Rn×n : det(X) = 1, −1}.

By Proposition 2, the determinant is a continuous function, therefore the set On is closed.

√
Moreover, On is bounded because kQk = n for Q ∈ On . In conclusion, On is compact.

8
2.3 Haar measure on orthogonal group
From Theorem 2, we know that On has an invariant measure µ, i.e., a measure on the Borel sets
B of On such that for every B, and Q ∈ On ,

µ(B) = µ(QB) = µ(BQ) and µ(On ) = 1.

This measure is known as the Haar measure. Note that the Haar measure is normalized to one,
hence it is a natural choice for a probability measure on On ,

Prob{X ∈ B} = µ(B).

Moreover, it is invariant under group operation. Technically speaking, the Haar measure is
analogue to the uniform density on a finite interval in the sense that it equally weighs different
regions of On [7].
We illustrate this by two simplest examples: O1 and U1 . The group O1 is

O1 = {−1, 1}.

Therefore the Haar measure is given by

1 1
µ({−1}) = and µ({1}) = .
2 2
It is a counting measure normalized to one, and equivalent to the discrete uniform distribution
of coin tossing.
The group U1 is actually the circle group:

U1 = {exp(iθ) : θ ∈ [0, 2π)}.

The Haar measure on U1 is a probability measure with constant density 1/(2π) such that
µ(U1 ) = 1. Roughly speaking, it returns the ratio of the arc length to 2π. Also note that it acts
like the standard Lebesgue measure because the group operation is a translation in this case. In
reality, we have
1
µ(B) = m(φ−1 (B)),
2π
where B is a Borel set in U1 , m is the restriction of Lebesgue measure to the interval [0, 2π),
and φ is a bijection that maps [0, 2π) to U1 .
Next we consider the Haar measure on On from a “construction” viewpoint. We first intro-
duce the concept of real Ginibre ensemble.

Definition 7 The set of all X ∈ Rn×n , whose entries xij are independent identically distributed
(i.i.d.) standard normal random variables, is called the real Ginibre ensemble.

Let X = (xij ) ∈ Rn×n be a random matrix taken from the real Ginibre ensemble. In other
words, the probability density function of xij is

1 2
g(xij ) = √ e−xij /2 .
2π

9
Since all the matrix entries are independent, the joint probability density function can be re-
garded as a function of X and equals to
 
n n
!
2
xij
1 Y 1 1 X
G(X) = √
n2 exp − = √
n2 exp − x2ij 
2π 2 2π 2
i,j=1 i,j=1
 
1 1
= √
n2 exp − Tr(X | X) ,
2π 2

where
n
X
Tr(X) = xii
i=1
is the trace of X. Now we use G(X) to construct a measure

dγ(X) = G(X)dX,

where
n
Y
dX = dxij
i,j=1

is the Lebesgue measure on Rn×n .

Remark: The measure γ is also known as the Gaussian measure on Rn×n . Note that the joint
probability density function G(X) is still a probability density. Therefore we have
Z Z
γ(Rn×n ) = dγ(X) = G(X)dX = 1.
Rn×n Rn×n

Lemma 1 The measure γ is invariant under orthogonal transformations, i.e., for any Q ∈ On ,

dγ(QX) = dγ(X) = dγ(XQ).

Proof: Suppose Q is an orthogonal matrix, we have

   
1 1 | | 1 1 |
G(QX) = √
n2 exp − Tr(X Q QX) = √
n2 exp − Tr(X X) = G(X).
2π 2 2π 2

We then show that the Jacobian of the map

Λ : X 7→ QX

is one. Since the map Λ is isomorphic to

Λ = Q ⊕ · · · ⊕ Q,
| {z }
n times

it follows that Λ is in fact an n2 × n2 orthogonal matrix, therefore | det(Λ)| = 1. Together we

have
dγ(QX) = dγ(X).
The proof of the right invariance is identical.
There is another important property regarding the real Ginibre ensemble.

10
Lemma 2 (Proposition 7.1 [4]) If X is a random matrix from the real Ginibre ensemble, then
X is invertible with probability 1.
Suppose we have chosen the first n − 1 columns of X from Rn , which span a linear subspace.
Note that a proper subspace of Rn has measure zero. Therefore the probability of the last vector
lying in the same subspace is zero. Hence all the singular matrices from a measure zero set of
Rn×n , and X is invertible with probability 1.
Let Gln be the set of all nonsingular matrices in Rn×n , and it is easy to check that Gln itself
is a group with On being its subgroup. Recall that any nonsingular matrix can be turned into
an orthogonal matrix under the Gram-Schmidt process. Suppose X ∈ Rn×n is invertible and
{x1 , x2 , . . . , xn } are its columns. Set y1 = x1 and
j−1 |
X xj yi
yj = x j − yi , j = 2, 3, . . . , n.
yi| yi
i=1
1
Let qi = yi /(yi| yi ) 2 .
Then Q = [q1 , q2 , · · · , qn ] is an orthogonal matrix. Therefore, the Gram-
Schmidt process defines a map Φ from Gln to On , which has the following properties.
Lemma 3 Suppose Φ is the map
Φ : Gln → On
defined by the Gram-Schmidt process. Let B ⊂ On ⊂ Gln be a Borel set . Then
(a) Φ is surjective;

(b) Φ(QB) = QΦ(B), Q ∈ On ;

(c) Φ−1 (QB) = QΦ−1 (B), Q ∈ On .

By Lemma 1, we know that the measure γ on the real Ginibre ensemble is invariant under
On . We aim to show that under the map Φ, the derived measure will still be invariant and equal
to the Haar measure.

Theorem 4 (Proposition 7.2 [4]) Given a matrix X ∈ Rn×n in the real Ginibre ensemble with
the Gaussian measure γ. Let Φ be the map of Gram-Schmidt process from Gln to On . Then the
pushforward measure of γ, which is defined as
µ(B) = γ(Φ−1 (B))
for any Borel set B of On , is the invariant Haar measure on On .

Proof: By Lemmas 1 and 3, we have

µ(QB) = γ(Φ−1 (QB)) = γ(QΦ−1 (B)) = γ(Φ−1 (B)) = µ(B),
for all Q ∈ On . Thus, µ is left invariant. The proof of the right invariance is similar. Also we
have
µ(On ) = γ(Φ−1 (On )) = γ(Gln ) = γ(Rn×n ) = 1.
The third equality above follows from Lemma 2. Hence the pushforward measure µ is the Haar
measure on On .
Numerically speaking, Theorem 4 provides a way to generate a random orthogonal matrix
which is Haar distributed.

11
References
[1] J.B. Conway, A Course in Functional Analysis, Springer-Verlag Inc., New York, 1990.

[2] P. Diaconis and M.Shahshahani, On the eigenvalues of random matrices, J. Appl.

Probab., 31A (1994), pp. 49–62.

[3] P. Diaconis, Patterns in eigenvalues: The 70th Josiah Willard Gibbs Lecture, Bull. Amer.
Math. Soc., 40 (2003), pp. 155–178.

[4] M. Eaton, Multivariate Statistics, Wiley, New York, 1983.

[5] M. Eaton, Group-Invariance Applications in Statistics, Regional Conference Series in

Probability and Statistics 1, Institute of Mathematical Statistics, Hayward, CA, 1989.

[6] P. R. Halmos, Measure Theory, Springer-Verlag, New York, 1974.

[7] F. Mezzadri, How to generate random matrices from the classical compact groups, Notices
Amer. Math. Soc., 54 (2007), pp. 592–604.

[8] M. Mehta, Random Matrices, Elsevier, Amsterdam, 2004.

[9] L. Nachbin, The Haar Integral, D. Van Nostrand, Princeton, New Jersey, 1965.

[10] J. von Neumann, Invariant Measures, American Mathematical Society, Providence,

Rhode Island, 1999.

[11] W. Rudin, Functional Analysis, McGraw-Hill Inc., New York, 1991.

12