Chapter 1 Integrals and The Haar Measure
Chapter 1 Integrals and The Haar Measure
Given a space X, let K(X) be the set of all continuous real valued functions f
defined on X which have compact support. Thus if f E K(X), there is a compact
set C c X (which may vary from one f to another) such that f(x) = 0 if x $. C.
For / 1, / 2 E K(X) and a 1, a 2 E R, it is clear that a 1 / 1 + a 2 / 2 E K(X) so that
K (X) is a real vector space.
DEFINITION 1.1.
R,
is called an integral.
An integral J is a linear function defined on the vector space K (X) which
maps nonnegative functions into nonnegative numbers. Assumption (iii) is to rule
out the trivial case J = 0. Examples of integrals are easily constructed using
Lebesgue measure when the space X is a subset of n dimensional Euclidean
space Rn.
EXAMPLE 1.1. Assume X<;;; Rn with its inherited topology is a space and let
dx denote Lebesgue measure restricted to X. If
dx > 0,
define J on K (X) by
fx f (X) dx.
J(f ) =
go( X) = {
~~~ '
X =!= 0,
0,
X=
JL(B)
Jg (x)dx.
0
Then JL is a well defined a-finite measure on X but JL does not define an integral
Vla
--7
jt(x)JL(dx)
jf(x)JL(dx)
+oo.
(1.1)
J(f)
jt(x)p.(dx),
K(X).
(i)
defined on G
G to G and
(ii)
defined on G to G are both continuous. Then G is a topological group.
Again, the assumption that the topology for G has a countable base is one of
convenience and is ordinarily not part of the definition of a topological group. In
almost all the examples considered here, G is an open or closed subset of a finite
dimensional vector space with the inherited topology. In the examples below, the
verification that G is a topological group is not too hard and most of the details
are left to the reader. Recall that H c G is a subgroup of G if H, with its
inherited operations from G, is a group.
EXAMPLE 1.2. Take G = Rn with addition as the group operation and the
usual topology on Rn. Obviously, 0 ERn is the identity and -x is the inverse of
x E Rn. This group is commutative, that is, x + y = y + x, for x, y E Rn. This
example clearly extends to the case where G is a finite dimensional real vector
space with the Euclidean topology. D
EXAMPLE 1.3. Let G = Gln, which is the group of all n X n nonsingular real
matrices with matrix multiplication as the group operation-commonly called
the general linear group. ForgE Gln, g- 1 means the matrix inverse of g so
that the identity in Gln is the n X n identity matrix. Let .!l'n n be the real
vector space of n X n real matrices so Gln ~ .!l'n n The determina~t function det
is defined on .!l'n n and is continuous. Since
'
Gln
= {xlx Efi>n,n,det(x)
=fo
0},
we see that Gln is an open subset of the Euclidean space fi>n, n That is, Gln is the
complement of the closed set {xlx E fi>n, n det(x) = 0}. Thus, Gln is a space with
the topology inherited from fi>n, n That the group operations are continuous is
not too hard to check. D
EXAMPLE 1.4. Let G be the group G;J, of n X n lower triangular matrices
whose diagonal elements are positive. The group operation is matrix multiplication and G;J, is easily shown to be closed under this operation. That g- 1 E G;J,
for g E G~ is most easily established by induction. For ~ E G;f., partition p as
g (;~~
=
g~J.
0 )
g:;2 1
a;.
a;.
'
a;
a;
a;
a;
The topology for Aln is that of the product space Gin X Rn, which is an open
subset of the coordinate space .Pn n X Rn. The continuity of the group operations is left to the reader to check.' D
Finally, we mention two discrete groups which arise in statistical applications.
The first of these is Pn, the group of permutation matrices. Elements of Pn are
n X n matrices g such that in each row and each column of g, there is exactly
one element which is equal to 1 and all the remaining elements are 0. Each
g E Pn is an orthogonal matrix because gx and x have the same length for all
x E Rn. Thus g- 1 = g' where g' is the transpose of g. That Pn is closed under
matrix multiplication is easily checked. Of course, Pn is a topological group with
the discrete topology. Obviously Pn has n! elements. Elements of Pn are called
permutation matrices because for each x E Rn, gx is a vector whose coordinates
are a permutation of the coordinates of x.
The group of coordinate sign changes Dn has elements g which are n X n
matrices with diagonal elements equal to 1 and off diagonal elements 0.
Obviously each g is an orthogonal matrix and Dn has 2n elements.
1.3. Haar measure. One of the most important theorems regarding topological groups asserts the existence and uniqueness (up to a positive constant) of
left- and right-invariant integrals (measures). This theorem together with the
relationship between left- and right-invariant integrals is described here. A more
extensive discussion of these results together with proofs can be found in
Nachbin (1965) and Segal and Kunze (1978).
To state things precisely let G be a topological group as in Definition 1.3. The
real vector space K(G) is the set of all continuous functions with compact
support defined on G. Given g E G, define the transformation Lg on K(G) to
K(G) by
for g, x E G and
f E K(G).
forgE G.
gE G,
for f E K(G) and hence for all JL-integrable
sometimes written less formally as
JL(d(gx))
JL(dx),
feK(G).
(1.2)
b.(gl )b.(g2).
1/b.(g).
PROOF. A proof of the continuity can be found in Nachbin [(1965), page 77].
To verify the functional equation for b., compute
b.(glg2) jf(x)vz(dx)
b.(gi)b.(g2) jf(x)vz(dx).
f(xg-
1 ).
Rg,g2
is valid. (More is said about this when group actions are discussed.)
DEFINITION 1.5. An integral J is right-invariant if J( f)
f E K (G) and g E G.
THEOREM 1.4.
J(f)
jf(x)v 1(dx)
J( fRg) for
and let
(1.3)
for f
K(G).
J1( fRg)
and the
f(xg
1)
~(g)Jf(x)
= ~(g)
A proof of the second equation can be found in Nachbin [(1965), page 78]. D
Equation (1.3) establishes a relationship between left and right Haar measure.
Namely, if v1 is a left Haar measure, then
(1.4)
is a right Haar measure. Further if vr is a right Haar measure, a similar
construction shows that
(1.5)
is a left Haar measure.
Before turning to a few examples, we discuss the important special case of
compact groups. Here is a characterization of compact groups in terms of Haar
measure.
THEOREM 1.5. In order that G be compact, it is necessary and sufficient that
there exist a finite Haar measure [that is, p,(G) < + oo].
PROOF.
EXAMPLE 1.7. Take G to be the vector space Rn with addition as the group
operation. Then ordinary Lebesgue measure on Rn is both left- and rightinvariant (the group here is commutative) and so the modulus is L1 = 1. 0
EXAMPLE 1.8. If G is any finite or denumerable discrete group, then counting measure is both left- and right-invariant and so the modulus is 1. 0
EXAMPLE 1.9. For this example, take G = Gln as in Example 1.3. Let dx
denote Lebesgue measure restricted to Gln with Gln regarded as a subset of .Pn n
Since G is open in .Pn, n Gln has positive Lebesgue measure. Define an inte~al
J by
dx
J(f)
jf(x)jdet(x)ln
The claim is that J is both left- and right-invariant. To see this, first consider
dx
J(Lgf)
jf(g-lx) jdet(x)jn
where xi is the ith column of x and do the same with y. In this notation, the
equation x = gy becomes
g
10
dx
ldet(g) Indy
J(Lgt) = jf(g-lx) ldet(x)ln = jf(y) ldet(gy)ln
dy
j f ( Y) Idet( y) (
.J(f )
J(f)
dxda
jt[(x,a)] ldet(x)ln+I"
(g, a)
J(Lhf)=
f f[h- (x,a)]
1
dxda
n+l"
ldet(x) I
Thus x = gy and a= gb +a. From the last example, the Jacobian of the
transformation x = gy is ldet(gW and it is obvious that the Jacobian of the
transformation a= gb + a is ldet(g)l. Making this change of variables, we have
J( L h f)
/det(g) /n+l
Jf[ ( Y, b)] /det(gy)
n dy db
/
dydb
f t[(y, b)] /det(y)/n+l
J(f).
+1
11
dxda
jt[(x,a)] !det(x)ln
J(Lgf)
jt(g- 1x)m(dx)
x(g) jt(x)m(dx)
x(g).J(f ).
1.6.
Let X be a multiplier on G.
J(f)
j f(x )x(x-
)m( dx)
is left-invariant.
(ii) If J( f)= ff(x)v 1(dx) is left-invariant, then
Jl(f)
jt(x)x(x)v1(dx)
x.
12
PROOF.
gE G,
J(Lgt)
Only the proof of (i) is given, as the proof of (ii) is similar. For
X ( g- 1 ) X (g)
f f (X ) X(X-
1)
m ( d.x )
J( f )
a;
1f(g- x) d.x
1
x(g)
1f(x) d.x
forgE G,
then
m(d.x)
x(x)v1(d.x),
which relates the left Haar measure v1 and the measure m which defines a
relatively invariant integral with multiplier X For this reason, relatively rightinvariant integrals are not mentioned explicitly.
It follows from Theorem 1.6 that the problem of describing all the relatively
invariant integrals on K(G) can be solved by first exhibiting one relatively
invariant integral and then describing all the multipliers on G. This program is
carried out in some examples that follow. Before turning to these examples, it is
useful to describe a few general facts concerning multipliers.
THEOREM 1.7. Suppose the topological group G is the direct product G 1 X G2
of two topological groups [that is, each g E G has the form g = (gp g 2 ) with
gi E G; and the group operation is go h = (g 1 h 1, g 2 h 2 ) where g = (g 1, g 2 ) and
13
(1.8)
X [ (gp
g2)] = Xl(gl)xig2),
x[(gv e2)],
gl E Gv
X2(g2)=x[(ePg2)],
g2EGz.
The continuity of Xi on Gi follows since xis continuous and Xi is x restricted to
a closed subset of G. That (1.8) holds follows since x is a homomorphism. D
X1(g1)
x~(x
+ y) = x~(x)x~(y).
f(x) =ex,
where c is some real number. Thus,
x 1(x)
exp[cx],
R\
exp[
is a
~ciul
where u E Rn has coordinates u 1, ... , un and c1, ... , en are fixed real numbers.
Thus, a Radon measure m is relatively invariant with multiplier x given above
iff
m(du) = c 0 x(u) du,
where c0 is a fixed constant and du is Lebesgue measure on Rn.
14
When G is the product multiplicative group (0, oo t, the continuous isomorphism cp: G ~ Rn given by
cp(x) =
(0,
X E
oot,
logxn
can be used to show that the measure
x is a
multiplier on G iff
x has
the
x(x)
f1 (x;)\
i~l
where c 1, , en are fixed real numbers. Thus, all the relatively invariant integrals on K (G) are known. D
EXAMPLE 1.12. Consider G
p
Gl n as in Example 1.3 so
dx
(dx) -
Idet( X ) In
is left- and right-invariant. To describe all the multipliers on Gln, the functional
equation
( 1.9)
X ( xy)
X (X )x ( y),
X,
Gl n,
aD/3,
x(/3)
1.
D = fiDi(dJ,
i=l
where di is the ith diagonal of D and Di( di) has ith diagonal equal to di and all
other diagonal elements equal to 1. Then, because Di(dJ E Gln,
x(D)
x(}]ni(dJ)
Qx(DJd;)).
15
E
Pn c 0 n
1, ... , n,
so that
x(Di(dJ) = x(D1(dJ),
because x(P)
i = 1, ... , n,
x(P') = 1. Thus,
x(D)
flx(Di(dJ)
i= 1
Tix(Dl(dJ)
i= 1
0
1
i\ > 0.
0
1
Since these matrices form a subgroup of Gln which is isomorphic to the multiplicative group (0, oo ), it follows from Example 1.11 that
x(Dl(i\))
i\c
for some fixed real number c. Now, retracing the above steps, we have
x(x)
x(D) = (l]dir
Idet( x) Ic.
Hence, all multipliers on Gln are some power of !det(x)l. The converse is clear.
Thus every relatively invariant measure has the form
c0 1 det( X)
dx
c
I
ldet(x) I
x must satisfy
16
so
x (e, g 1x)
x (e, x)
1 = x(e,O)
x(e, x),
X E
x implies that
R".
Hence
x(g,x) = x(g,O),
But g
Gl,.
for some fixed real number c. Conversely, it is clear that such a function on Al,
is a multiplier. Thus, all the multipliers on Al, have been given. D
ExAMPLE 1.14. The final example of this chapter concerns a; c Gl,. The
left and right Haar measures, the modular function and all the multipliers for
a; are derived here. We are going to use the method described after the proof of
Theorem 1.6. For x E a;j., x has the form
x=
Xu
X21
X22
0
0
x,l
x,2
x,,
with X;;> 0 and x; 1 E R", i > j. Let dx denote Lebesgue measure restricted to
the set of such x's in n(n + 1)/2 dimensional space. Define an integral Jon
K(a;j.) by
J(f )
and forgE
f f (X) dx
a;, consider
+ 1)/2
Xo(g)
TI gt;,
i~l
(1.10)
a;
17
is a left Haar measure on G;j. c Gln. The modulus of G~ is computed using the
definition (1.2). Thus, consider
dx
xg 1 so x
xl(g)
ngg-i+l,
i= 1
where gw ... , gnn are the diagonal elements of g. Making this change of
variables in the above integral yields
jf(xg- 1 )vz(dx)
jf(y)x 1(g)
t)
Xo yg
X1(g)
jf(y)~
Xo(g)
Xo(Y)
X1(g)
f f(x)v (dx).
= - (-)
Xo g
(l.ll)
b.(g)
X1(g)
xo(g)
Ilgg-2i+l
i= 1
x(g)
n (gi;)\
i=l
= (
~~~ X~2)'
18
(::: x~J = ( x~ 1
Thus, if x is a multiplier,
0 )(Xu0 0).
1
~ x: x~J
)(
x 22
Xu
x120 )x(xu0 ~)
X( 0
so that
x(
o ) x(
X21
X22
x: x~2 )
= X(
X21Xu
x implies that
~ x~2 )
x22~x(~ x~J
defines a multiplier on (0, oo ), so
0 )
x22
X( 0
(x22rn+I
x 11
~ X ( x~ 1 ~)
a;
defines a multiplier on
c Gln. Since the diagonal elements of Xu are
x 22 = gn+ 1, n+ 1, using the induction hypothesis, we have
!=1
n g{i
n+ 1
f1g~ g~Yi
i=l