The Peter-Weyl Theorem for Compact Groups

The following notes are from a series of lectures I gave at Dartmouth College in the
summer of 1989. The general outline is provided by an introductory section in 1] but with
considerable detail added by myself. The mistakes of course are mine.
Dana P. Williams
Hanover, June 1991
x1 Preliminaries.
We begin with some warm-up exercises on locally compact groups
a.k.a., a long series of
denitions! At this point, G is meant to be an arbitrary locally compact group.
Denition 1: A (unitary) representation of G is a continuous homomorphism  from G
to the unitary group U(H ) on a (complex) Hilbert space H equipted with the strong
operator topology.
Remark 2: The condition that  be continuous merely means that g 7! (g) is contin-
uous from G to H for each  2 H . There are many equivalent conditions: the weakest
I'm aware of is to insist that g 7! h(g) i be Borel for each   2 H .
Example 3: Let H = L2(G). The left regular representation  : G ! B(H) is given by
(g)f (t) = f (g;1 t).
Denition 4: Two representations 1 and 2 are said to be (unitarily) equivalent if there
is a unitary operator U : H 1 ! H 2 such that 1 (g) = U2 (g)U  for all g 2 G. In this
event, we write 1  = 2 and let ] denote the (unitary) equivalence class of .
If  : G ! U(H ) is a representation of G, then we'll write d for the dimension (in
f 0 1 2 : : :  1g) of H . Fortunately, whenever 1 = 2 , then it is clear that d 1 = d 2 .
Thus we will often write d ] to denote the dimension of each representation in the same
equivalence class as .
Denition 5: A non-zero representation  is called irreducible if H has no no-trivial
closed invariant subspaces
Remark 6: If d < 1, then the word \closed" in redundent in the above denition.
Denition 7: The symbol Gb is used to denote the collection of equivalence classes of
irreducible representations of G.
Example 8: If G is abelian, then every irreducible representation is one-dimensional. In
particular, Gb coincides with the character group of G.
1
Denition 9: If  and  are representations of G, then    denotes the representation
on H  H dened by
  (g)( ) = ((g) (g) ):
If n 2 Z+  f1g, then n   = ni=1.
Denition 10: If  is a representation of G, then
(G)0 = f A 2 B(H ) : A(g) = (g)A for all g 2 G g:
Remark 11: Since  is unitary, (G)0 is a self-adjoint. In fact it is not hard to check that
(G)0 is a -subalgebra of B(H ) which is closed in the weak operator topology.
Theorem 12: Suppose that  is a representation of G. Then the following are equivalent.
(1)  is irreducible.
(2) (G)0 = C I .
(3) Every non-zero  2 H is cyclic for  (i.e., (G) = spanf (g) : g 2 G g = H ).
Proof: Suppose that  is irreducible. Let A 2 (G)0 and suppose for the moment that
A is normal: A A = AA . Then the norm closed unital *-subalgebra generated by A|
that is the C -algebra C (I A)|is contained in (G)0 . Since A is normal, C (I A) is
commutative and is isomorphic to C
(A) by the spectral theorem. If
(A) 6= f pt g,
; 

then we can nd nonzero self-adjoint operators B1 and B2 in C (I A) so that B1B2 =

B2B1 = 0. Thus hB1  B2i = 0 for all   2 H . In particular, the closures V1 and V2 of
the ranges of B1 and B2, respectively, are closed, non-zero, orthogonal, invariant subspaces
for . This contradics the irreducibility of 
therefore
(A) must be a single point and
therefore A = I for some  2 C . For a general A 2 (G)0 , we apply the above reasoning
to AA and A A. Thus, we have AA = I and A A = I with   > 0 (since A 6= 0).
Since A = A(A A) = A, we have  =  and A is normal. This shows that (1) implies
(2).
Since V = (G) is a closed, non-zero, invarinat subspace, in order to show that
(2) implies (3) it will suce to show that the orthogonal projection P onto any closed,
non-zero, invarinat subspace V is in (G)0 . But since  is unitary, V ? is also invariant.
Thus if   2 H , then
hP(g) i = h(g) Pi
= h(g)P + (g)(I ; P ) Pi
which, since (g)(I ; P ) 2 V ? and P 2 V ,
2
 = h(g)P Pi
= hP(g)P i
= h(g)P i
because (g)P 2 V .
That (3) implies (1) is clear.

3
x2 The Peter-Weyl Theorem.
Now we'll specialize to compact groups G. Compact groups are characterized by the fact
that any Haar measure  on G satises (G) < 1. It is customary to normalize Haar
measure on a compact group by choosing the unique measure such that (G) = 1. Since
there is now no possibility of confusion, I'll simply write
Z
f (g) dg
G

for the integral of f 2 L1(G).

Now suppose that  is a nite dimensional representation of G. If b = f e1  : : :  ed g
is an orthonormal basis for H , then for each g 2 G the matrix of (g) with respect to b
has ij th coordinate h(g)ej  ei i. The function ij (g) = hei  (g)ej i is called a coordinate
function for . (It will be convenient to use this convention|even though ij is the
complex conjugate of what you might expect. This usage and terminolgy will be justied,
somewhat, by Remark 14 below.) Notice that ij 2 C (G). We'll write EG , or just E when
no confusion is likely to arise, for the linear span of the all the functions (g) = h (g)i,
where  ranges over all irreducible representations of G and  and  range over H .
Since every nite dimensional representation is the direct sum of irreducibles, notice that
(g) = h (g)i denes an element of E for every nite dimensional representation |
irreducible or not.
Remark 13: When G is abelian, and sometimes in general, the functions in E are called
the trigonometric polynomials. The motivation for this probably comes from the case
where G = T = R=2Z. Then each f 2 ET has the form
nX
=k k
X
f () = cn exp(in) = dn cos(n) + bn sin(n):
n=;k n=0

Remark 14: It is clear that E is self-adjoint

that is, if f 2 E, then so is f  2 E,
where f  (g) = f (g;1 ). This is because matrix coecients are themselves self-adjoint:
ij = (ij ) . It is also true that if f 2 E, then so is f, where f(g) = f (g;1 ). To see
this we need to introduce the conjugate Hilbert space H e to a given Hilbert space H.
The space H e coincides with H as an additive group. If j : H ! H e denotes the identity
map, then the Hilbert space structure on H e is given by the formulas j ( ) = j ( ), and

4
hj() j()iHe = h iH. If  is a given representation of G on H, then we can dene a
; 
representation e on He in the obvious way:  e(g )j ( ) = j  (g ) . The assertion follows
from the fact that ije (g) = ij (g;1). Since ij (g;1 ) = ji (g), it is reasonable to call
g 7! h (g)i a coordinate function.

Denition 15: Let Mn be the n  n complex matrices. If A = (aij ) 2 Mn , then the

Hilbert-Schmidt norm of A is

k(aij )k : : = X jaij j :
HS
2
ij

Remark 16: If A 2 Mn, then kAk : : = tr(A A). In particular, if B = U  AU for some
HS
unitary matrix U , then kAkH:S: = kBkH:S: . It follows that if  is nite dimensional, then
k(g)kH:S: is well dened and depends only on ].
Our object here is to prove the following theorem known as the Peter-Weyl Theo-
rem.
Theorem 17: Let G be a compact group.
(1) Every irreducible representation of G is nite dimensional.
(2) If  is the left-regular representation of G, then

 d 
M
=
 ]2G
b

(3) Given g 2 G, there is a ] 2 Gb such that (g) 6= I .

(4) E is dense in C (G) (and hence in Lp(G) for 1 p < 1).
(5) If f 2 L2(G), then

kf k2
2 =
X
d  tr (f )(f ) ) =
; X
d  k(f )k : : :
HS
 ]2G
b  ]2G
b

We'll need the following preliminary results, some of which may be of interest by
themselves.
5
Lemma 18: Let 1 and 2 be representations of a locally compact group G. If A :
H1 ! H2 is a bounded linear operator satisfying A1 (g) = 2 (g)A for all g 2 G, then
A A1 (g) = 1 (g)A A for all g 2 G.
Proof: One computes as follows:
hA A (g) i = hA (g) Ai
1 1

= h (g)A Ai
2

= hA  (g; )Ai 2

1

= hA A (g; )i 1

1

= h (g)A A i
1

Lemma 19: Suppose that 1 and 2 are representations of a locally compact group G,
that 1 is irreducible, and that A : H1 ! H2 is any non-zero bounded linear operator such
that A1 (g) = 2 (g)A for all g 2 G. Then AH1 is a closed invariant subspace for 2 , and
1 
= 2 jAH1 , where 2 jAH1 is the subrepresentation of 2 corresponding to AH1.
Proof: By Lemma 18, A A 2 1(G)0 . By Theorem 12, A A = I . Thus B = ; A is 1
2

an isometery, and hence a unitary from H1 onto AH1. (Note that AH1 is actually closed
in H2 since A is a multiple of an isometery.) Again Lemma 18 shows that BB 2 2(G)0 .
Since BB is the orthogonal projection onto BH1 = AH1, it follows that AH1 is invariant
for 2 and the assertion follows.
The next result is crucial, and depends heavily on the fact that G is compact.
Proposition 20: Suppose that 1 and 2 are irreducible representations of a compact
group G. Fix orthonormal bases f eik gdk=1i for Hi, and put ikl(g) = heik  i (g)eil i.
(1) If 1 
6 2 , then
= Z
1ij (g)2jl(g) dg = 0
G
for all 1 i j d 1 , and 1 k l d 2 .
(2) If 1 is nite dimensional, then
Z
1ij (g)1kl (g) dg = ik jl d1 
G 1

6
 for all 1 i j k l d 1 .
Proof: Let B be any bounded linear operator from H1 to H2. Then we can dene
Z
A= 2 (g)B1 (g;1) ds:
G
Then Z
A1 (r) = 2 (g)B1 (g;1 r) dg
G
Z
= 2 (rg)B1 (g;1) dg = 2 (r)A:
G
If 1 
6 2, then A = 0 by Lemma 19. Now suppose that B = Bij is the rank-one operator
=
dened by Bij () = h e1j ie2l . Then,
Z
0 = hAei  ek i =
1 hBij  (g; )ei   (g; )ek i dg
2
1
1 1
2
1 2
G
Z
= h (g; )ei  ej ihel   (g; )ek i dg
1
1 1 1 2
2
1 2
G
Z
= 1ij (g)2kl (g) ds:
G
This proves (1).
Now assume that d 1 < 1. By the above and Lemma 19,
Z
A= 1 (g)B1 (g;1 ) dg = I
G
for any B. Taking traces,
d1
tr(A) =
X
hAek  ek i
1 1

k=1
d1
Z X
=
G k=1
hB (g; )ek   (g; )ek i dg
1
1 1
1
1 1

Z
= tr(B) dg = tr(B):
G
Since tr(A) = tr(I ) = d 1 and tr(Bjl ) = jl , we have  = jl d11 when B = Bjl. On the
other hand, Z
1ij (g)1kl(g) dg = hAe1i  e1k i = he1i  e1k i = ik jl d1 :
G 1

7
Proof of Theorem 17: Let Ef be the subset of E consisting of the collection of matrix
coecients of the form (g) = h (g)i for  irreducible and d < 1. The rst part of
the proof will consist of showing that it suces to show that Ef is dense in C (G). Since
C (G) is dense in L2(G), it follows that L2(G) = H has an orthonormal basis consisting of
normalized matrix coecients d 2 ij (g) = d 2 hei  (g)ej i (where, of course, f e1  : : :  ed g
1 1

denotes an orthonormal basis for H ).

In fact, if  is any irreducible representation of G and if f e g2A is an orthonormal ba-
sis for H , then it follows from Proposition 20 that  ? ij , where  (g) = he  (g)e i
and d < 1. Thus we must have  = 0, and (1) follows.
p
If d < 1, then Proposition 20 shows that the d2 functions f d ij g are an or-
thonormal basis for a subspace H ] of L2(G). Now observe that
ij (g;1t) = h(g)ei  (t)ej i
d
=
X
h(g)ei  ek ihek  (t)ej i
k=1
Xd
= ik (g;1 )kj (t):
k=1
Therefore if, for each 1 j d , we dene Aj : H !H  ] by Aj ei = ij , then
(g)(Aei )(t) = ij (g;1t)
d
X
= ik (g;1)kj (t)
k=1
d
=
X
hei  (g; )ek iAj ek (t)
1

k=1
d
X 
=A h(g)ei  ek iek (t)
k=1
; 
= A (g)ei (t):
That is, A intertwines the irreducible representation  and the subrepresentation jH ]j ,
where H ]j = spanf 1j  2j  : : :  d j g. Therefore Lemma 19 implies that  
= jH ]j ,
and thus, jH ] 
= d  . Since we're assuming that Ef is dense, we have
H
M
= H ] 
 ]2G
b

8
and we have proved (2).
We have now shown that f d2 ] ij g ]2Gb forms an orthonormal basis for L2(G). (Under
1

the assumption that Ef is dense in C (G).) Thus, if f 2 L2(G), we can write

X ;  1
f= c ] i j d2 ] ij :
Furthermore,
d ]
kf k2
2 =
X X
jc; ] i j j :
2

b ij =1
 ]2G

Now (5) follows from the fact that

Z
;  1
c ] i j = f (g)d2 ] ij (g) dg
G
Z
f (g)h(g)ej  ei i dg
1
= d 2
]
G
= d2 ] h(f )ej  ei i dg
1

1  
= d2 ] (f ) ij :

Of course, (3) follows from (4) (otherwise, E wouldn't separate e and g), so it only
remains to prove that Ef is dense in C (G). Towards this end, we need to recall some basic
facts about so-called Hilbert-Schmidt operators. If (X M ) is a measure space and if
K 2 L2(X  X   ), then we can dene a bounded linear operator T : L2(X ) ! L2(X )
by Z
Tf (x) = K (x y)f (y) dy:
X
It is not hard to see that T is self-adjoint if K (x y) = K (y x). An operator of this
form is called a Hilbert-Schmidt operator and all such operators are self-adjoint compact
operators(1). In particular, each eigenspace
H = f f 2 L2(X ) : Tf = f g
(1) In our case, we'll only be interested in the case where X = G,  is normalized Haar measure, and K
is continuous. Then the Stone-Weierstrass Theorem implies that there are functions i 2 C (G) such
that
K (x y) =
X
i i (x)i (y) ()
i2I

9
is nite dimensional and there is an orthonormal sequence f i g of eigenvectors with eigen-
values i so that every f 2 L2(X ) can be written uniquely as
X
f= cii + 0
where T0 = 0 and ci = hf i i.
Our interest in such operators is as follows. Let k be any element of C (G) which
satises k(g) = k(g;1). Therefore
Z Z
f k(g) = f (r)k(r;1 g) dr = K (g r)f (r) dr
G G
where K (g r) = k(r;1 g), is a self-adjoint Hilbert-Schmidt operator. Let C = kK k1 =
maxxy2G jk(y;1 x)j. Notice that kTf k1 C kf k1 C kf k2 . A moments re ections allows
one to see that this implies that Tf 2 C (G). (Of course, Tf is only dened almost
everywhere, but I mean it agrees almost everywhere with a continuous function on G.
Since this function is uniquely determined, it makes sense to treat Tf as a continuous
function. This is standard practice.) It follows that each eigenfunction of T is continuous.
Lemma 21: Let k be as above and let T be the Hilbert-Schmidt operator on L2(G) dened
by Tf = f k. Then for each  2 C r f 0 g
H = f f 2 L2(G) : Tf = f g  Ef :
Proof: By the above remarks, H is; nite dimensional and consists of continuous func-
tions. Suppose that f 2 H. Then T (g)f = (g)f k = (g) f k = (g)(Tf ) =
;  ; 

 (g)f . That is, H is invariant for . Let f f1  : : :  fr g be an orthonormal basis for

; 

H. Dene continuous functions ki by

ki (g) = h(g)fi  fk i: (1)
Since (g)fi 2 H,
Xr
;
fi (g t) = ki (g)fk (t):
1
(2)
k=1

KF (x y) =
X
uniformly. Notice that for each nite subset F  I the operator TF corresponding to
i i (x)i (y)
i2F
is a nite rank operator. Since the convergence in () is uniform, it follows that TF ! T in the
operator norm
hence T is compact. (The remaining assertions in the paragraph follow from the
Spectral Theorem.)

10
(A priori, Equation (2) is an equality in L2 (G), and so would yeild pointwise equality
only almost everywhere. But since both sides are continuous, the equality must hold
everywhere.) Now dene (g) to be the operator on H whose r  r matrix with respect
to the basis f f1  : : :  fr g is ij . Since we have ij (g) = ji (g;1 ), it follows from
; 

Equation (1) that (g) = (g;1 ). Similarly,

ij (gt) = h(t)fj  (g;1 )fi i
r
=
X
hfk  (g; )fi ih(t)fj  fk i
1

k=1
Xr
= ik (g)kj (t):
k=1

Therefore (gt) = (g)(t). It follows that  is a nite dimensional (unitary) represention

of G. Using Equation (2), we see that
r
X Xr
;
fi (g) = ki (g )fk (e) = fk (e)(g)
1

k=1 k=1

where ki(g) = ki (g;1) = hfi  (g)fk i. We have shown that each fi , and hence H, is
in the span of the matrix coecients of nite dimensional representations of G. Since
every nite dimensional representation is the direct sum of irreducible (nite dimensional)
representaions, we have H  Ef as desired.
Lemma 22: Let k and T be as above. If f 2 L2(G), then Tf 2 Ef .
Proof: Let f i g be a complete orthonormal set of eigenvectors for the eigenspaces with
non-zero eigenvalues of T . By the spectral theorem, we can write
X
f= ck k + 0
k

where T0 = 0 and kf k22 k jck j2. Let Tk = k k . Given  > 0, there is an N so that
P
 
k k>N ckkk2 < . In particular, kT k>N ckk k2 C. But
P P

N  N
k ck k 2 Ef
X X
T ck k =
k=1 k=1
11
by Lemma 21. This suces as
N X 
kTf ; X
k ck k k1 = kT ck k k1 C:
k=1 k>N

The Peter-Weyl theorem now follows as C (G) always contains a self-adjoint approxi-
mate identity. Specically, we have the following.
Proposition 23: If G is a compact group, then there is a net f k g in C (G) which satises
k = k, and such that both f k f g and f f k g converge uniformly to f for each
f 2 C (G).
Proof: For each neighborhood U of e 2 G, let kU be a continuous non-negative function
which satises kU (e) = 1, G kU (g) dg = 1, kU = kU , and supp k  U . Then f kU gU is a
R

net in C (G) directed by reverse inclusion: i.e., U  V if and only if U  V .

Fix f 2 C (G). Since G is compact, given  > 0, there is a neighborhood W of e 2 G
such that jf (t;1 g) ; f (g)j <  for all g 2 G and t 2 W . Therefore if U  W , then
Z 
jkU f (g) ; f (g)j = 


G
; 
kU (t)f (t;1 g) dt f (g)
Z
G kU (t) f (t; g) ; f (g) dt
1

Z
 G kU (t) dt = :
It follows that kU f ! f uniformly, as claimed. On the other hand, (kU f ) = f  kU
implies that f kU ! f uniformly as well. This proves the lemma.
Let TU be the Hilbert-Schmidt operator corresponding to kU . Thus, TU f converges
to f uniformly for each f 2 C (G). Thus, f 2 Ef by Lemma 22, and Theorem 17 is proved.

12
 References

1] Elias Stein, Topics in harmonic analysis, related to the Littlewood-Paley theory,
Annals of mathematics studies, no. 63, Princeton University Press,
New Jersey, 1970

13