February 3, 2009

CHAPTER 1
Introduction to Topological Groups and the
Birkho-Kakutani Theorem
1. Introduction
For us, a topological group is a group G that is equipped with a topology that makes the func-
tions (x, y) xy from GG to G and x x
1
from G to G continuous.
Here are some basic observations regarding topological groups; they follow simply and directly
from the denition
given a, b G, should ab nd itself in an open set U then there are open sets V and W
such that a V, b W and V W = xy : x V, y W U.
given a G, should U be an open set containing a
1
, then there is an open set V
containing a so that V
1
= v
1
: v V U.
given a, b G, should U be an open set containing ab
1
, then there are open sets V and
W such that a V, b W and V W
1
U.
given a, b G, should U be an open set containing a
1
b, then there are open sets V and
W such that a V, b W and V
1
W U.
Each of the mappings from G to G
l
a
: G G, l
a
(x) = ax
r
a
: G G, r
a
(x) = xa
inv : G G, inv(x) = x
1
is a homeomorphism of G onto G.
If F is a closed subset of G then so are aF, Fa, and F
1
for any a G.
If U is an open subset of G and S is a non-void subset of G then the sets S U, U S, and
U
1
are open subsets of G.
G is homogeneous: if p, q G then there is a homeomorphism of G onto G such that
(p) = q. Indeed = l
qp
1 will do the trick.
The homogeneity of topological groups has consequences regarding its topological structure. Heres
one such
Theorem 1.0.1. Let G be a topological group. If U is an open set containing the identity e then
there is an open set V containing e such that e V V U. Consequently, a T
0
topological group
is regular and so Hausdor.
Proof. First things rst: Let U be an open set that contains the identity e. By continuity of
multiplication there is an open set W containing e such that W W U. If we set V = W W
1
then we have an open set that contains e, is symmetric (V = V
1
) and satises V V U.
1
February 3, 2009
2 1. INTRODUCTION TO TOPOLOGICAL GROUPS AND THE BIRKHOFF-KAKUTANI THEOREM
We claim V U. Take x V . Then xV is an open set that contains x and so xV V ,= ,
so there must be v
1
, v
2
V so that xv
1
= v
2
. But then
x = v
2
v
1
1
V V
1
= V V U.
The homogeneous structure of G now tells us that whenever x G and U is an open set containing
x, then there is an open set V such that x V V U. Regularity follows from this and Gs
homogeneity. Again, homogeneity of G tells us that if x, y G are distinct and theres a neighbor-
hood of x that doesnt contain y, then we have a neighborhood of y that doesnt contain x. In other
words, T
0
topological groups satisfy the T
1
axiom.
Henceforth, we assume that all topological groups are Hausdor.
The surprising conclusion reached in Theorem 1.0.1 is a typical product of the mix of the alge-
bra and topology in topological groups. Heres another:
Proposition 1.0.2. Every open subgroup of a topological group is closed.
Proof. Let H be an open subgroup of the topological group G. Take g H. Every open set
that contains g intersects H; gH is such an open set. Therefore gHH ,= . Since cosets are either
the same or disjoint, gH = H. Thus
g = ge gH = H,
and H H.
Exercise 1. If G is a connected topological group then any neighborhood of the identity is a
system of generators for G.
2. The Classical (locally compact) groups
R
n
and C
n
will denote, as usual, real and complex nspaces, respectively. M
n
will denote the
linear algebra of all nn matrices with complex entries. We can associate with any (a
ij
) M
n
the
point (b
1
, . . . , b
n
2) C
n
2
, where
b
i+(j1)n
= a
ij
;
this establishes a bijective correspondence between M
n
and C
n
2
, a correspondence we use to equip
M
n
with the Euclidean topology of C
n
2
.
Of course if , M
n
then M
n
, too, with
c
ij
=
n

k=1
a
ik
b
kj
whenever = (a
ij
), = (b
ij
), and = (c
ij
). It is easy to see that the operation (, )
is continuous from M
n
M
n
to M
n
.
In addition to addition and multiplication, M
n
is endowed by nature with a couple of other natural
operations: transposition and conjugation: if = (a
ij
) M
n
then
t
= (a
ji
) and = ( a
ij
), where
a is the complex conjugate of the complex number a. Both of these operations are homeomorphisms
of M
n
onto itself and each is of order two, that is,
tt
= and

= .
February 3, 2009
2. THE CLASSICAL (LOCALLY COMPACT) GROUPS 3
Some members of M
n
have a multiplicative inverse; the collection of all such matrices will be de-
noted by GL(n, C) and called the General Linear Group. Of course, M
n
belongs to GL(n, C)
precisely when det ,= 0. Now it is easy to believe and also true that det is a continuous
function of the coordinates of and so GL(n, C) is an open subset of M
n
. If , GL(n, C) then
GL(n, C) and ( )
1
=
1

1
. This is elementary linear algebra; further if GL(n, C)
then
1
has coordinates b
ij
where
b
ij
=
p
ij
()
det()
where p
ij
() is a polynomial with coordinates of . It follows that the operation
1
of
GL(n, C) onto itself is also a homeomorphism (of order two).
Corollary 2.0.3. GL(n, C) is a locally compact metrizable topological group.
Proof. After all,
GL(n, C) = det

(z C : z ,= 0),
and so GL(n, C) is homeomorphic to an open subset of a locally compact metric space, C
n
2
. Our
comments about continuity of the operations (, ) and
1
nish the proof.
Inside GL(n, C) we can nd other classical topological groups. Here are a few of them.
O(n) : GL(n, C) is orthogonal if = = (
1
)
t
.
O(n, C) : GL(n, C) is complex orthogonal if = (
1
)
t
.
U(n) : GL(n, C) is unitary if = (
1
)
t
.
Since the mappings
1
and (
1
)
t
are continuous in GL(n, C), each of the groups
O(n), O(n, C) and U(n) are closed subgroups of GL(n, C).
GL(n, R) : M
n
is real if = ; denote the set of real members of M
n
by M
n
(R) so
GL(n, R) = GL(n, C) M
n
(R).
SL(n, C) : the members of GL(n, C) with determinant 1, called the Special Linear Group.
SL(n, R) : SL(N, C) M
n
(R).
SO(n) : SL(n, C) O(n).
SU(n) : SL(n, C) U(n).
Again, SL(n, C), SL(n, R), SO(n), SU(n) are closed subgroups of GL(n, C)). Actually more can
be said.
Theorem 2.0.4. The groups U(n), O(n), SU(n) and SO(n) are compact metric topological
groups.
Proof. Each of O(n), SU(n), SO(n) are closed subgroups of U(n), so its enough to establish
that U(n) is compact. Now U(n) precisely when
t
= id[
C
n. This in turn is the same as saying
if = (a
ij
) that for each 1 i, k n,

j
a
ji
a
jk
=
ik
.
Now the left side is a continuous function of so U(n) is closed not just in GL(n, C) but even in
M
n
. Moreover

j
a
ji
a
ji
= 1
February 3, 2009
4 1. INTRODUCTION TO TOPOLOGICAL GROUPS AND THE BIRKHOFF-KAKUTANI THEOREM
ensures that [a
ij
[ 1 for 1 i, j n. So the entries of any U(n) are bounded. But this ensures
that U(n) is homeomorphic to a closed bounded subset of C
n
2
.
3. The Birkho-Kakutani Theorem
Our next result is technical but its an investment well-worth the price.
Theorem 3.0.5 (Garrett Birkho/Shizuo Kakutani). Let (U
n
) be a sequence of symmetric open
sets each containing the identity e of the topological group G. Suppose that for each k N,
U
k+1
U
k+1
U
k
.
Let H =
k
U
k
. Then there is a left invariant pseudo-metric on GG such that
(i) is left uniformly continuous on GG;
(ii) (x, y) = 0 is and only if x yH;
(iii) (x, y) 4 2
k
, if x y U
k
;
(iv) 2
k
(x, y), if x , yU
k
.
Proof. We start by reinterpreting the sequences descending character with an eye toward
dening . For each k, set
V
2
k = U
k
.
Next dene V
r
for r a dyadic rational number with 0 < r < 1 as follows: if
r = 2
l
1
+ 2
l
2
+ + 2
l
n
,
with 0 < l
1
< < l
n
, all positive integers, let
V
r
= V
2
l
1
V
2
l
2
V
2
l
n
.
For dyadic rs greater than or equal to 1, set V
r
= G. Heres whats so
(3.1) r < s V
r
V
s
;
further for any l N, we have
(3.2) V
r
V
2
l V
r+2
l+2.
We put o the somewhat tedious yet clever proofs of the indicated relationships between the V s
(equations (3.1) and (3.2)) until the end of our general discussion.
With these in hand we go forth to dene . First, for x G, let
(x) = infr : x V
r
.
Plainly (x) = 0 if and only if x H. Now for x, y G dene
(x, y) = sup[(zx) (zy)[ : z G.
Plainly, (x, y) = (y, x) and (x, x) = 0. Its easy to see that (x, u) (x, y) + (y, u) for any
x, y, u G, and the fact that (ax, ay) = (x, y) for all a G is obvious. So is a left invariant
pseudometric.
Now we join the hunt.
February 3, 2009
3. THE BIRKHOFF-KAKUTANI THEOREM 5
Let l N, and suppose u V
2
l and z G. If z V
r
then z u V
r+2
l+2 thanks to (3.2).
Hence
(z u) r + 2
l+2
;
this is true whenever z V
r
, so using the denition of ,
(z u) (z) + 2
l+2
.
Similarly, if z u V
r
then
z V
r
u
1
V
r
V
1
2
l
= V
r
V
2
l V
r+2
l+2,
again by (3.2). It follows that
(z) r + 2
l+2
,
so we see that
(z) (z u) + 2
l+2
.
So
(z) (z u) + 2
l+2
and (z u) (z) + 2
l+2
.
The only conclusion that we can make is that for u V
2
l , and z G,
[(z) (z u)[ 2
l+2
.
From this we see that
(u, e) 2
l+2
, for u V
2
l .
The third statement of the theorem follows from this and s left invariance: If x yU
k
then
y
1
x U
k
= V
2
k so
(x, y) = (y
1
x, e) 2
k+2
= 4 2
k
.
Next we deal with s uniform continuity. Suppose x, y, x, and y satisfy
y
1
x V
2
l1 and y
1
x V
2
l1.
Then
x
1
yy
1
x V
1
2
l1
V
2
l1 = V
2
l1 V
2
l1 V
2
l ,
and so
[(x, y) ( x, y)[ = [(y
1
x, e) ( y
1
x, e)[
[(y
1
x, y
1
x)[
= [( x
1
yy
1
x, e)[
2
l+2
, (by (iii))
and this is what we mean by is left uniformly continuous.
For (iv), suppose y
1
x , U
l
= V
2
l . Then (y
1
x) 2
l
and so
(x, y) = (y
1
x, e) (y
1
x) 2
l
,
February 3, 2009
6 1. INTRODUCTION TO TOPOLOGICAL GROUPS AND THE BIRKHOFF-KAKUTANI THEOREM
where the last inequality follows since for any a G,
(a, e) = sup[(za) (z)[ : z G
[(a) (e)[ = [(a)[, (since e V
r
for every r.)
Finally, (ii) is an easy consequence of (iii) and (iv).
The hard work of Birkho and Kakutani pays o in a couple of fundamental consequences, conse-
quences which underscore the special character of topological groups.
Corollary 3.0.6. Let G be a topological group. If G has a countable neighborhood base at e
then G is metrizable. In this case the metric can be taken to be left invariant.
Proof. Suppose V
n
: n N is a countable open base at e. Let U
1
= V
1
V
1
1
, and U
2
be a
symmetric open neighborhood of e such that such that U
2
U
1
V
2
and U
2
U
2
U
1
. Continuing,
let U
n
be a symmetric open neighborhood of e such that
U
n
U
1
U
2
U
n1
V
n
,
and U
2
n
U
n1
.
The family U
k
: k N satises the conditions set forth in the Birkho-Kakutani theorem. Fur-
ther, H =
n
U
n
= e. Let be the left pseudo-metric introduced in the theorem. In fact, is a
true metric; (x, y) = 0 if and only if x = y, since after all, H = e!
Since
x G : (x, e) < 2
k
U
k
x G : (x, e) 2
k+2
,
the topology dened by coincides with the given topology of G.
Corollary 3.0.7. Let G be a topological group, a G, and F be a closed subset of G such
that a , F. Then there is a continuous real function on G such that (a) = 0 and (x) = 1 for
all x F. Consequently, every topological group is completely regular.
Proof. Let U
1
be a symmetric neighborhood of e such that (aU
1
) F = . Choose a sequence
(U
n
: n 2) of open neighborhoods of e such that each U
n
is symmetric, U
n+1
U
n+1
U
n
, and let
H =
n
U
n
. Apply Birkho-Kakutani theorem to the U
n
s. For x G, dene (x) by
(x) = min1, 2(a, x)
where is the left-invariant, uniformly continuous pseudometric produced in the Birkho-Kakutani
theorem. is continuous by (i) of our theorem, and (a) = 0. If x F then a
1
x a
1
F, a set
disjoint from U
1
; consequently, a
1
x , U
1
and (a, x) 2
1
by (iv) of our theorem. It follows that
(x) = 1 for all x F.
Corollary 3.0.8. Let G be a locally compact group and suppose e is the intersection of
countably many open sets. Then G admits a left-invariant metric compatible with its topology.
Proof. Suppose e =
n
U
n
, where U
n
is open in G. Choose open sets V
1
, V
2
, . . . , V
n
, . . . ,
each containing e with V
n
a compact subset of V
n1
U
n
for each n. We claim that V
n
: n N
is an open base at e.
February 3, 2009
3. THE BIRKHOFF-KAKUTANI THEOREM 7
Let W be an open set in G that contains e. Could it be that no V
n
W? Well lets suppose
this to be the case. Then V
n
W
c
: n 2 enjoys the nite intersection property; indeed,
V
1
. . . V
n
W
c
V
n+1
W
c
V
n+1
W
c
,= .
It follows that
n
V
n
W
c
,= . But

n2
_
V
n
W
c
_
=
_
_

n2
V
n
_
_
W
c

n
U
n
_
W
c
= e W
c
= .
OOPS! There must be some V
n
thats contained in W.
Heres another feature of topological groups that distinguishes them from general topological spaces;
they admit apt notions of uniform continuity, for instance.
Theorem 3.0.9. Let G be a topological group and M be a non-empty compact subset of G.
Then any continuous function f : G R is left uniformly continuous on M. i.e., given > 0 there
is an open set V containing the identity of G so that if x, y M and x yV then [f(x) f(y)[ .
Proof. Let > 0 be given. For each a M there is an open set V
a
that contains the identity
such that if x M and x aV
a
then [f(x) f(a)[

2
. Since e e = e, there is an open set W
a
that
contains the identity e and satises W
a
W
a
V
a
. Now if a M then a a W
a
so a W
a
: a M
covers M; we can nd a
1
, . . . , a
n
M so that
M (a
1
W
a
1
) (a
n
W
a
n
).
Look at
V = W
a
1
W
a
n
.
Then V is open and contains e. Let x, y M with x yV . To see that [f(x) f(y)[ note
that if y M then y a
i
W
a
i
for some i = 1, . . . , n. It follows from this and W
a
V
a
that
[f(y) f(a
i
)[

2
. Also
x yV a
i
W
a
i
V a
i
W
a
i
W
a
i
a
i
V
a
i
,
and so [f(x) f(a
i
)[

2
too.
In sum, if x, y M with x yV then
[f(x) f(y)[ [f(x) f(a
i
)[ +[f(a
i
) f(y)[

2
+

2
= .

Well be spending much of our time in a locally compact setting and again, theres more than
meets the eye because of the groups structure.
Theorem 3.0.10. Any locally compact topological group is paracompact, hence normal.
Proof. Let G be a locally compact topological group, and let V be an open set in G containing
the identity of G and having a compact closure V . By thereom 1.0.1 there exists a symmetric open
set U such that
e U U U U V.
February 3, 2009
8 1. INTRODUCTION TO TOPOLOGICAL GROUPS AND THE BIRKHOFF-KAKUTANI THEOREM
Look at H =
n
U
n
, where U
n
= U U
. .
n
. Then H is an open subgroup (hence closed) of G. Also
H is -compact since
H =
_
n
U U
. .
2n

_
n
U U
. .
n
H,
a countable union of compact subsets. Hence H is Lindelof and regular, so H is paracompact [?].
Let | be an open cover of G. Each coset xH of H is also -compact and so there is a count-
able subfamily V
(n)
xH
: n N of | that covers xH.
Naturally,
V
(n)
xH
xH : n N
is an open cover of xH and since xH is paracompact, there is a locally nite open cover of xH that
renes V
(n)
xH
xH : n N, call it J
(n)
xH

n=1
. Note that for each n = 1, 2, . . .
J
(n)
xH
xH.
For each n = 1, 2 . . . , let
J
n
=
_
xHG/H
J
(n)
xH
,
so J
n

n=1
is a locally nite open cover. Therefore J =
n
J
n
is an open cover of G thats
plainly -locally nite and renes |. Thus each open cover | of G admits a -locally nite open
cover J that renes | so G is paracompact and therefore normal ([?], Theorem 5.28, Corollary 5.32).
Warning: Not every topological group is normal as exhibited by the following classical example.
Example 3.0.11. Let m be an uncountable cardinal number. Then Z
m
is a T
0
group, hence
completely regular. But Z
m
is not normal.
Proof. Well write Z
m
as

iI
Z
i
where each Z
i
is Z and [I[ = m. For the sake of the present
eorts, let A, B Z
m
be given as follows
A = (x
i
) Z
m
: for any n ,= 0, there is at most one index i for which x
i
= n,
and
B = (x
i
) Z
m
: for any n ,= 1, there is at most one index i for which x
i
= n.
Then A and B are disjoint. If (x
i
)
iI
, A then there are i
0
, i
1
I, i
0
,= i
1
such that for some
n Z, n ,= 0, x
i
0
= x
i
1
= n. The set
(y
i
)
iI
Z
m
: y
i
0
= y
i
1
= n
is an open set containing (x
i
)
iI
but no point of A, so A is a closed set. Similarly B is a closed set.
Let U, V be open subsets of Z
m
such that A U and B V. We claim that U V ,= , which will
show that Z
m
is not normal.
Let (x
(1)
i
)
iI
Z
m
be dened by x
(1)
i
= 0 for each i I. Since (x
(1)
i
) A U, there are dis-
tinct indices i
1
, . . . , i
m
1
I such that
(x
(1)
i
)
iI
(x
i
)
iI
Z
m
: x
i
1
= x
i
2
= x
i
m
1
= 0 U.
February 3, 2009
3. THE BIRKHOFF-KAKUTANI THEOREM 9
Let (x
(2)
i
)
iI
be dened by
x
(2)
i
=
_
k for i = i
k
, k = 1, . . . , m
1
0 otherwise.
Since (x
(2)
i
) A U there exists i
m
1
+1
, . . . , i
m
2
I, distinct indices from each other and from
i
1
, . . . , i
m
1
such that
(x
(2)
i
)
iI
(x
i
)
iI
Z
m
: x
i
1
= 1, x
i
2
= 2, . . . , x
i
m
1
= m
1
, x
i
m
1
+1
= 0, . . . , x
i
m
2
= 0 U,
Continue in this manner.
Dene (y
i
)
iI
Z
m
as follows: y
i
k
= k for any k and y
i
= 1 if i ,= i
k
. Plainly, (y
i
)
iI
B.
Hence for some nite subset J of I,
(x
i
)
iI
Z
m
: x
i
= y
i
, for i J V.
But J is nite so there is an n
0
such that i
k
, J whenever k > m
n
0
. Look at (z
i
)
iI
Z
m
, where
z
i
=
_
_
_
k if i = i
k
, k m
n
0
0 if i = i
k
, m
n
0
+ 1 k m
n
0
+1
1 otherwise
So
(z
i
)
iI
(x
i
)
iI
Z
m
: x
i
= y
i
, for i J V.
So
(z
i
)
iI
(x
i
)
iI
Z
m
: x
i
1
= 1, x
i
2
= 2, . . . , x
i
m
0
= m
0
, x
i
m
n
0
+1
= x
i
m
n
0
+2
= x
i
m
n
0
+1
= 0 U.
So (z
i
)
iI
V, (z
i
)
iI
U, and U V ,= . Therefore Z
m
is not normal.
After thoughts: The missing steps of the Birkho-Kakutani Theorem. To prove (3.1)
r < s V
r
V
s
,
suppose s < 1, and write r, s in dyadic forms
r = 2
l
1
+ 2
l
2
+ + 2
l
n
, 0 < l
1
< l
2
< < l
n
s = 2
m
1
+ 2
m
2
+ + 2
m
p
, 0 < m
1
< m
2
< < m
p
,
where l
1
, . . . , l
m
, m
1
, . . . , m
p
are all positive integers. There must be a k so that
l
1
= m
1
, l
2
= m
2
, . . . , l
k1
= m
k1
but l
k
> m
k
.
Let
W = V
2
l
1
V
2
l
2
V
2
l
k1
.
February 3, 2009
10 1. INTRODUCTION TO TOPOLOGICAL GROUPS AND THE BIRKHOFF-KAKUTANI THEOREM
Then
V
r
= W V
2
l
k
V
2
l
k
+1 V
2
l
n
W V
2
l
k
V
2
l
k
1 V
2
l
k
2 V
2
l
n
+1 V
2
l
n
V
2
l
n
W V
2
l
k
V
2
l
k
1 V
2
l
k
2 V
2
l
n
+1 V
2
l
n
+1
(since after all V
2
l
n
V
2
l
n
= U
l
n
U
l
n
U
l
n
1
= V
2
l
n+1
)
.
.
.
W V
2
l
k
V
2
l
k
W V
2
l
k
+1
W V
2
m
k
V
2
l
1
V
2
l
2
V
2
l
k
1 V
m
k
V
2
m
1
V
2
m
2
V
2
m
k1
V
2
m
k
V
2
m
1
V
2
m
2
V
2
m
k
V
2
m
p
= V
s
.
With the same representation of r in mind we set o to prove (3.2)
V
r
V
2
l V
r+2
l+2.
We suppose that r + 2
l+2
< 1. If l > l
n
then
V
r
V
2l
= V
r+2
l ,
and all is well. So we look to the case that l l
n
.
Let k be the positive integer such that
l
k1
< l l
k
, (l
0
= 0).
Let r
1
be given by
r
1
= 2
l+1
2
l
k
2
l
k
+1
2
l
n
,
and
r
2
r +r
1
.
Its plain that
r < r
2
< r + 2
l+1
,
so
V
r
V
2
l V
r
2
V
2l
= V
r
2
+2
l V
r+2
l+1
+2
l V
r+2
l+2,
and thats that.
February 3, 2009
CHAPTER 2
Lebesgue Measure in Euclidean Space
By an interval in R
n
we mean any set I of the form
I = I
1
I
k
where I
1
, , I
k
are nite intervals in R. We do not ask that I
1
, , I
k
all be open, closed or
half-open/ half-closed mixtures are just ne. Each I
j
R has a length l(I
j
) and with this mind we
dene volume of I by
vol(I) =

jk
l(I
j
)
let A R
k
. We dene the outer measure or Lebesgue outer measure of A, m

(A) by
m

(A) = inf
_

n
vol(I
n
) : I
n
is an interval in R
k
, A
n
I
n
_
.
Regarding edges: There is a great deal of latitude with regards to the nature of the edges of the
intervals in the coverings of a set A R
k
that are used to compute m

(A).
For instance if we wish we can assume each edge has length less than for some xed > 0.
This is plain since any interval I in R
k
is the union of overlapping intervals all of whose edges have
length less than and the sum of whose volume totals Is volume.
More, we can assume were covering A by open intervals, that is, all the edges are open. In fact, if
(I
j
) is a covering of A by intervals and > 0 then for each j we can enlarge I
j
to an open interval
J
j
, I
j
J
j
and
vol(J
j
) < vol(I
j
) +/2.
It follows that in computing m

(A), if (I
j
) is a covering of A by intervals then we can nd a sum

vol(J
j
) that is as close as we please to

vol(I
j
) where (J
j
) is a covering of A by open intervals.
Hence we can restrict our attention to nding the inmum of such sums

vol(J
j
) where (J
j
) is an
open covering of A by intervals.
1
0
If A B then
m

(A) m

(B).
2
0
If A =
n
A
n
then
m

(A)

n
m

(A
n
)
11
February 3, 2009
12 2. LEBESGUE MEASURE IN EUCLIDEAN SPACE
We can, and do, assume that

n
m

(A
n
) < . Indeed, let > 0 be given and choose for each
n a sequence (I
n
j
)of intervals that cover A
n
and satisfy

j
vol(I
n
j
) < m

(A
n
) +

2
n
j
.
Since A =
n
A
n

n,j
I
n
j
,
m

(A)

n,j
vol(I
n
j
)

n
_
m

(A
n
) +

2
n
_

n
m

(A
n
) +.
3
0
For any interval I,
m

(I) = vol(I).
Let > 0 be given and let (I
j
) be an open covering of I such that

j
vol(I
j
) < m

(I) +.
Take any closed subinterval J of I. Since J is compact there is a j
0
such that J I
1
I
j
0
.
Lets look closely to the intervals I
1
, , I
j
0
, J. Each (k 1) dimensional face of these intervals
lies in a (k 1) dimensional hyperplane in R
k
; in turn, these hyperplanes divide

I
1
, . . . ,

I
j
0
into
closed intervals K
1
, . . . , K
n
j
0
; similarly J is divided into closed intervals J
1
, . . . , J
m
j
0
by the same
hyperplane. (Think of the case n = 3.) Since J
jj
0
I
j
each J
m
is one of the K
n
s so that
vol(J) =

mm
j
0
vol(J
m
)

nn
j
0
vol(K
n
)

jj
0
vol(I
j
)
m

(I) +.
This is so for every closed subinterval J of I so
vol(I) m

(I) +;
epsilonics soon tell us that
vol(I) m

(I).
The reverse is plain.
Some ground work is needed to prepare the way for measurable sets.
A. If F
1
and F
2
are disjoint closed bounded sets then
m

(F
1
F
2
) = m

(F
1
) +m

(F
2
).
February 3, 2009
2. LEBESGUE MEASURE IN EUCLIDEAN SPACE 13
Let > 0 be chosen so that no interval of diameter less than meets both F
1
and F
2
(e.g.,
<
1
2
d(F
1
, F
2
).) Let > 0 be given. Pick a sequence (I
i
) of intervals of diameter less than such
that
F
1
F
2

i
I
i
, and

i
vol(I
i
)

m

(F
1
F
2
).
Denote by (I
(1)
k
) those intervals among the I
i
s that meet F
1
and by (I
(2)
k
) those that meet F
2
.
F
1

j
I
(1)
j
and F
2

I
(2)
j
and by our judicious concerns over . Alas,
m

(F
1
) +m

(F
2
)

j
vol(I
(1)
j
) +

j
vol(I
(2)
j
)

i
vol(I
i
) m

(F
1
F
2
) +;
Epsilonics to the rescue: m

(F
1
) +m

(F
2
) m

(F
1
F
2
).
B. If G is a bounded open set then for each > 0 there is a closed set F G such that
m

(F) > m

(G) .
Represent G =
i
I
i
where I
i
s are non-overlapping intervals, and let > 0 be given; of course
m

(G)

i
vol(I
i
) and so there is an n
0
so that

in
0
vol(I
i
) > m

(G)

2
.
For each i n
0
let J
i
be a closed subinterval of the interior of I
i
with
vol(J
i
) > vol(I
i
)

2
i
.
Then F =

in
0
J
i
is a closed subset of G and
m

(F) = m

_
_
_
in
0
J
i
_
_
=

in
0
m

(J
i
) (by A)
=

in
0
vol(J
i
)
>

in
0
_
vol(I
i
)

2
i
_

in
0
vol(I
i
)

2
> m

(G) .
NB: The openness of G was used to represent G in an appropriate way.
C. If F is a closed subset of an open bounded set G then
m

(GF) = m

(G) m

(F).
February 3, 2009
14 2. LEBESGUE MEASURE IN EUCLIDEAN SPACE
Let > 0 be given. Using B, chose a closed set F
1
GF so that
m

(F
1
) > m

(GF) .
Notice that
m

(F) +m

(GF) m

(F) + (m

(F
1
) +)
= m

(F
1
F
2
) +
m

(G) +.
Epsilonics take over to say
m

(F) +m

(GF) m

(G).
Of course, 2
0
takes care of the reverse inequality and with it, C.
A subset E of R
k
is Lebesgue measurable if given an > 0 there is a closed set F and open set
G such that
F A G, and m

(GF) < .
By the complementary nature of open and closed sets E is Lebesgue measurable if and only if E
c
is:
F E G G
c
E
c
F
c
and F
c
G
c
= GF.
4
0
If A and B are measurable then so is A B.
Pick F
A
, F
B
closed and G
A
, G
B
open such that
F
A
A G
A
and m

(G
A
F
A
)

2
F
B
B G
B
and m

(G
B
F
B
)

2
F = F
A
F
B
is closed, G = G
A
G
B
is open, F A B G, and GF (G
A
F
A
) (G
B
F
B
)
so that
m

(GF) m

(G
A
F
A
) +m

(G
B
F
B
)

2
+

2
= .
5
0
A bounded set B is measurable if for each > 0 there is a compact set K B such that
m

(K) > m

(B) .
Suppose the bounded set B satises the conditions set forth and let > 0 be given. We can nd a
compact set K B such that
m

(K) m

(B)

2
.
But m

(B) < (why is that?). So we can cover B by a sequence (I

j
) of open intervals each of
diameter less than
1
37
such that

j
vol(I
j
) < m

(B).
February 3, 2009
2. LEBESGUE MEASURE IN EUCLIDEAN SPACE 15
Let G be the union of all those I
j
s that meet B. K B G and G is bounded. So our preparatory
... tells us that
m

(GK) = m

(G) m

(K)

j
m

(I
j
) m

(K)
=

j
vol(I
j
) m

(K)
m

(B) +

2
m

(K)
= m

(B) m

(K) +

2

2
+

2
= .
6
0
(Finite) intervals are measurable.
After all, if I is a (nite or bounded) interval m

(I) = vol(I) and so we can plainly approxi-

mate I from the inside by compact intervals.
7
0
Sets of outer measure zero are measurable.
If m

(N) = 0 and > 0 is given there must be a sequence (I

j
) of open intervals so N
j
I
j
and

j
vol(I
j
) m

(N) + = .
Then G =
j
I
j
and F = soon show the way to Ns measurability.
Another technical rest stop.
D. If (A
n
) is a sequence of disjoint measurable sets of the interval I then
n
A
n
is measurable,
too, and
m

(
n
A
n
) =

n
m

(A
n
).
Let > 0 be given. Choose compact sets F
n
A
n
so
m

(F
n
) > m

(A
n
)

2
n+1
.
Since
m

(
n
A
n
)

n
m

(A
n
)
there is an n N so that

nn
0
m

(A
n
) > m

(
n
A
n
)

2
.
If F =
nn
0
then F is compact (its closed, and being a subset of I, bounded). Hence by A
m

(F) =

nn
0
m

(F
n
) >

nn
0
m

(A
n
)

2
> m

(
n
A
n
)

2
.
February 3, 2009
16 2. LEBESGUE MEASURE IN EUCLIDEAN SPACE
Weve just taken the bounded set
n
A
n
and for each > 0 found a compact set K, contained in

n
A
n
so that
m

(F) > m

(A) .

n
A
n
is measurable thanks to 5
0
. Lets check the sums: for any n N

nm
m

(A
n
) <

nm
_
m

(F
n
) +

2
n+1
_

nm
m

(F
n
) +

2
= m

(
nm
F
n
) +

2
= m

(
n
A
n
) +

2
.
Epsilonics assure us that

nm
m

(A
n
) m

(
n
A
n
)
and this is so for each n. It follows that

n
m

(A
n
) m

_
_
n
A
n
_
.
Weve already seen the reverse so thats all she wrote for D.
8
0
If (A
n
) is any sequence of disjoint measurable set of R
k
then
n
A
n
is measurable and
m

(
n
A
n
) =

n
m

(A
n
).
Well bootstrap our way from D to 8
0
. To start, let (I
m
) be a sequence of disjoint intervals whose
union is R
k
and such that any bounded set in R
k
is covered by nitely many I
m
s.
For each m, n N let
A
m,n
= I
m
A
n
be that part of A
n
inside I
m
. Each A
m,n
is measurable (6
0
and 4
0
) and then A
m,n
s are pairwise
disjoint. Look to

A
m
=
n
A
m,n
,
the part of
n
A
n
in I
m
. By D,

A
m
is measurable. Further, the

A
m
s are pairwise disjoint and

A
m
=
n
A
n
.
Let > 0 be given. For each m, choose a closed set F
m


A
m
and an open set G
m
, which is
a bounded open set,

A
m
G
m
such that
m

(G
m
F
m
) <

2
m
.
Look at F =
m
F
m
and G =
m
G
m
. F is closed (if (x
n
) is a convergent sequence of points of
F then (x
n
) is bounded and so for some m
0
, (x
n
) is a sequence on
mm
0
F
m
hence visits one of
February 3, 2009
2. LEBESGUE MEASURE IN EUCLIDEAN SPACE 17
F
1
, . . . , F
m
0
innitely often - which ever F
j
it visits so often contains its limit and G is open.
F =
m
F
m

m

A
m
=
m
A
m
=
m

A
m

m
G
m
= G.
Further,
GF =
m
(G
m
F)
m
(G
m
F
m
)
so
m

(GF)

m
m

(G
m
F
m
)

2
m
= ,
and
n
A
n
is measurable.
Now
A
n
= A
m,n
so
m

(A
n
)

m
m

(A
m,n
).
It follows that

n
m

(A
n
)

m
m

(A
m,n
)
=

n
m

(A
m,n
)
=

m
m

A
m
),
by D. Take m N. Then

jm
m

A
j
)

jm
(m

(F
j
) +m

(G
j
F
j
))
= m

_
_
_
jm
F
j
_
_
+

jm
m

(G
j
F
j
) (by A)
m

_
_
_
jm
F
j
_
_
+

jm

2
j
m

_
_
n
A
n
_
+.
The usual epsilonics leads us to conclude that

m
m

A
m
) m

_
_
n
A
n
_
and in tandem with what gone on before we see

n
m

(A
n
) m

_
_
n
A
n
_
.
Again the reverse holds without assumption so 8
0
is proved.
Theorem 0.0.12 (The Fundamental Theorem of Lebesgue Measure). In summary
February 3, 2009
18 2. LEBESGUE MEASURE IN EUCLIDEAN SPACE
(i) m

is a non-negative, extended real-valued function dened for every subset of R

k
which
assigns
to each interval, a value equal to its volume,
to each set, a value common to all its translates,
to bigger sets, bigger values
to compact sets, nite values
to non-empty sets, positive values
and is countably subadditive in doing so. For any A R
k
,
m

(A) = infm

(G) : G is open A G.
(ii) The Lebesgue measurable subsets of R
k
form a eld / of sets containing every open
set, closed set, interval, and set of outer measure zero; E /if and only if Es translates
are members of /.
(iii) m

is countably additive on / and for E /,

m

(E) = supm

(K) : K is compact , K E.
February 3, 2009
CHAPTER 3
Invariant Measures on R
n
1. Introduction
As we proceed in our study of invariant measures we will encounter theorems that assert the
uniqueness of such measures in varying degrees of generality and in diering senses.
On compact object the uniqueness will be with regard to Borel probabilities (positive Borel
measures of total mass 1).
If G is a locally compact group then well see that there is a unique left invariant regular Borel
measure acting on (the Borel subsets of G) with uniqueness taken to mean up to multiplicative
constants.
In concrete cases it occurs that those multiplicative constants can themselves be of considerable
interest, representing the rate of exchange between various natural geometric view points of the
groups under consideration.
In this chapter we encounter an early example of seemingly mulitple invariant measures on a con-
crete group and compute the constants that allow one to convert from one viewpoint to another.
Our setting with be R
n
. We will be dealing with Lebesgue nmeasure
n
on R
n
and Hausdor
nmeasure
(n)
on R
n
. Each is translation invariant and, sure enough, for each n, there is a
constant
n
such that

(n)
=
n

n
.
The constant
n
is in fact the precise rate of exchange that allows one to move from a rectalinear
view of R
n
(as expressed by
n
) to a spherical view (as found in
(n)
.)
Along the way to establishing the existence of
n
and of computing it, we meet some of the real
stalwarts of measure theory.
In the rst section we give the elegant Hadwiger-Ohlmann proof of the Brunn-Minkowski The-
orem, a geometric version of the Arthimetic-Geometric Mean Inequality. As a simple consequence
of this we derive the Isodiametric Inequality which says that among Borel sets in R
n
of the same
diameter, the ball has the greatest volume.
In the next section we derive the still-wonderful Covering Theorem of Vitali for Vitali fam-
ilies of balls. We follow this with a short introduction to Hausdor measure on R
n
,
(n)
. This leads
19
February 3, 2009
20 3. INVARIANT MEASURES ON R
n
to the main course of this chapter, the proof that for each n there is a
n
> 0 such that

(n)
=
n

n
.
2. The Brunn-Minkowski Theorem
Theorem 2.0.13 (The Brunn-Minkowski Theorem). Let n 1, and let
n
denote Lebesgue
measure on R
n
. If A and B are compact subsets of R
n
then
(
n
(A+B))
1/n
(
n
(A))
1/n
+ (
n
(B))
1/n
(BM)
where
A+B = a +b : a A, b B.
Notice that (BM) is a geometric generalization of the Arithmetic-Geometric Mean Inequality
for if A and B are rectangles with sides of length (a
j
)
n
j=1
and (b
j
)
n
j=1
respectively, then (BM) looks
like
_
n

i
(a
j
+b
j
)
_
1/n

_
n

1
a
j
_
1/n
+
_
n

1
b
j
_
1/n
. (BM)

.
Homogeneity lets us reduce this to the case where a
j
+b
j
= 1 for each j. But now the Arithmetic-
Geometric Mean Inequality assures that
1
n
n

j=1
a
j

_
n

1
a
j
_
1/n
, and
1
n
n

j=1
b
j

_
n

1
b
j
_
1/n
.
So (noting that

n
j=1
a
j
+

n
j=1
b
j
=

n
j=1
(a
j
+b
j
) =

n
j=1
1 = n)
1 =
1
n
n
_
n

1
a
j
_
1/n
+
_
n

1
b
j
_
1/n
,
which is (BM)

. Thus we have proved (BM) for boxes, rectangular parallelpipeds whose sides are
parallel to the coordinate hyperplanes.
To prove (BM), note that A and B each are the union of nitely many rectangles whose interi-
ors are distinct. We proceed by induction on the total number of rectangles in A and B. It is
important to realize that the inequality is inaected if we translate A and B independently: in fact,
replacing A by A + h and B by B + k replaces A + B by A + B + h + k and the corresponding
measures are the same as what we started with. If R
1
and R
2
are essentially disjoint rectangles
in the collection making up A then they can be separated (a translation may be necessary) by a
coordinate hyperplane x
j
= 0 say. Thus we may assume that R
1
lies in A

= A x
j
0 and
R
2
lies in A
+
= Ax
j
0. Notice that each A
+
and A

contain at least one rectangle less than

does A and A = A
+
A

.
What to do with B? Well, slide B over so if B
+
= B x
j
0 then

n
(B
+
)

n
(B)
=

n
(A
+
)

n
(A)
;
of course this entails

n
(B

n
(B)
=

n
(A

n
(A)
February 3, 2009
2. THE BRUNN-MINKOWSKI THEOREM 21
as well. But
(A
+
+B
+
) (A

+B

) A+B
and the union on the left hand side is essentially disjoint. Moreover the total number of rectangles
in either A
+
and B
+
or in A

and B

is less than that in A and B. Our induction hypothesis applies.

The result?

n
(A+B)
n
(A
+
+B
+
) +
n
(A

+B

)
(
n
(A
+
)
1/n
+
n
(B
+
)
1/n
)
n
+ ((
n
(A

)
1/n
+
n
(B

)
1/n
)
n
(induction hypothesis)
=
n
(A
+
)
_
1 +
_

n
(B)

n
(A)
_
1/n
_
n
+
n
(A

)
_
1 +
_

n
(B)

n
(A)
_
1/n
_
n
=
n
(A)
_
1 +
_

n
(B)

n
(A)
_
1/n
_
n
= (
n
(A)
1/n
+
n
(B)
1/n
)
n
and thats all she wrote. Thus we have (BM) for nite unions of boxes.
If A and B are open sets of nite measure then once given a margin of error, > 0, we can
nd unions A

, B

of essentially disjoint rectangles such that A

A, B

B, and

n
(A)
n
(A

) +,
n
(B)
n
(B

) +.
Once done, A

+B

A+B and we can apply the work of the previous paragraphs:

n
(A+B)
1/n

n
(A

+B

)
1/n
(
n
(A) )
1/n
+ (
n
(B) )
1/n
.
Since > 0 was but a margin of error and we can assume arbitrary small errors and made at worst,
it follows that (BM) holds for open sets A and B of nite measure.
If A and B are compact sets then A + B is compact as well. Look at A

= [d(x, A) < ]. Then

A

is open, contains A and A

A. Similarly B

is dened analogously. What can we say about

(A+B)

? The same conclusions can be drawn. Further

A+B A

+B

(A+B)
2
.
Applying (BM) to A

, B

we get

n
((A+B)
2
)
1/n

n
(A

+B

)
1/n

n
(A

)
1/n
+
n
(B

)
1/n

n
(A)
1/n
+
n
(B)
1/n
.
Letting 0 gives (BM) for A, B, A+B.
The general situation of A, B and A + B measurable follows by approximating from within: if

A,

B are compact sets with

A A,

B B then

A+

B is a compact set inside A+B, and

n
(A+B)
n
(

A+

B) (
n
(

A)
1/n
+
n
(

B)
1/n
)
n
.
February 3, 2009
22 3. INVARIANT MEASURES ON R
n
Let
n
(

A)
n
(A),
n
(

B)
n
(B) and be done with it.
With the Brunn-Minkowski inequality in hand we can easily nd our way to the isodiametric
inequality.
Theorem 2.0.14. For any Borel set B R
n
we have

n
(B)
_
diam B
2
_
n

n
(B
n
),
where as usual B
n
denotes the closed unit ball of R
n
.
Proof. Without loss of sleep we can assume that
d := diam B < .
This in mind, realize that if x, y, B then [[x y[[ d. Hence
B B d B
n
,
and so
2
n
(B)
1/n
=
n
(B)
1/n
+
n
(B)
1/n
=
n
(B)
1/n
+
n
(B)
1/n

n
(B B)
1/n
(Brunn-Minkowski to the rescue!)

n
(dB
n
)
1/n
= d
n
(B
n
)
1/n
.
Its easy to deduce the conclusion of the theorem from this inequality:
2
n
(B)
1/n
diam(B)
n
(B
n
)
1/n
.
3. Vitalis Covering Theorem
Theorem 3.0.15 (Vitali). Let T be a family of closed balls in R
n
that covers a set E in the
sense of Vitali. i.e., given an > 0 and x E there is a B T such that the diameter of B is less
than and x B. Then T contains a disjoint sequence that covers
n
almost all of E.
Proof. First we suppose E is bounded and contained in the bounded open set G. We disregard
any members of T that arent contained in G as well as those that dont intersect E. The resulting
family, which we will still refer to as T covers E in the sense of Vitali. Our proof will be by
exhaustion. To start let R
1
be
R
1
:= supradius(B) : B T.
Choose B
1
T, say centered at a, with radius r
1
so that
R
1
2
r
1
.
Next let R
2
be
R
2
= supradius (B) : B T, B
1
B
2
= .
Choose B
2
T, say centered at a
2
with radius r
2
so that B
2
B
1
= and
R
2
2
r
2
.
February 3, 2009
3. VITALIS COVERING THEOREM 23
Continue down this primrose lane. After k steps, let R
k+1
be
R
k+1
= supradius(B) : B T, B (B
1
B
k
) = .
Choose B
k+1
T with center a
k+1
, radius r
k+1
so
B
k+1
(B
1
B
k
) =
and
R
k+1
2
r
k+1
.
Naturally,
R
1
R
2

More is so. The B
k
s are pairwise disjoint so

n
(B
k
) =
n
_
_
k
B
k
_

n
(G) < ;
hence
lim
k

n
(B
k
) = 0.
But

n
(B
k
) = c
n
r
n
k
(and we know c
n
but we dont need to right now) so
lim
k
r
k
= 0.
Since 0 R
k
2r
k
we see
lim
k
R
k
= 0.
The B
k
s eat up E. How? What does it mean for x E not to be in

kK
B
k
? Well x ,

kK
B
k
means
d
_
_
x,
_
kK
B
k
_
_
= > 0.
Choose B T so x B and B (which is centered at say a with radius r) has radius less than /2.
Each point of B is within 2r of x and 2r < . Hence B is disjoint from B
1
B
K
. B is
one of those balls belonging to T that take part in dening R
k+1
; in particular, r R
k+1
. But
lim
j
R
j
= 0
and the R
j
s are monotone and non-increasing. Hence there is a j (which is necessarily bigger than
K) so that
R
j+1
< r R
j
.
Since r > R
j+1
, B must meet one of the balls B
1
, . . . , B
j
; Suppose B B
m
,= . Since B does not
meet B
1
B
K
, m > K. It follows that for any c B B
m
[[x a
m
[[ [[x a[[ +[[a c[[ +[[c a
m
[[ (where a
m
is the center of B
m
and a is the center of B)
r +r +r
m
= 2r +r
m
.
February 3, 2009
24 3. INVARIANT MEASURES ON R
n
So
[[x a
m
[[ 2r +r
m
,
where
r R
j
R
m
2r
m
.
WHY?
Hence
[[x a
m
[[ < 2r +r
m
2 (2r
m
) +r
m
5r
m
.
To summarize: if x E
kK
B
k
then x belongs to the union
k>K

B
k
of those closed balls

B
k
,
where

B
k
shares the center a
k
with B
k
but

B
k
has ve times the radius of B
k
.
In other words,

n
_
_
E
_
kK
B
k
_
_

kK

n
(

B
k
)
=

k>K
5
n

n
(B
k
) 0, as K,
since, as we have already noted,

n
(B
k
) =
n
_
_
k
B
k
_

n
(G) < .
For general E, look at the sets
E
m
= x E : m1 [[x[[ < m, m N.
Of course,

n
(E
_
m
E
m
) = 0
and each E
m
is as weve discussed above. Whats more, the G we chose at the very start of our
proof can be chosen to be the open set
x R
n
: m1 < [[x[[ < m
so the disjointness is achieved by passing from one E
m
to the next.
4. Hausdor nmeasure on R
n
A Hausdor gauge function is a map
h : [0, ][0, ]
such that h(0) = 0, h(t) > 0 whenever t > 0, h() , h is ascending and right-continuous from
[0, ) to [0, ).
Let (X, ) be a metric space. For E X let h(E) = h(diam E), h() = 0. Notice that h so
February 3, 2009
4. HAUSDORFF nMEASURE ON R
n
25
dened is a premeasure on any class ( containing . Let > 0. Consider the following classes ( of
subsets of X:
( = ( = G X : G is open
( = T = F X : F is closed
( = T = S X.
On each of these classes, we have intermediate measures

,
h

,
h

, and
h

(E) := inf
_

n
h(G
n
) : G
n
(, E
n
G
n
, diam G
n

_
,

(E) := inf
_

n
h(F
n
) : F
n
T, E
n
F
n
, diam F
n

_
,

(E) := inf
_

n
h(S
n
) : E
n
S
n
, diam S
n

_
,
and

(E) := inf
_

n
h(S
n
) : E =
n
S
n
, diam S
n

_
.
Proposition 4.0.16. If 0 < < then

=
h

=
h

.
Corollary 4.0.17. If
h
,
h
and
h
are the measures derived from h and the classes (, T and
T a la Method II then

h
=
h
=
h
.
Moreover

R
(E) = sup
>0

(E).
As C.A. Rogers noted, this shows that the
h
measure of E can be dened in terms of the
diameters and covering properties of subsets of E. Thus the
h
measure of E is an intrinsic
property of E as a set with the metric . So
h
(E) doesnt change when E is put (isometrically)
insider another metric space (X

) than (X, ).
Proof. Several assertions of the proposition are clear or, at the very least, easy-to-see. So
h

re-
quires covering by closed sets of diameter and so
h

. On the other hand, diam S = diam

S
so in fact,
h

=
h

.
Again its plain that
h

; yet if (S
n
) is a sequence of sets that covers E with each S
n
having
diameter and T
n
= S
n
E then T
n
also covers E, diam T
n
and
n
T
n
= E; it soon follows
that

(E)

n
h(T
n
)

n
h(S
n
),
February 3, 2009
26 3. INVARIANT MEASURES ON R
n
and so
h

. In sum:

=
h

.
Naturally,
h

and so the only missing ingredient to a complete proof is that

h

when
< . Here some epsilonics are called for. Let > 0. Cover E by a sequence (F
n
) of closed sets
with diameter so that

(E)

n
h(F
n
)
h

(E) +
(we clearly need only consider what happens when
h

(E) < ).
Choose > 0 so that
h(diam F
n
+ 2
n
) < h(F
n
) +

2
n
and
+ 2
n
< .
Let
U
n
= [d(x, F
n
) <
n
];
each U
n
is open, contains F
n
and
diam U
n
diam F
n
+
n
+
n
+ 2
n
< .
Further
h(U
n
) h(diam + 2
n
) h(F
n
) +

2
n
.
Hence E
n
U
n
each U
n
open, has diameter less than and

n
h(U
n
) h(diam F
n
+ 2
n
)

n
h(F
n
) +

2
n

h

(E) +.
It follows that

(E)
h

+.
Since > 0 was arbitrary, the proof is complete.
Let h(t) = t
n
. The Hausdor measure generated by h is denoted by
(n)
. Keep in mind that
the Hausdor measure is a misnomer; after all
(n)
makes sense in any metric space (X, ).
A comparison of
(n)
and ndimensional Lebesgue measure
n
on R
n
is worth our while. The
comparison proceeds in two steps. In the rst, well argue to their proportionality, in the second
step well complete the constant of proportionality.
Theorem 4.0.18. Let
C
0
= x = (x
1
, . . . , x
n
) R
n
: 0 x
i
< 1, i = 1, 2, . . . , n
and

n
=
(n)
(C
0
).
Then

(n)
(E) =
n

n
(E).
February 3, 2009
4. HAUSDORFF nMEASURE ON R
n
27
Proof. Both
(n)
and
n
are invariant under translations and are homogeneous of order n.
(i.e., for any t 0,
(n)
(tE) = t
n

(n)
(E) and
n
(tE) = t
n

n
(E).) It follows that

(n)
(C) =
n

n
(C)
for any cube of the form
C = x = (x
1
, . . . , x
n
) : a
i
x
i
< a
i
+s
where (a
1
, . . . , a
n
) R and s > 0.
Consider the collection ( of cubes of the form
_
x = (x
1
, . . . , x
n
) :
r
i1
2
n
x
i
<
r
i
2
n
_
where k 0 is an integer and r
1
, . . . , r
n
Z. ( enjoys the following property: if two members of (
have a point in common, then one is a subset of the other.
So what? Well if G is any open set then each point g G lies in a cube C ( that is con-
tained entirely within G and is maximized in the sense that C is not a subset of any larger member
of ( that is entirely inside G. Hence G is the union of the maximal cubes from ( and these cubes
are disjoint. Lets write
G =
m
C
m
,
where (C
m
) is the sequence of maximal cubes just described. Everything were talking about is
measurable and so

(n)
(G) =

(n)
(C
m
) =

n
(C
m
) =
n

n
(G).
So

(n)
(G) =
n

n
(G)
for open sets G R
n
.
Next we bootstrap this equalitys verity to (

sets. Let H be a (

subset of R
n
. If
(n)
(H)
and
(n)
(H) are both equal to + then theres little that need be said. Naturally, theyre also
nite at the same time. In fact, if
(n)
(H) < then we can cover H by a sequence (G
m
) of open
sets with

n
(diam G
m
)
n
< ; if we enclose each G
m
in an open cube U
m
where each edge of U
m
is twice the diameter of G
m
then H
m
U
m
and

n
(
m
C
m
)

n
(C
m
)

m
(2 diam G
m
)
n
2
n

m
(diam G
m
)
n
< .
On the other hand, if
n
(H) < then H can be covered by an open set of nite
n
measure and
so (since for open sets
(n)
and
n

n
agree),

(n)
(H)
(n)
(G) =
n

n
(G)
for any appropriate open set H G and
(n)
(H) < follows.
February 3, 2009
28 3. INVARIANT MEASURES ON R
n
Okay, suppose
(n)
(H) and
n

n
(H) are both nite. H is a (

set so we can write H in the

form
H =
m
G
m
,
where (G
m
) is a descending sequence of open sets each of nite measure (be that measure
(n)
or

n
!). It follows that

(n)
(H) = inf
m

(n)
(G
m
) = inf
m

n
(G
m
) =
n

n
(H),
thanks to the measurability of all sets involved. So

(n)
(H) =
n

n
(H)
for any (

set H R
n
.
Now look at any E R
n
. Choose (

sets H
1
, H
2
R
n
with E H
1
H
2
so

(n)
(E) =
(n)
(H
1
), and
n
(E) =
n
(H
2
).
Of course

(n)
(E)
(n)
(H
1
H
2
)
(n)
(H
1
) =
(n)
(E)
and

n
(E)
n
(H
1
H
2
)
n
(H
2
) =
n
(E)
so

(n)
(E) =
(n)
(H
1
H
2
) and
n
(E) =
n
(H
1
H
2
).
It soon follows that

(n)
(E) =
(n)
(H
1
H
2
) =
n

n
(H
1
H
2
) =
n

n
(E)
since H
1
H
2
is a (

set.
Theorem 4.0.19.

n
=
_
4

_
n/2

_
n + 2
2
_
.
Proof. To compute
n
we need to know the volume,
n
(B), of the closed unit ball in R
n
;
cutting to the quick, this is given by

n
(B) =

n/2

_
n+2
2
_.
Lets see what is involved. Let (S
k
) be a sequence of subsets of R
n
that cover the cube C
0
. Then
1 =
n
(C
0
)
n
(
k
S
k
)

n
(S
k
)

k
_

4
_
n/2
(diam(S
k
))
n

_
n+2
2
_
by the isodiametric inequality. It follows that
1
_

4
_
n/2

(n)
(C
0
)

_
n+2
2
_ =
_

4
_
n/2

_
n+2
2
_.
February 3, 2009
5. NOTES AND REMARKS 29
On the other hand, if > 0 is given to us then Vitalis covering theorem provides us with a sequence
(B
k
) of disjoint balls each of diameter less than and each contained in C
0
with

n
(C
0
) =

n
(B
k
), and
n
(C
0

k
B
k
) = 0.
It follows that

(n)
(C
0

k
B
k
) = 0
Hence

(n)

(C
0
)
(n)

(
k
B
k
) +
(n)

(C
0

k
B
k
)
=
(n)

(
k
B
k
)

k
(diam B
k
)
n
=

k
_
4

_
n/2

_
n + 2
2
_

n
(B
k
)
=
_
4

_
n/2

_
n + 2
2
_

n
(B
k
)
=
_
4

_
n/2

_
n + 2
2
_

n
(C
0
)
=
_
4

_
n/2

_
n + 2
2
_
.
So

n
=
(n)
(C
0
)
_
4

_
n/2

_
n + 2
2
_
,
as well. Therefore

n
=
_
4

_
n/2

_
n + 2
2
_
.

5. Notes and Remarks

The isoperimetric inequality
February 3, 2009
February 3, 2009
CHAPTER 4
Measures on Metric Spaces
Definition 0.0.20. A function dened on the sets of a space is called a measure on if
it satises the conditions
(i) (E) 0, (E) + for each E ;
(ii) () = 0;
(iii) (E) (F) if E F;
(iv) if (E
n
) is a sequence of subsets of then
(
n
E
n
)

n
(E
n
).
If is a measure on the space then E is said to be measurable if whenever A E and
B E
c
then
(A B) = (A) +(B).
Notice that E is measurable precisely when given A E and B E
c
,
(0.1) (A B) (A) +(B);
after all, (iv) assures that the reverse inequality holds regardless of E, A or B. In fact, we really
need only show (0.1) for sets A, B of nite measure. (WHY?)
Theorem 0.0.21 (Caratheodory). Let be a measure on the set . Then
(i) if (N) = 0 then N is measurable;
(ii) E is measurable if and only if E
c
is;
(iii) if (E
n
) is a sequence of measurable sets then
n
E
n
and
n
E
n
are too;
(iv) if (E
n
) is a sequence of pairwise disjoint measurable sets then
(
n
E
n
) =

n
(E
n
).
Proof. (of (i)) Suppose A N and B N
c
. Of course (A) = 0. Hence
(B) (A B) (A) +(B) = (B).
Squeezy says all are equal and so N is measurable. (ii) is plain. To start the proof of (iii) and
(iv), well show that the union of two measurable sets E, F is measurable. Take A and B to
be sets of nite measure with
A E F, B (E F)
c
.
Notice
A B = (A E) ((A B) E
c
).
31
February 3, 2009
32 4. MEASURES ON METRIC SPACES
Since
A E E, and (A B) E
c
E
c
,
by Es measurability,
(A B) = ((A E) ((A B) E
c
)) = (A E) +((A B) E
c
).
At the same time
(A B) E
c
= (A E
c
) B
with
A E
c
F
and
B (E F)
c
;
so Fs measurability assures that
((A B) E
c
) = (A E
c
) B) = (A E
c
) +(B).
But Es measurability is still in eect and says that
(A E) +(A E
c
) = (A),
so
(A B) = (A E) +((A B) E
c
)
= (A E) +(A E
c
) +(B)
= (A) +(B).
Therefore if E and F are measurable then so is EF. Next we consider a sequence (E
n
) of pairwise
disjoint measurable sets and write E =

n
E
n
. Let A and B be sets (of nite measure) with
A E, B E
c
.
Each of
n
m=1
E
m
is measurable and
(A B) ([A
n
m=1
E
m
] B)
= (A (
n
m=1
E
m
)) +(B),
because, after all,
A (
n
m=1
E
m
)
n
m=1
E
m
,
and
B (
n
E
n
)
c
=
n
E
c
n

n
m=1
(E
c
m
) = (
n
m=1
E
m
)
c
.
Since the sets E
n
, E
n1
, . . . , E
1
are pairwise disjoint and measurable,
(A ((
n
m=1
E
m
))) = ([A (
n1
m=1
E
m
)] [A E
n
]
= (A (
n1
m=1
E
m
)) +(A E
n
)
= = (A (
n2
m=1
E
m
)) +(A E
n1
) +(A E
n
)
= = (A E
1
) +(A E
2
) + +(A E
n
).
So we see that
(A B) (A (
n
m=1
E
n
) +(B)
=
n

m=1
(A E
m
) +(B),
February 3, 2009
4. MEASURES ON METRIC SPACES 33
and this is so for each n. Thus
(A B)

n
(A E
n
) +(B)
(A (
n
E
n
)) +(B) (since is subadditive)
= (A E) +(B)
= (A) +(B)
(A B).
All are equal; in particular, (AB) and (A) +(B) are equal and
n
E
n
is measurable. Next
if B = we get for any subset A of
n
E
n
,
(A) =

n
(A E
n
).
Letting A =
n
E
n
gives
(
n
E
n
) =

n
(E
n
).
This gives (iv) and (iii), at least in the case the sets are pairwise disjoint.
Alas from (ii) and the fact that any nite union of measurable sets is measurable, we see
that any nite intersection of measurable sets is also measurable .

To establish (iv) is now

It seems that
this sentence
belongs in the
previous
paragraph.
clear sailing: let (E
n
) be a sequence of (not necessarily pairwise disjoint) measurable sets. Write

n
E
n
as follows:

n
E
n
= E
1
(E
1
E
2
) (E
3
(E
1
E
2
)) ;
the result is what we wanted -
n
E
n
is the union of a sequence of pairwise disjoint measurable
sets and so is itself measurable.
Definition 0.0.22. A function dened on a class ( of subsets of a space is a premeasure
if
(i) (;
(ii) 0 (C) for all C (;
(iii) () = 0.
Theorem 0.0.23. If is a premeasure dened on a class ( of subsets of the space then the
set function
(E) := inf
_

i
(C
i
) : C
i
(, E
_
i
C
i
_
is a measure.
Here, as usual, the inmum over an empty set is +. The measure so constructed from is
said to be constructed by Method I.
Proof. It is plain and easy to see that the set function as dened from satises properties
(i)-(iii) of the denition of a measure. Only countable subadditivity requires a bit more proof. So
let (E
n
) be a sequence of subsets of the space , and lets show that
(
n
E
n
)

n
(E
n
).
February 3, 2009
34 4. MEASURES ON METRIC SPACES
For sure we can assume

n
(E
n
) < and so, in particular, each (E
n
) < as well. Now for
each n, choose a covering of E
n
, say C
j
(n)
j
from ( so that for the always present > 0 we have
(E
n
)

j
(C
j
(n)) (E
n
) +

2
n
.
Now C
j
(n) : j, n N covers
n
E
n
and so
(
n
E
n
)

j,n
(C
j
(n))

j
(C
j
(n))

n
_
(E
n
) +

2
n
_
=

n
(E
n
) +.
As expected, > 0 was arbitrary and its arbitrariness leads us to the conclusion that
(
n
E
n
)

n
(E
n
).
Let be a premeasure on the family ( of subsets of the metric space . Let > 0 and denote by
(

= C ( : diameter C .
Denote by

the restriction of to C

. The result is a premeasure that generates a measure

on
by Method I. A look at

is worth taking:

:= inf
_

i
(C
i
) : E
i
C
i
, C
i
(, diam C
i

_
.
Notice that

is a measure by Method I. Its plain that as gets smaller there are fewer sets with
diameter so

(E) gets bigger. Hence we can dene (E) by

(E) := sup
>0

(E),
and we say that generates by Method II.
Method I applies in a general setting, as opposed to Method II which relies on the metric structure
present in the underlying structure set .
Theorem 0.0.24. constructed by Method II is a measure.
The only possible stumbling point to this is the countable subadditivity so lets see why is
countably subadditive. To this end, let (E
n
) be a sequence of subsets of and consider these
quantities
(
n
E
n
) and

n
(E
n
).
Obviously the latter exceeds the former if its so we may as well assume

n
(E
n
) < .
February 3, 2009
4. MEASURES ON METRIC SPACES 35
Now for each > 0,

is a known measure so

(
n
E
n
)

(E
n
),
which in turn is

n
(E
n
).
The countable subadditivity follows from this.
A key ingredient to our mix is provided in the following. Comment on this theorem?
Theorem 0.0.25. Let be a measure on a metric space derived from the premeasure by
Method II. If A and B are non-empty subsets of that are positively separated then
(A B) = (A) +(B).
Here A and B positively separated means that there is a > 0 so that for any a A and
b B,
(a, b) .
Measures that enjoy this additive property are called metric measures and their importance lies
in the fact that these are precisely the measures on a metric space for which every Borel set is
measurable. Dene Borel sets here or somewhere else?
Proof. We need to show that if A and B are positively separated then
(A B) (A) +(B)
where all the terms involved are nite. The idea of the proof is to cover A, B and A B with very
ne covers from the domain ( of , so ne that we can distinguish between which sets are needed
to cover A and which to cover B.
More precisely, suppose
(a, b) > 0
for any a A and b B. Let > 0 announce its presence. By the rules of engagement
(A B) = sup
d>0
inf
_

i
(C
i
) : diamC
i
d, C
i
(, A B
_
i
C
i
_
.
So we can choose a sequence (C
i
) in ( so that
diam C
i
are all small;
A B
i
C
i
;

(C
i
) (A B) +.
Small ? Well to set things up for computation of the measure of A and of B through the
intermediaries

(A),

(B), we let
1
,
2
be positive numbers each less than diam , and let be
the number
= min
_

1
,
2
,

3
_
.
Choose as our model of small.
February 3, 2009
36 4. MEASURES ON METRIC SPACES
Heres the rst punchline; if each C
i
has diameter then each has diameter

3
and so
a given C
i
can intersect A or B but not both. It follows that

C
i
A=
(C
i
) +

C
i
B=
(B
i
)

i
(C
i
) (A B) +.
But each C
i
that intersects A has diameter
1
so knowing, as we do, that A is a subset of
_
C
i
A=
C
i
,
we get

1
(A)

C
i
A=
(C
i
).
Similarly,

2
(B)

C
i
B=
(C
i
).
The result:

1
(A) +

2
(B)

C
i
A=
(C
i
) +

C
i
B=
(C
i
) (A B) +,
and so Method II leads us to believe
(A) +(B) (A B) +.
Since > 0 was arbitrary, this proof is done.
Metric measures enjoy some very strong continuity properties. Heres one of them. Comment
on the importance of this proposition.
Proposition 0.0.26. Let be a metric measure on a metric space . Suppose (A
n
) is an
increasing sequence of subsets of so that A
n
and A
c
n+1
are positively separated. Then
(
n
A
n
) = sup
n
(A
n
).
We defer the proof until after giving the Proposition a chance to show o.
Theorem 0.0.27. If is a metric measure on the metric space then every closed subset of
is measurable.
Comment that this tells us that Borel sets are measurable.
Proof. Suppose F is a closed subset of the metric space , and let A F and B F
c
be
non-empty sets. For each n let
B
n
:=
_
x B : inf
yF
(x, y) >
1
n
_
.
Notice that B
n
B.
By their very denition, each B
n
is positively separated from A. Well show that each B
n
is
also positively separated from B
c
n+1
. To this end, let x B
n
and u B
c
n+1
. Now u , B
n+1
so
1
n + 1
inf
yF
(x, y);
February 3, 2009
4. MEASURES ON METRIC SPACES 37
There is a y
0
F so that
1
n +
1
2
(u, y
0
).
Now if
(x, u)
1
2
n
_
n +
1
2
_,
then
inf
yF
(x, y) (x, y
0
)
(x, u) +(u, y
0
)

1
2
n(n +
1
2
)
+
1
_
n +
1
2
_,
which by the method of common denominators, is
=
1
2
+n
n
_
n +
1
2
_ =
1
n
.
Hence x , B
n
. Our conclusion? If x B
n
and u B
c
n+1
, then
(x, u) >
1
2
n
_
n +
1
2
_ =
1
n(2n + 1)
.
i.e., B
n
and B
c
n+1
are positively separated.
Lets compute (A B) :
(A B) sup
n
(A B
n
)
= sup
n
(A) +(B
n
)
= (A) + sup
n
(B
n
)
= (A) +(
n
B
n
) by Proposition 0.0.26
= (A) +(B),
and F is measurable.
Now we look to the rather elegant proof of Proposition 0.0.26:
Proof. We have
A
1
A
2
A
n

with A
n
and A
c
n+1
positively separated for each n. We want to show that (
n
A
n
) sup
n
(A
n
)
and in this eort we may plainly suppose sup
n
(A
n
) < since otherwise all is okay.
First we look at the dierence sequence
D
1
= A
1
, D
2
= A
2
A
1
, D
3
= A
3
A
2
, . . . , D
n
= A
n
A
n1
, . . .
Of course, D
n
A
n
but further
D
2
A
c
1
, D
3
A
c
2
A
c
1
, D
4
A
c
3
A
c
2
A
c
1
, . . .
February 3, 2009
38 4. MEASURES ON METRIC SPACES
So
D
1
A
1
and D
3
, D
4
, . . . A
c
2
,
D
2
A
2
and D
4
, D
5
, . . . A
c
3
,
etc., etc., etc. In particular
D
1
A
1
, D
3
A
c
2
D
2
A
2
, D
4
A
c
3
D
1
D
3
A
3
, D
5
A
c
4
D
2
D
4
A
4
, D
6
A
c
5
.
.
.
D
1
D
3
D
2n1
A
2n1
, D
2n+1
A
c
2n
D
2
D
4
D
2n
A
2n
, D
2n
A
c
2n+1
.
It follows (inductively if you must know) that
(D
1
) +(D
3
) + +(D
2n1
) +(D
2n+1
) = (
n+1
k=1
D
2k1
),
and
(D
2
) +(D
4
) + +(D
2n
) +(D
2n+2
) = (
n+1
k=1
D
2k
).
Each of
n+1
k=1
D
2k1
and
n+1
k=1
D
2k
are subsets of A
2n+2
and so all above nd themselves
(A
2n+2
) sup
n
(A
n
) < .
Conclusion: both series

n
(D
2n1
) and

n
(D
2n
) converge. Now
(
n
A
n
) = (A
n
D
n+1
D
n+2
) Not sure about this one.
(A
n
) +(D
n+1
) +(D
n+2
) +
sup
n
(A
n
) +

k=n+1
(D
k
);
if > 0 be given then there is an n so that the latter sum is less than so
(
n
A
n
) sup
n
(A
n
) +.
Enough said.
Weve seen that every Borel set in a metric space is measurable whenever is a metric
measure on . Its worthwhile to note that this is not accidental; its part and parcel of being a
metric measure. Indeed if we suppose is a measure on the metric space for which every closed
subset of is measurable and if A, B are positively separated (so A B = ), then
(A B) = ((A B) A) +((A B) A
c
)
by the measurability of A. But
(A B) A = A and (A B) (A)
c
= B
by the positive separation of A and B. So
(A B) = ((A B) A) +((A B) (A)
c
= (A) +(B),
and is a metric measure.
February 3, 2009
4. MEASURES ON METRIC SPACES 39
We turn now to regularity properties of measures on a metric space, that is, how well we can
approximate a typical value (E) of by values at good sets - closed, open, T

, (

or even com-
pact. REWRITE THIS SENTENCE: Before we go any further we hasten to say that the problems
were concerned with are previously with metric measures that are not nite. To be sure, we em-
phasize that if measure is nite then good things happen (at least in the world of Borel sets).
Heres (some of) whats so.
Theorem 0.0.28. Let (X, ) be a metric space. Let be a countably additive map from the
Borel eld B/(X) to [0, ). Dene this eld?
(i) For any Borel set B in X, we have
sup(F) : F B, F closed = (B) = inf(G) : B G, G open.
Should we change G to U since we use U in the proof ?
(ii) If X is Polish (that is, there is a complete metric that generates Xs topology and X is
separable) then for any Borel set B in X
(B) = sup(K) : K B, K compact.
This result nds so much use in what we do, we submit a proof in detail.
Proof. (of (i)) Heres the trick to nd an approach to general Borel sets; we consider the
collection o of Borel sets in X that satisfy the conclusion of (i).
Open sets belong to o. If U is open then its trivial to approximate (U) from above by s
values on open super sets of U. How to approximate from within by s values at closed subsets?
Well, notice that if U is an open set in X then U
c
is closed and so x U
c
precisely when d(x, U
c
) = 0
where
d(x, U
c
) = infd(x, y) : y U
c
.
But d(x, U
c
) is a continuous function of x X and so we see that
U =
_
nN
_
d(, U
c
)
1
n
_
,
and U is an T

set. If we let F
n
=
_
d(, U
c
)
1
n

then each F
n
is closed,
F
1
F
2
F
n
U,
and so
(U) = lim
n
(F
n
).
Open sets are T

s and so each open set belongs to o.

Now o, by its very denition, can be described in a manner suited to epsilonics. Indeed, a Borel
set B o precisely when B satises
given > 0 there is an open set U containing B and a closed set F contained in B such that
(UF) < .
Its plain from this that o is closed under complements.
February 3, 2009
40 4. MEASURES ON METRIC SPACES
What about countable unions? Well suppose (B
k
) is a sequence of Borel sets each one of which is
in o. For each n, pick U
n
and F
n
so that U
n
is open, F
n
is closed, F
n
B
n
U
n
and
(U
n
F
n
) <

2
n+1
.
Let U =
n
U
n
, F =
n
F
n
. Then
F
n
B
n
U.
Moreover
(UF) = (
n
U
n

n
F
n
) (
n
(U
n
F
n
))

n
(U
n
F
n
)

2
n+1
=

2
.
Now U is open and F is almost closed. In fact, F =
n
F
n
so there is an n
0
so that
(F
n>n
0
F
n
) <

2
;
The result:
n
0
_
n=1
F
n

_
n
B
n
U,
where

n
0
n=1
F
n
is closed, U is still open and
(U
n
0
n=1
F
n
) < .
Therefore o is a collection of Borel sets containing Xs topology, (containing?) closed under com-
plementation and the taking of countable unions. Thus o = B
0
(X).
(ii) hints at the plentitude of compact sets in Polish spaces. This comes from the (existence of a)
complete metric; in complete metric spaces, subsets are relatively compact precisely when theyre
totally bounded. So to prove (ii) well use the separability to construct good closed and totally
bounded approximants. We start with a complete separable metric space (X, ) and show (ii) holds
for the Borel set X itself. Let x
n
: n N be a countable dense subset of X. For each n N
X =
k
B
1/n
(x
k
)
where
B

(x) = y X : (x, y) < .

For each n we can nd a k
n
so that

_
k
n
_
k=1
B
1/n
(x
k
)
_
is within

2
n
of (X). Look at
K =

n
_
k
n
_
k=1
B
1/n
(x
k
)
_
.
Since K is complete, its closed. K is totally bounded, too: after all, for any n,
K
k
n
_
k=1
B
1/n
(x
k
)
February 3, 2009
4. MEASURES ON METRIC SPACES 41
so each point of K is within 2/n of an x
k
, k = 1, . . . , k
n
. Therefore K is compact.
(K
c
) =
__

n
_
k
n
_
k=1
B
1/n
(x
k
)
__
c
_
=
_
_
n
_
k
n
_
k=1
B
1/n
(x
k
)
_
c
_

__
k
n
_
k=1
B
1/n
(x
k
)
_
c
_

__
k
n
_
k=1
B
1/n
(x
k
)
_
c
_

2
n
= .
Generally if we apply (i) we can nd inside any Borel subset B of X a closed set F with (F) within
/2 of (B); then dene on the Borel sets of X by
(E) = (E F).
Our opening salvo applied to gives a compact K X so (K) is within /2 of (X) = (F).
Then K F does the dirty deed since (K F) = (K), which is within of (B).
Before entering into the subject of regularity of metric measures, there are a few apt comments
to be made regarding the theorem above. First regarding (ii) an example might highlight the com-
pleteness hypothesis. Suppose is a subset of [0, 1] for which the inner measure
k
() is 0 and
the outer measure
k
() is 1. Of course is not Lebesgue measurable and hopelessly incomplete.
However is separable. Now the Borel subsets of are just those subsets of of the form B
where B is a Borel subset of [0, 1]. If we dene P(B ) to be the Lebesgue measure of B then
P is

[
B
0
(). P is a probability Borel measure on the separable metric space . However, if K is
a compact subset of then K is also a compact subset of [0, 1] and so a Borel subset of [0, 1]; it
follows that
P(K) =

(K) =

(K) = 0.
P is not a regular Borel measure on . Comments on this example:
Should we say something about inner measure and outer measure? What is
k
?
How do we know anything about ? eg Why is not Lebesgue measurable and incom-
plete?
What is

?
Should we dene probability here or somewhere prior to this?
Some kind of completeness assumption is needed in (ii). How about separability? Well most
probabilities have separable support. Should we dene separable support? Heres the fact.
Theorem 0.0.29. Let be a metric space. Then in order that every probability Borel measure
on have a separable support it is both necessary and sucient that each discrete subset of have
non-measurable cardinal.
February 3, 2009
42 4. MEASURES ON METRIC SPACES
Recall that a set S has measurable cardinal if one can dene on 2
S
a probability measure that
vanishes on singletons; otherwise S has non-measurable cardinal. Its unknown if sets measurable
cardinals exist. It is known that if they do, then they must be huge! So the comforts of nite
measure aside, lets discuss general metric measures and approximation properties thereof. First, a
general denition: if is a measure on a space and is a family of subsets of then we call
-regular if for any E there is an R so that E R and (E) = ().
Now some facts.
If is a measure on a space generated by a premeasure on a collection ( with (
by Method I, then is C

regular. What does this mean? Since (, is certainly

C

and so for sets E with (E) = , E serves us well. So we look at Es with

(E) < . In this case, epsilonics enter the foray: for each n there is a sequence (C
n
i
)
i
of members of ( such that
E
i
C
n
i
, and (E)

i
(C
n
i
) (E) +
1
n
.
Set
D =
n

i
C
n
i
C

.
Then E D and for each n
(E) (D)

i
(C
n
i
) (E) +
1
n
.
It follows that (E) = (D).
Let be a measure on a metric space generated by a premeasure on a class ( which
contains via Method II. Then is C

regular.
Proof. As before we use ( to rid us nuisances: take E ; if (E) =
then ( C

contains E and plainly (E) = () = . So we can restrict our

attention to E such that (E) < . But now each

(E) is also nite since

for
any > 0. Because each

(E) < we can also nd a C

such that E C

and

(E) =

(C
n
). To keep tabs on (E) for each n let C
n
C

be so that E C
n
and

1/n
(E) =
1/n
(C
n
).
Put
C =
n
C
n
C

.
E C and

1/n
(E)
1/n
(C)
1/n
(C
n
) =
1/n
(E) (E).
Choosing carefully from this we see

1/n
(E)
1/n
(C) (E).
Taking n , we get
(E) = (C),
and thats all she wrote.
Each of these facts has interesting consequences in case ( is special.
February 3, 2009
4. MEASURES ON METRIC SPACES 43
Corollary 0.0.30. Suppose is a measure on a topological group generated by a premeasure
dened on the topology of the space by Method I. Then is (

regular. If is a measure on a
metric space generated by a premeasure dened on the topology of the space by Method II, then is
(

regular.
Corollary 0.0.31. Suppose is a measure on a topological space generated by a premeasure
dened on the Borel subsets of by Method I. Then is Borel-regular. Suppose is a measure on
a metric space generated by a premeasure dened on the Borel subsets of by Method II. then
is Borel-regular.
We have the following general principle for approximating from within:
If is an -regular measure on a space and E is a measurable subset of with (E) <
then there is a set R
1
R
2
where R
1
, R
2
, contained in E so that
(R
1
R
2
) = (E).
Choose R
1
with E R
1
so that
(E) = (R
1
).
E is measurable so
(E0 = (R
1
) = (R
1
E) +(R
1
E
c
) = (E) +(R
1
E
c
).
But (E) < so
(R
1
E
c
) = 0.
is still regular so there is an R
2
so that
R
1
E
c
R
2
and 0 = (R
1
E
c
) = (R
2
).
Now
R
1
R
2
(R
1
E
c
)
= R
1
(R
1
E6c)
c
= R 1 (R
c
1
E)
= (R
1
R
c
1
) (R
1
E) = E.
Also (R
2
) = 0. So
(R
1
R
2
) = (R
1
R
2
) +(R
2
)
((R
1
R
2
) R
2
)
(R
1
)
= (E)
(R
1
R
2
).
Voila!
February 3, 2009
February 3, 2009
CHAPTER 5
Banach and Measure
Like most (abstract) analysts of his day, Banach also took a keen interest in developments
related to the existence, uniqueness and uses of Haar measure. In this chapter we recount Banachs
views on the subject. Naturally we open with a bit of functional analysis and present the real case
of the Hahn-Banach theorem. We follow this with a derivation of the existence of Banach limits.
After a brief discussion of weak topologies, we pass to Banachs remarkable characterization of
weakly null sequences of bounded functions on a set. That is followed by Banachs characterization
of weakly null sequences in the space C(Q), where Q is a compact metric space, a result that was
derived before Banach knew what C(Q)

was! It wasnt long before Banach was able to compute

C(Q)

. We follow his way of doing this, beneting from his view of abstract Lebesgue integration,
`a la a Daniell type construction. Finally we present Banachs construction of invariant measures.
His proof was a source of inspiration for many and will reappear in our discussion of Steinlages
description of existence and uniqueness of measures on locally compact spaces that are invariant
under group actions.
1. A bit of Functional Analysis
Theorem 1.0.32 (Hahn-Banach Theorem). Let X be a real linear space, and let S be a linear
subspace of X. Suppose that p : X R is a subadditive positively homogeneous functional and
f : S R is a linear functional with f(s) p(s) for all s S. Then there is a linear functional F
dened on all of X such that F(x) p(x) for all x X, and F(s) = f(s) for all s S.
Proof. Our rst task is to see how to extend a functional like f one dimension at a time
preserving the domination by p. With this in mind, let x XS and notice that for any linear
combination s + x of a vector in S and x, whatever the linear extension Fs value at x (say that
value is c) we must have
F(s +x) = F(s) +F(x) = f(s) +c.
So it must be (if we are to have F dominated by p) that
f(s) +c p(s +x)
holds for all s S and all real s. Well follow where this leads us; for all R and s S we have
to have
c p(s +x) f(s).
For > 0 this tells us that
c
1

(p(s +x) f(s))

= p
_
s

+x
_
f
_
s

_
;
45
February 3, 2009
46 5. BANACH AND MEASURE
while if < 0 then ( > 0 and)
c f(s) p(s +x),
or since > 0,
c
1

f(s)
1

p(s +x)
= f
_
s

_
p
_
s

x
_
Taking into account the linearity of S we see that what we seek is a c R (which will be F(x)) so
that regardless of s, s

S satises
f(s

) p(s

x) c p(s +x) f(s).

Does such a c exist? You bet! After all if s, s

S then
f(s) +f(s

) = f(s +s

) p(s +s

) p(s x) +p(s

+x).
So for all s, s

S
f(s) p(s x) p(s

+x) f(s

),
and we can chose c in an appropriate manner.
Now we know that we can extend linear functionals one dimension at a time while preserving
ps domination. Its time for some transnite hijinks. We consider the collection of all linear func-
tionals g dened on a linear subspace Y of X such that S Y , g[
S
= f and on Y , g p. We
partially order this collection by saying that g
1
g
2
if g
2
is an extension of g
1
(so g
1
is dened
on a linear subspace Y
1
that is contained in g
2
s domain).
Hausdors maximal principle ensures us that there is a maximal linearly ordered subfamily g

of linear extensions of f so that on g

s domain Y

(which contains S), g

p. We dene F on
the linear space thats the union of the domains of the g

s as one mine expect! F(x) = g

(x) if x
is in g

s domain. Because of the ordering described above, the domain of F is linear subspace of
X and on that domain F is linear and dominated by p. Of course, F is a linear extension of f and
the domain of F must be all of X - this is assured us by the opening salvo.
We put the Hahn-Banach theorem to immediate use by establishing the existence of generalized
limits or Banach limits, as well refer to them henceforth. We will call on two spaces: l

, the space
of bounded, real-valued sequences, and the linear subspace c of l

consisting of all the convergent

sequences. Typically if x l

then
[[x[[

= sup(x
n
) : x N.
Theorem 1.0.33 (Banach). There exists a linear functional LIM on l

such that
(i) LIM(x) 0 if x = (x
n
) l

and x
n
0 for all n;
(ii) [LIM(x)[ [[x[[

, for all x l

;
(iii) If x l

and Tx = (x
2
, x
3
, . . .) for x = (x
1
, x
2
, . . .), then LIM(Tx) = LIM(x);
(iv) For any x l

liminf
n
x
n
LIM(x) limsup
n
x
n
.
February 3, 2009
1. A BIT OF FUNCTIONAL ANALYSIS 47
Proof. Let p : l

R be given by
p(x) = limsup
n
x
1
+ +x
n
n
.
It is easy to see that p is subadditive and positively homogeneous. Next, let f : cR be the linear
functional
f(x) = lim
n
x
n
.
Since f and p agree on c, f(x) p(x) is trivially satised. LIM is any Hahn-Banach extension of f
to all of l

. LIM is a linear functional on l

such that for any x l

LIM(x) p(x).
To see (i), take x l

and suppose x
n
0 for all n. Then
LIM(x) = LIM(x) p(x),
so
LIM(x) p(x) = limsup
n
x
1
x
2
x
n
n
= liminf
n
x
1
+x
2
+ +x
n
n
0,
because x
n
0 for all n.
(ii) is plain since p(x) [[x[[

for all x l

. To see (iii), take x l

, then
p(x Tx) = limsup
n
x
1
x
n+1
n
= 0,
since x is bounded. It follows that
LIM(x Tx) p(x Tx) = 0.
So for any x l

(1.1) LIM(x) LIM(Tx).

This applies as well to x so
LIM(x) = LIM(x) LIM(Tx) = LIM(Tx),
and so
(1.2) LIM(x) LIM(Tx)
as well. LIM(x) = LIM(Tx) is the only conclusion that can be drawn from (1.1) and (1.2).
For (iv), let > 0. Find N N so
inf
n
x
n
x
N
inf
n
x
n
+.
Then
x
n
+ x
N
inf
n
x
n
+ x
N
0.
Hence
0 LIM(x + x
N
) = LIM(x) + x
N
February 3, 2009
48 5. BANACH AND MEASURE
by (i). Hence
inf
n
x
n
x
N
LIM(x) +.
Since > 0 was arbitrary
inf
n
x
n
LIM(x).
For any k
inf
nk
x
n
= inf
nk
T
k
x
n
LIM(T
k
x) = LIM(x);
but
liminf
n
x
n
= sup
n
inf
kn
x
k
sup
n
LIM(x) = LIM(x).
Again using LIMs linearity, we see that
limsup
n
x
n
= liminf
n
(x
n
) LIM(x) = LIM(x).
The proof of the existence of Banach limits appears.
Let X be a normed linear space. A linear functional on X is bounded if there is a c > 0 so
that [f(x) c[[x[[ for all x X. The appellation bounded pertains to fs boundedness on the
closed unit ball
B
X
= x X : [[x[[ 1
of X. It is easy to show that a linear functionals boundedness is tantamount to its continuity and
this, in turn, is assured by fs continuity at the origin. Then
[[f[[ = sup
xB
X
[f(x)[
is a norm on X

and is in fact, complete. So Cauchy sequences in X

with respect to this just-

dened norm, converge.
The Hahn-Banach Theorem provides any normed linear space with lots of linear functionals in
its dual.
Let S be a linear subspace of the real normed linear space X, and let s

be a continuous
linear functional on S. Then there is an extension x

of s

to all of X so that [[s

[[ = [[x

[[.
We simply let p(x) = [[s

[[ [[x

[[ and apply the Hahn-Banach Theorem to nd a linear

functional F dened on all of X with F(x) p(x) for all x X. Because
F(x) p(x) = p(x),
we see that
F(x) p(x) = [[s

[[ [[x

[[
for all x X and so F = x

is in X

and [[F[[ [[s

[[. Since F extends s

, [[s

[[ [[F[[
too.
Let x be a member of the (real) normed linear space X. Then there is an x

with
[[x

[[ = 1 so that x

(x) = [[x[[. Indeed, let

S = x : R,
dene s

(x) = [[x[[, apply the Hahn-Banach Theorem and get x

, the norm-preserving
extension of s

.
February 3, 2009
1. A BIT OF FUNCTIONAL ANALYSIS 49
It is frequently the case that for a particular normed linear space the norm-topology is too coarse
to uncover special and delicate phenomena peculiar to that space. Oft times such phenomena are
best described using the weak topology.
Definition 1.0.34. A base for the weak topology of the normed linear space X is given by
sets of the form
U(x; x

1
, x

2
, . . . , x

n
) = y X : [x

i
(x y)[ < , i = 1, . . . n
where x X, x

1
, . . . , x

n
X

, and > 0.
The topology generated by this base is easily seen to be a Hausdor linear topology so the
operations of (x, y)x+y and (, x)x of XX to X and RX to X, respectively, are continuous.
This topology is never metrizable and never complete for the innite-dimensional normed linear
spaces! Usually closures are not determined sequentially. Nevertheless, weak sequential convergence
in special spaces often holds secrets regarding the inner nature of such spaces. To be sure, if X
is a normed linear space and (x
n
) is a sequence of vectors in X then we say that (x
n
) converges
weakly to x X (sometimes denoted by x = weak- lim
n
x
n
) if for each x

,
lim
n
x

(x
n
) = x

(x).
It is an integral part of basic functional analysis to compute the duals of special normed linear
spaces and using the specic character of the spaces involved to characterize when sequences are
weakly null (tend to zero weakly).
We now turn our attention to life inside spaces of the form l

(Q), where Q is a set and l

(Q)
denotes the normed linear space of all bounded real-valued functions x dened on Q, where xs
norm is given by
[[x[[

= sup[x(q)[ : q Q.
Now to bring tools like the Hahn-Banach Theorem to bear on the study of l

(Q), we need to know

something about l

(Q)

. Banach knew a great deal about this (as did F. Riesz before him); he
didnt formulate an exact description of l

(Q)

but nevertheless, understood the basics. To begin,

if x

(Q)

, then x

is entirely determined by its values at members of l

(Q) of the form

E
where E Q; after all, simple functions are dense in l

(Q). Now if F(E) = x

(
E
) then F is a
bounded nitely additive real-valued measure on 2
Q
, the collection of all subsets of Q. If we dene
[F[ by
[F[(E) = supF(S) : S E
for E Q then [F[ is a non-negative real-valued function dened on 2
Q
and [F(E)[ [F[(E) for
each E Q. (Remember:

= 0 so F() = x

) = x

(0) = 0.) Whats more, [F[ is also nitely

additive! If G E
1
E
2
where E
1
and E
2
are disjoint subsets of Q then G = (G E
1
) (G E
2
),
and so
F(G) = F((G E
1
) (G E
2
))
= F(G E
1
) +F(F E
2
) [F[(E
1
) +[F[(E
2
);
It follows that
[F[(E
1
E
2
) [F[(E
1
) +[F[(E
2
).
February 3, 2009
50 5. BANACH AND MEASURE
On the other hand, if E
1
and E
2
are disjoint subsets of Q then for any > 0 we can pick G
1
E
1
and G
2
E
2
so
[F[(E
1
) F(G
1
) +

2
, [F[(E
2
) F(G
2
) +

2
.
But now
[F[(E
1
) +[F[(E
2
) F(G
1
) +

2
+F(G
2
) +

2
= F(G
1
) +F(G
2
) +
= F(G
1
G
2
) + (since F is nitely additive)
[F[(E
1
E
2
) +;
since > 0 was arbitrary, we see that
[F[(E
1
) +[F[(E
2
) [F[(E
1
E
2
)
whenever E
1
and E
2
are disjoint subsets of Q.
Heres the punch line: if x

(Q)

then x

denes a bounded nitely additive measure on

2
Q
- call this measure F. From F we generate [F[, all of whose values are non-negative; [F[ F is
also a non-negative bounded real-valued nitely additive map on 2
Q
and F = [F[ ([F[ F). So F
is the dierence of non-negative bounded nitely additive maps on 2
Q
. In turn, such non-negative
additive bounded maps on 2
Q
dene positive linear functionals on l

(Q), functionals that are nec-

essarily bounded linear functionals.
Why is this last statement so? Well suppose G : 2
Q
[0, ) is nitely additive. If A B Q then
B = A (BA) so
G(B) = G(A (BA)) = G(A) +G(BA) G(A);
it follows that for any E Q, G(E) G(Q), and G is bounded by G(Q). Moreover if E
1
, . . . E
k
are pairwise disjoint subsets of Q and a
1
, . . . , a
n
R then

in
a
i
G(E
i
)

in
[a
i
[G(E
i
) sup
1in
[a
i
[

in
G(E
i
)
=

in
a
i

E
i

G
_
_
_
in
E
i
_
_
G(Q)

in
a
i

E
i

and so G determines a linear functional

a
i

E
i

a
i
G(E
i
)
on the simple functions which is bounded there on. This bounded linear functional extends to
l

(Q) in a bounded linear fashion, a positive functional to be sure.

Now x B
l

(Q)
means [x(q)[ 1 for all q Q; it follows that 1 x(q) 1 for all q Q.
February 3, 2009
1. A BIT OF FUNCTIONAL ANALYSIS 51
If f is a positive linear functional dened on l

(Q), so f(x) 0 whenever x(q) 0 for all q Q,

then f(1) f(x) f(1), or [f(x)[ f(1), whenever x B
l

(Q)
. Positive linear functionals on
l

(Q) are bounded linear functionals.

Theorem 1.0.35 (Banach). Let Q be a (non-empty) set and (x
n
) be a (uniformly) bounded
sequence in l

(Q). Then (x
n
) is weakly null if and only if
(1.3) lim
n
liminf
k
[x
n
(q
k
)[ = 0
for each sequence (q
k
) of points in Q.
Proof. Necessity: Suppose to the contrary that there is a sequence of points (q
k
) in Q such
that
limsup
n
liminf
k
[x
n
(q
k
)[ > > 0
for some . Then unraveling the meaning of limsups, we can nd a strictly increasing sequence
(n
j
) of positive integers such that
liminf
k
[x
n
j
(q
k
)[ > > 0
for each j. Now turning to the meaning of liminf, we nd a subsequence (q
k
m
) of (q
k
) such that
[ lim
m
x
n
j
(q
k
m
)[ > > 0
for each j. Let x

(Q) be given by
x

(x) = LIM((x(q
k
m
)
m
))
where LIM l

(N)

is a Banach limit. Then for each j,

[x

(x
n
j
)[ >
and so
limsup
n
[x

(x
n
)[ > > 0.
It follows that (x
n
) is not weakly null in l

(Q).
Suciency: By remarks preceding the statement of this theorem, to test (x
n
)s weak nullity it
suces to check the action of x

(Q)

, for x

a positive linear functional of norm 1, on the

sequence (x
n
). So suppose x

is such a functional, with

limsup
n
x

(x
n
) > > 0
where (1.3) holds:
lim
n
liminf
k
[x
n
(q
k
)[ = 0
for each sequence (q
k
) of points in Q. Let s
n
be the sequence
s
n
(q) =
_
x
n
(q) if x
n
(q) 0
0 otherwise
and let t
n
= x
n
s
n
.
One of limsup
n
x

(s
n
) and limsup
n
x

(t
n
) must exceed

2
; after all, x
n
= s
n
+t
n
so
x

(x
n
) = x

(s
n
) +x

(t
n
)
February 3, 2009
52 5. BANACH AND MEASURE
ensuring that
limsup
n
x

(x
n
) = limsup
n
(x

(q
n
) +x

(t
n
)) limsup
n
x

(s
n
) + limsup
n
x

(t
n
).
If we clip o the bottoms of s
n
by dening
y
n
(q) =
_
s
n
(q) if s
n
(q)

6
0 otherwise
then
[[s
n
y
n
[[

6
;
whats more,
limsup
n
x

(y
n
) = limsup
n
x

(s
n
(s
n
y
n
))
= limsup
n
(x

(s
n
) x

(s
n
y
n
))
so that
limsup
n
x

(y
n
)

2


6
=

3
.
Let
S
n
=
_
q Q : [x
n
(q)[

6
_
,
and look at
S
n
. Since
[[y
n
[[

[[s
n
[[

[[x
n
[[

M,
say, we see that for any q Q

S
n
(q)
y
n
(q)
M
so that
x

(M
S
n
) x

(y
n
).
From this we see that
limsup
n
x

(
S
n
)

3M
=: > 0.
For E Q, let F(E) = x

(
E
); of course F l

(Q)

and
limsup
n
F(S
n
) > .
Let n
1
be the smallest positive integer such that
limsup
n
F(S
n
S
n
1
) > 0.
Such an n exists by the way! This is the crux of the matter! In fact, otherwise,
lim
n
F(S
n
S
k
) = 0
for each k so (because F is additive)
lim
n
F(
k
j=1
(S
n
S
j
) = 0,
for each k as well. Let k
1
= 1. Pick m
1
> k
1
so large that
F(S
m
1
) >
and
F(S
m
1
S
k
1
) <

2
.
February 3, 2009
1. A BIT OF FUNCTIONAL ANALYSIS 53
Let k
2
> m
1
. Pick m
2
so large that
F(S
m
2
) >
and
F(S
m
2
(S
1
S
k
2
)) <

2
.
Continuing in this fashion, producing k
1
< m
1
< k
2
< m
2
< , with
F(S
m
j
) >
and
F(S
m
j
(S
1
S
k
j
)) <

2
.
Now disjointivity: let T
j
be given by
T
j
= S
m
j
[S
m
j
(
k
j
i=1
S
i
)].
By construction
F(T
j
) >

2
and this is a no-no since F takes disjoint sequences to 0.
So n
1
does indeed exist such that
limsup
n
F(S
n
S
n
1
) > 0.
Believe it or not.
Once faith has been established for n
1
we see that there are n
2
< n
3
< < n
k
< so that
limsup
n
F(S
n
S
n
1
S
n
k
) > 0
for each k. The all-important point here is that for each k there is at least one point q
k
so
q
k
S
n
1
S
n
2
S
n
k
.
Of course if j k then q
j
S
n
k
and so by how the S
n
s were dened
[x
n
k
(q
j
)[

6
.
It soon follows that
limsup
n
liminf
j
[x
n
(q
j
)[

6
.
This contradicts (1.3) and thus, the suciency is proven.
This result is remarkable because it characterizes weak convergence in a highly non-separable, non-
metrizable situation in terms of sequences q
k
in Q.
Suppose Q is a non-void compact metric space. Then the space C(Q) of all continuous functions
real-valued functions dened on Q, equipped with the norm
[[x[[

= sup[x(q)[ : q Q
is a closed linear subspace of l

(Q), the space of all bounded real-valued functions on Q. We can

give a much more succinct characterization of when a bounded sequence (x
n
) is weakly null in C(Q)
then just what Banach has above. Indeed, and again this was observed by Banach.
February 3, 2009
54 5. BANACH AND MEASURE
Theorem 1.0.36 (Banach). A (uniformly) bounded sequence (x
n
) in C(Q) is weakly null if and
only if
lim
n
x
n
(q) = 0
for each q Q.
Proof. If (x
n
) C(Q) is weakly null and q Q then using the point evaluation

q
C(K)

, where
q
(x) = x(q)
we see that
0 = lim
n

q
(x
n
) = lim
n
x
n
(q).
Now assume that (x
n
) is a (uniformly) bounded sequence in C(Q) for which lim
n
x
n
(q) = 0 for each
q Q and imagine that (x
n
) is not weakly null in C(Q). Of course the Hahn-Banach theorem tells
us that (x
n
) is not weakly null in l

(Q) either. So by Theorem 1.0.35 there must be a subsequence

(x

n
) of (x
n
) and a sequence of points (q
k
) in Q and an > 0 so that for each n
(1.4) liminf
k
[x

n
(q
k
)[ > 0.
But (q
k
) is a sequence in the compact (hence sequentially compact) metric space Q and so (q
n
) has
a subsequence (q

k
) that converges to some q
0
Q. By (1.4) it must be that for all n
[x

n
(q
0
)[ > 0.
OOPS.
2. The Lebesgue Integral on Abstract Spaces
In this section we present Banachs approach to the Lebesgue integral in abstract spaces.
Banachs approach is a Daniell-like construction built using his clear and deep understanding of
limsups and liminfs. His starting point is a positive linear functional f acting on a vector lattice
( of real-valued functions dened on some set K; we suppose (with Banach) that the functional
satises a kind of Bounded Convergence Theorem (BC) on (. It is important to note that in the
previous section we presented Banachs famous result characterizing weakly convergent sequences
in spaces C(Q), Q a compact metric space; it follows from this that should the initial vector lattice
( be such a C(Q), then every positive linear functional satises the (BC) hypothesis.
After an initial discussion of technical consequences of the (BC) hypothesis involving limsups
and liminfs of functions in (, Banach introduces an upper and a lower integral. Were we doing
measure theory, this piece of the puzzle would be concerned with properties of outer and inner
measures generated from an initial set function.
Next the class of integrable functions is isolated, being identied as those real-valued func-
tions for which the upper and lower integrals coincide and are simultaneously nite. The classical
Monotone and Dominated Convergence Theorems are derived and all is well with the world.
We follow a discussion of what the construction does in the all important case that the ini-
tial vector lattice ( = C(Q), Q a compact metric space.
It is noteworthy that this construction of Banach led him to a description of C(Q)

. It is
February 3, 2009
2. THE LEBESGUE INTEGRAL ON ABSTRACT SPACES 55
impossible to tell for certain (but easy to imagine) what Saks thought of Banachs construction of
C(Q)

. It appeared after all, as an appendix to Saks classic monograph Theory of the Integral,
yet uses practically none of the material from the monograph! Whatever Saks though, he soon
presented an elegant proof of Banachs result about C(Q)

in the Duke Mathematics Journal. We

present Saks proof in an appendix to this chapter.
2.1. A Start. The numbering and notation throughout this section is consistent with Ba-
nachs. Let ( denote a vector lattice of real-valued functions dened on a set Q. (i.e., If x, y (
then so is x y = infx, y and x y = supx, y.) A linear functional f on ( is a positive linear
functional if f(x) 0 for any x (, x 0.
Throughout this section, we will suppose f is a positive linear functional on ( satisfying
BC
_
if (x
n
) C, M C, with |x
n
| M and lim
n
x
n
(t) = 0 for all t Q, then
lim
n
f(x
n
) = 0.
If ( = C(Q), Q a compact metric space then any positive linear functional f on ( satises (BC)
thanks to Theorem 1.0.36. This is an example well worth keeping in mind.
1

If (x
n
) (, m (, z 0, x
n
m, and liminf
n
x
n
z then
lim
n
x
n
[x
n
[ = 0.
Now liminf
n
x
n
z ensures liminf
n
x
n
(t) 0 for each t Q, that is, for each t Q,
lim
n
infx
n
(t), x
n+1
(t), . . . 0.
So, given > 0 theres n = n() so that for all k n,
(2.1) x
k
(t) .
Look at x
n
[x
n
[:
(x
n
[x
n
[)(t) =
_
0, if x
n
(t) 0
2x
n
(t), if x
n
(t) < 0;
naturally, (x
n
[x
n
[)(t) 0 for all t. So, if lim
n
(x
n
[x
n
[) exists, it must be less than or equal to
0.
Lets check on contrary possibilities. Can it be that
liminf
n
(x
n
[x
n
[)(t
0
) < 0
for some t
0
Q? Lets suppose that this is possible. Notice since m (
liminf
n
(x
n
[x
n
[)(t
0
) inf2m(t
0
), 0 > .
If
liminf
n
(x
n
[x
n
[)(t
0
) < 0
its because theres a subsequence (x

n
) of x
n
so that for some
0
> 0
2x

n
(t
0
) = x

n
(t
0
) [x

n
(t
0
)[ <
0
,
for all n. It follows that
x

n
(t
0
) <

0
2
February 3, 2009
56 5. BANACH AND MEASURE
for all n. But if n is BIG we can arrange
x

n
(t
0
) >

0
4
(thats what our opening words (2.1) of wisdom guarantee). Ah ha! While
x

n
(t
0
) >

0
4
,
we also have
x

n
(t
0
) <

0
2
.
Drawing the conclusion that

0
2
<

0
4
is easy from this and leaves us with a clear-cut contradiction
to all that is right in our world. We conclude that
liminf
n
(x
n
[x
n
[) 0.
We know that
limsup
n
(x
n
[x
n
[) 0,
and so 1

follows.
2

If (x
n
) (, m (, x
n
m, and liminf
n
x
n
0 then
liminf
n
f(x
n
) 0.
Again start with a look at x
n
[x
n
[; as before since 2m (
< inf2m, 0 x
n
[x
n
[ 0.
By 1

, we know that
lim
n
(x
n
[x
n
[)(t) = 0
for each t Q, and since ( is a lattice, we can use the the (BC) condition on f to conclude
lim
n
f(x
n
[x
n
[) = 0.
Can liminf
n
f(x
n
) < 0? If so then its because there is a subsequence (x

n
) of (x
n
) and an
0
> 0 so
f(x

n
) <
0
for all n. We can (and do) assume that
lim
n
f(x

n
)
exists as well. But now
lim
n
f(x

n
), lim
n
f(x

n
[x

n
[)
both exist and so
lim
n
f([x

n
[)
exists, too, with
lim
n
f([x

n
[) = lim
n
(f([x

n
[ x

n
) +f(x

n
))
= lim
n
f(x

n
)
0
,
which is not possible, and so we have 2

.
February 3, 2009
2. THE LEBESGUE INTEGRAL ON ABSTRACT SPACES 57
2.2. Upper and lower integrals. Let L

denote the set of all real-valued functions z on Q

for which there exist two sequences (x
n
), (y
n
) ( such that
limsup
n
y
n
z liminf
n
x
n
.
L

is a linear space containing ( since ( is a vector lattice.

Given z L

, the upper integral of z,

_
(z), is dened by
_
(z) = infliminf
n
f(x
n
) : there exists m (, (x
n
) (, x
n
m, liminf
n
x
n
z;
the lower integral of z,
_
(z), is dened by
_
(z) = suplimsup
n
f(x
n
) : there exists M (, (x
n
) (, x
n
M, limsup
n
x
n
z.
Obviously
_
z =
_
(z).
Note: in each of the above denitions we can suppose that lim
n
f(x
n
) exists and is real valued
since liminfs and limsups are taken over sequences which are eventually nite. So we replace
liminf
n
f(x
n
) and limsup
n
f(x
n
) with lim
n
f(x
n
) throughout.
From the denitions we have
3

If z L

, z 0, and
_
(z) < P < then we can nd (x
n
) (, x
n
0, liminf x
n
z
with f(x
n
) < P for all n.
The value of 3

is found in the accessibility it aords us to epsilonics; since (with Banach) we

have frequent call on computing limsups and liminfs. This is a critical aid.
Lemma 2.2.1. For any x (,
_
(x) = f(x).
Proof. On the one hand, we can let x
n
= x for all n and m = x; this done, we plainly have
liminf
n
x
n
x, and x
n
m.
Hence
_
(x) liminf
n
f(x
n
) = f(x).
On the other hand, if (x
n
) ( and m ( with
liminf
n
x
n
x, and x
n
m,
then
liminf
n
(x
n
x) 0, and x
n
x mx.
2

steps in to say
0 liminf
n
f(x
n
x) = liminf
n
f(x
n
) f(x);
we see that
f(x) liminf
n
f(x
n
)
February 3, 2009
58 5. BANACH AND MEASURE
and with this we conclude
f(x)
_
(x).

Lemma 2.2.2. If z
1
, z
2
L

with
_
(z
1
),
_
(z
2
) < , then
_
(z
1
+z
2
)
_
(z
1
) +
_
(z
2
).
Proof. Suppose P
1
, P
2
are numbers such that
_
(z
1
) < P
1
, and
_
(z
2
) < P
2
.
There are sequences (x
(1)
n
), (x
(2)
n
) ( and functions m
1
, m
2
( such that
liminf
n
x
(1)
n
z
1
, x
(1)
n
m
1
for all n,
liminf
n
x
(2)
n
z
2
, x
(2)
n
m
2
for all n,
and
lim
n
f(x
(1)
n
) < P
1
, lim
n
f(x
(2)
n
) < P
2
.
Letting x
n
= x
(1)
n
+x
(2)
n
and m = m
1
+m
2
, we see that
liminf
n
x
n
z
1
+z
2
, x
n
m.
It follows that
_
(z
1
+z
2
) lim
n
f(x
n
) = lim
n
f(x
(1)
n
) + lim
n
f(x
(2)
n
) < P
1
+P
2
.
Enough said.
Lemma 2.2.3. For any z L

,
_
(z)
_
(z).
Proof. There is nothing to prove if
_
(z) = +. If
_
(z) = + then
_
(z) =
_
(z) so
again there is nothing to prove. If
_
(z),
_
(z) < then Lemma 2.2.2 kicks in to give
0 = f(0) =
_
(0) =
_
(z z)
_
(z) +
_
(z)
so that
_
(z) =
_
(z)
_
(z).

Lemma 2.2.4. If z L

and
_
(z) < , then
_ _
z +[z[
2
_
< ,
and
_
(z) =
_ _
z +[z[
2
_
+
_ _
z [z[
2
_
.
February 3, 2009
2. THE LEBESGUE INTEGRAL ON ABSTRACT SPACES 59
Proof. Suppose
_
(z) < P < . Find m ( and (x
n
) ( so that x
n
m for all n, and
liminf
n
x
n
z, lim
n
f(x
n
) < P. Notice that if x
n
m then
x
n
[x
n
[
2

m[m[
2
;
this can be seen by a simple analysis of cases.
Hence
f
_
x
n
+[x
n
[
2
_
= f
_
x
n

_
x
n
[x
n
[
2
__
= f(x
n
) f
_
x
n
[x
n
[
2
_
f(x
n
) f
_
m[m[
2
_
and
_ _
z +[z[
2
_
liminf
n
f
_
x
n
+[x
n
[
2
_
lim
n
f(x
n
) f
_
m+[m[
2
_
< .
Now
P > lim
n
f(x
n
)
liminf
n
f
_
x
n
+[x
n
[
2
_
+ liminf
n
f
_
x
n
[x
n
[
2
_

_ _
z +[z[
2
_
+
_ _
z [z[
2
_
;
it follows from Ps arbitrary nature among members >
_
(z) that
_
(z)
_ _
z +[z[
2
_
+
_ _
z [z[
2
_
.
Lemma 2.2.2 tells the rest of this tale.
Two more lemmas are plain and worth mentioning.
Lemma 2.2.5. If z
1
, z
2
L

satisfy z
1
< z
2
then
_
(z
1
)
_
(z
2
); in particular if z L

and
z 0 then
_
(z) 0.
Lemma 2.2.6. If z L

then
_
(z) =
_
(z) for any real number 0.
2.3. The Integral. Let L be the set
L = z L

:
_
(z) =
_
(z), with both nite.
Lemma 2.3.1. L is a linear space and
_
is a linear functional on L. Moreover ( L and
_
extends f.
February 3, 2009
60 5. BANACH AND MEASURE
Lemma 2.3.2. If z L then [z[ L, that is, L is a vector lattice.
Proof. Since
[z[ =
_
z +[z[
2
_
+
_
[z[ z
2
_
,
its enough (thanks to Ls linearity) to show that
z+|z|
2
,
z|z|
2
both belong to L if z L. We recall
that Lemma 2.2.4 ensures that if z L then
_
_
z+|z|
2
_
< and
_
_
z|z|
2
_
> , as well as
_
(z) =
_ _
z +[z[
2
_
+
_ _
z [z[
2
_
.
Symmetry (applying Lemma 2.2.4 to z and using
_
z =
_
z) shows that
_
_
z+|z|
2
_
< and
_
_
z|z|
2
_
> as well as
_
(z) =
_ _
z +[z[
2
_
+
_ _
z [z[
2
_
.
But z L says that
_
z =
_
z and so
__ _
z +[z[
2
_

_ _
z +[z[
2
__
+
__ _
z [z[
2
_

_ _
z [z[
2
__
=
__ _
z +[z[
2
_
+
_ _
z [z[
2
__

__ _
z +[z[
2
_
+
_ _
z [z[
2
__
=
_
z
_
z = 0.
Now Lemma 2.2.3 kicks in to say that the nite quantities
_
_
z+|z|
2
_
,
_
_
z+|z|
2
_
must, in fact, be
equal and the nite quantities
_
_
z|z|
2
_
,
_
_
z|z|
2
_
follow suit.
An old friend is next on the agenda.
Lemma 2.3.3 (Monotone Convergence Theorem).
MC
_
if (z
n
) L, z
n
z
n+1
for all n and z = lim
n
z
n
with
lim
n
_
(z
n
) < , then z L and
_
(z) = lim
n
_
(z
n
).
Proof. We can, and do, assume z
1
= 0; otherwise subtract z
1
from each of the z
n
s. Next note
that z z
n
for all n and so since
_
(z
n
) =
_
(z
n
) we have
(2.2)
_
(z) lim
n
_
(z
n
) = lim
n
_
(z
n
).
Let > 0. Let w
n
= z
n+1
z
n
0. For each n nd (w
(n)
k
) ( with w
(n)
k
0, lim
k
w
(n)
k
w
n
and
f(w
(n)
k
)
_
(w
n
) +

2
n
,
say. Write
y
n
= w
(1))
n
+ +w
(n)
n
;
February 3, 2009
2. THE LEBESGUE INTEGRAL ON ABSTRACT SPACES 61
so
y
1
= w
(1)
1
(
y
2
= w
(1)
2
+w
(2)
2
(
.
.
.
y
n
= w
(1)
n
+w
(2)
n
+ +w
(n)
n
(
Since y
n
= w
(1)
n
+w
(2)
n
+ +w
(n)
n
w
1
+w
2
+ w
n
,
liminf
n
y
n
liminf
n
(w
1
+w
2
+ +w
n
) =

n
w
n
= z.
At the same time,
f(y
n
) = f(w
(1)
n
) + +f(w
(n)
n
)

_
(w
1
) +

2
+
_
(w
2
) +

2
2
+ +
_
(w
n
) +

2
n
<
_
(w
1
) + +
_
(w
n
) +
=
_
(z
2
) +
_
(z
3
z
2
) + +
_
(z
n+1
z
n
) +
=
_
(z
2
+ (z
3
z
2
) + + (z
n+1
z
n
)) +
=
_
(z
n+1
) + lim
n
_
(z
n
) +. (by Lemma 2.2.5)
It follows that
_
(z) liminf
n
f(y
n
) lim
n
_
(z
n
) +,
and by s arbitrary nature,
0
_
(z) lim
n
_
(z
n
)

_
(z) by (2.2)

_
(z) < .
Therefore
_
(z) = lim
n
_
(z
n
).

Another old friend.

Lemma 2.3.4 (Dominated Convergence Theorem). Suppose (z
n
) L, M L and [z
n
[ M.
Then
g = liminf
n
z
n
, h = limsup
n
z
n
L
February 3, 2009
62 5. BANACH AND MEASURE
with
_
(g) liminf
n
_
(z
n
) limsup
n
_
(z
n
)
_
(h).
Consequently
DC
_
Suppose (z
n
) L, M L satisfy [z
n
[ M.
If z(t) = lim
n
z
n
(t) for each t Q then z L and
_
(z) = lim
n
_
(z
n
).
Proof. For each i and for each j i, write
g
ij
= minz
i
, z
i+1
, . . . , z
j
.
Then the sequence (g
ij
)

j=i
is decreasing, each member belongs to L and so the sequence (Mg
ij
)

j=i
is an increasing sequence of members of L. (MC) guarantees that if
g
i
= lim
j
g
ij
,
then
M g
i
L
and
_
(M g
i
) = lim
j
_
(M g
ij
);
that is, g
i
L and
_
(g
i
) = lim
j
_
(g
ij
).
Applying (MC) again reveals
g = liminf
n
z
n
= lim
j
g
ij
L
with
_
(g) = lim
i
_
(g
i
) liminf
n
_
(z
n
).

Lemma 2.3.5. If z L, z 0 and

_
z = 0 then whenever the function x satises [x[ z we
have that x L and
_
x = 0.
This is an immediate consequence of lemma 2.2.5.
2.4. We have the integral... Now lets turn to Banachs approach to integration. Start with
( = C(Q), Q a compact metric space. If f is a positive linear functional on ( then [f(x)[ is bounded
by f(1) so long as x B
C
. Why? Well if x B
C
then [x[ 1 and [x(q)[ 1 for all q Q. So
x B
C
means 1 x(q) 1; from this it follows that f(1) = f(1) f(x) f(1). Okay?
Any positive linear functional f on C(Q) is a bounded linear functional with norm f(1). By Ba-
nachs theorem (Theorem 1.0.36), if (x
n
) B
C
and [x
n
[ M ( with lim
n
x
n
(q) = 0 for all q Q,
then (x
n
) is weakly null in C(Q), hence lim
n
f(x
n
) = 0. This is (BC) in Banachs integration theory!
February 3, 2009
2. THE LEBESGUE INTEGRAL ON ABSTRACT SPACES 63
So we can extend f to a vector lattice L of real-valued functions dened on Q in a linear fash-
ion so that the extension, denoted by
_
df, enjoys the fruits of (MC) and (DC), that is
if (z
n
) L, (z
n
) an ascending sequence with z = lim
n
z
n
, z L whenever lim
n
_
z
n
df < ;
in this case,
_
z = lim
n
_
z
n
. (MC)
Suppose (z
n
) L and M L satisfy [z
n
[ M. Then liminf z
n
, limsup z
n
L and
_
liminf z
n
liminf
_
z
n
limsup
_
z
n

_
limsupz
n
. (DC)
So should (z
n
) L, / L satisfy [z
n
[ /, if z(q) = lim
n
z
n
(q) for each q Q then z L and
_
z = lim
n
_
z
n
.
What does this provide us with? To start, let F be a closed subset of Q. Consider the continuous
function
F
(q) = d(q; F); notice that q F precisely when
F
(q) = 0. If we consider
n
C(Q) to
be

n
(q) = inf
_

F
(q),
1
n
_
,
then lim
n

n
(q) =
F
(q) for each q Q. Moreover, [
n
[ 1 ( L; hence
F
L and
_

F
df = lim
n
_

n
df be either (MC) or (DC), take your pick.
Suppose E Q and
E
L. Then 1 =
E
+
E
c so
E
c L, too afterall 1 ( and
E
L. If
A, B Q and
A
,
B
L then since
AB
= inf
A
,
B
, we see that
AB
L. Because

A
+
B
=
AB

AB
we see that should
A
,
B
L then
AB
L as well. Therefore
E Q :
E
L
is an algebra of sets containing all the closed subsets of Q.
Suppose (E
n
) is an ascending sequence of subsets of Q such that
E
n
L for each n. Then

n
E
n
= lim
n

E
n
and
[
E
n
[ 1 ( L
so

n
E
n
L too - here you can appeal to (MC) or (DC), take your choice.
Therefore E Q :
E
L is a algebra of subsets of Q containing each and every closed
subset of Q. Therefore for any Borel set B Q,
B
L. Thus we have proved
Theorem 2.4.1. L contains all of the indicator functions on Borel sets.
Suppose (B
n
) is a sequence of pairwise disjoint Borel subsets of Q. Then
B
n
,

n
B
n
L and

n
B
n
=

B
n
;
an appeal to (MC) or (DC) soon reveals that
_

n
B
n
df =

n
_

n
df.
February 3, 2009
64 5. BANACH AND MEASURE
_
df acts in a countably additive fashion on Bo(Q)! It is a measure on Bo(Q). i.e.,

f
=
_
df, where
f
(E) =
_

E
df, whenever
E
L
is a countably additive measure dened on some eld that contains the Borel eld. The
total mass of
f
is f(1).
3. A Brief Intermission
While Banach did not base his derivation of C(Q)

on the material of Saks monograph, he

did nd it useful in his development of Haar measure for compact metric groups. In particular, he
called on metric outer measures as a guiding light for his passage to Haar measure.
What follows is a presentation of the basics of metric outer measures, enough to see us through
Banachs proof of the existence of an invariant (outer) measure and, afterwards, Saks elegant proof
that C(Q)

is what it is.
Preceding via the standard path, we suppose that we have a metric space with metric d
and a premeasure dened on the family T of subsets of . Let > 0 and denote by
T

= C T : diameter C .
Denote by

the restriction of to C

. The result is a premeasure that generates an outer measure

on by Method I, namely for E ,

:= inf
_

(E
n
) : E
n
T

, E
n
E
n
_
= inf
_

n
(E
n
) : E
n
T, E
n
E
n
, diam(E
n
)
_
.
Its plain that as gets smaller there are fewer members of T having diameter so

(E) gets
bigger. Hence
(E) := sup
>0

(E) = lim
0

(E),
exists and is well-dened.
Theorem 3.0.2. is a measure.
The only possible stumbling point to this is the countable subadditivity so lets see why
is countably subadditive. To this end, let (E
n
) be a sequence of subsets of and consider the
quantities
(
n
E
n
) and

n
(E
n
).
Obviously the latter exceeds the former if its

n
(E
n
) = so we may as well assume

n
(E
n
) <
.
Now for each > 0,

is a known outer measure so

(
n
E
n
)

(E
n
),
which in turn is

n
(E
n
).
February 3, 2009
3. A BRIEF INTERMISSION 65
The countable subadditivity follows from this.
A key ingredient to our mix is provided in the following.
Theorem 3.0.3. If A and B are non-empty subsets of that are positively separated then
(A B) = (A) +(B).
Here A and B positively separated means that there is a > 0 so that for any a A and
b B,
(a, b) .
Proof. We need to show that if A and B are positively separated then
(A B) (A) +(B)
where all the terms involved are nite. The idea of the proof is to cover A, B and A B with very
ne covers from the domain T of , a cover so ne that we can distinguish which members of the
cover touch A from those that touch B.
More precisely, suppose
0
> 0 is so small that
d(a, b)
0
for any a A and b B. Let > 0 announce its presence, <
0
/3. Let

> 0 be such that

< . Let = min

. Since
(A B) = sup
>0
inf

n
(C
n
),
where the inmum is taken over all sequences (C
n
) of members of T such that each C
n
has diameter
and A B
n
c
n
, we can choose (C
n
) from T in such a way that
A B
n
C
n
,
diam C
n
,
and

n
(C
n
) (A B) +.
By choosing C
n
s this way we see that a given C
n
can intersect A or B but not both.
Heres the rst punchline; if each C
i
has diameter then each has diameter

3
and so
a given C
i
can intersect A or B but not both. It follows that

C
i
A=
(C
i
) +

C
i
B=
(B
i
)

i
(C
i
) (A B) +.
But each C
i
that intersects A has diameter
1
so knowing, as we do, that A is a subset of
_
C
i
A=
C
i
,
we get

1
(A)

C
i
A=
(C
i
).
February 3, 2009
66 5. BANACH AND MEASURE
Similarly,

2
(B)

C
i
B=
(C
i
).
The result:

1
(A) +

2
(B)

C
i
A=
(C
i
) +

C
i
B=
(C
i
) (A B) +,
and so Method II leads us to believe
(A) +(B) (A B) +.
Since > 0 was arbitrary, this proof is done.
The property of an outer measure on a metric space that we have isolated above is important
enough to earn a special designation: an outer measure dened on the subsets of a metric space
(, d) is called a metric outer measure if whenever A and B are non-empty subsets of which are
positively separated then (AB) = (A) +(B). Their importance lies in the fact that these are
precisely the measures on a metric space for which every Borel set is measurable. Metric measures
enjoy some very strong continuity properties. Heres one of them.
Proposition 3.0.4. Let be a metric measure on a metric space . Suppose (A
n
) is an
increasing sequence of subsets of so that A
n
and A
c
n+1
are positively separated. Then
(
n
A
n
) = sup
n
(A
n
).
Proof. We have
A
1
A
2
A
n

with A
n
and A
c
n+1
positively separated for each n. We want to show that (
n
A
n
) sup
n
(A
n
)
and in this eort we may plainly suppose sup
n
(A
n
) < since otherwise all is okay.
First we look at the dierence sequence
D
1
= A
1
, D
2
= A
2
A
1
, D
3
= A
3
A
2
, . . . , D
n
= A
n
A
n1
, . . .
Of course, D
n
A
n
but further
D
2
A
c
1
, D
3
A
c
2
A
c
1
, D
4
A
c
3
A
c
2
A
c
1
, . . .
So
D
1
A
1
and D
3
, D
4
, . . . A
c
2
,
D
2
A
2
and D
4
, D
5
, . . . A
c
3
,
etc., etc., etc. In particular
D
1
A
1
, D
3
A
c
2
D
2
A
2
, D
4
A
c
3
D
1
D
3
A
3
, D
5
A
c
4
D
2
D
4
A
4
, D
6
A
c
5
.
.
.
D
1
D
3
D
2n1
A
2n1
, D
2n+1
A
c
2n
D
2
D
4
D
2n
A
2n
, D
2n
A
c
2n+1
.
February 3, 2009
3. A BRIEF INTERMISSION 67
It follows (inductively if you must know) that
(D
1
) +(D
3
) + +(D
2n1
) +(D
2n+1
) = (
n+1
k=1
D
2k1
),
and
(D
2
) +(D
4
) + +(D
2n
) +(D
2n+2
) = (
n+1
k=1
D
2k
).
Each of
n+1
k=1
D
2k1
and
n+1
k=1
D
2k
are subsets of A
2n+2
and so all above nd themselves
(A
2n+2
) sup
n
(A
n
) < .
Conclusion: both series

n
(D
2n1
) and

n
(D
2n
) converge. Now
(
n
A
n
) = (A
n
D
n+1
D
n+2
) Not sure about this one.
(A
n
) +(D
n+1
) +(D
n+2
) +
sup
n
(A
n
) +

k=n+1
(D
k
);
if > 0 be given then there is an n so that the latter sum is less than so
(
n
A
n
) sup
n
(A
n
) +.
Enough said.
A dividend paid by considering metric outer measures is found in the following.
Theorem 3.0.5. If is a metric measure on the metric space (, d) then every closed subset
of is measurable. Consequently every Boreal subset of is measurable.
Comment that this tells us that Borel sets are measurable.
Proof. Suppose F is a closed subset of the metric space , and let A F and B F
c
be
non-empty sets. For each n let
B
n
:=
_
x B : inf
yF
d(x, y) >
1
n
_
.
Notice that (B
n
) is an ascending sequence of subsets of B with B =
n
B
n
.
By their very denition of B
n
, each B
n
is positively separated from A. In fact, each B
n
is positively
separated from B
c
n+1
. Indeed if x B
n
and x

B
c
n+1
its because
inf
yF
d(x, y) >
1
n
, and inf
yF
d(x

, y)
1
n + 1
.
In the latter situation, it must be that for any > 0 there is a y

F so
d(x

, y

) <
1
n + 1
+;
of course d(x, y

) >
1
n
. It follows from the triangle inequality that
d(x, x

) d(x, y

d(x

, y

)
>
1
n

_
1
n + 1
+
_
=
1
n

1
n + 1
.
February 3, 2009
68 5. BANACH AND MEASURE
So if we choose
0
= (
1
n

1
n+1
)/2 we see that
d(x, x

) >
_
1
n

1
n + 1
_
/2.
This we can do for any x B
n
and x

B
c
n+1
.
Lets compute (A B). By our previous theorem
(A B) sup
n
(A B
n
)
= sup
n
(A) +(B
n
)
= (A) + sup
n
(B
n
)
= (A) +(
n
B
n
) by Proposition ??
= (A) +(B),
and it follows that F is measurable.
Since the collection of measurable sets which contain each closed set in is a elf, it must
be that each Borel set belongs. That Borel sets are measurable whenever is a metric outer
measure on is not accidental; its part and parcel of being a metric measure. Indeed if we suppose
is an outer measure on the metric space (, d) for which every closed subset of is measurable
and suppose that A and B are positively separated subsets of . Of course A B = so
(A B) = ((A B) A) +((A B) A
c
)
by the measurability of A. But
(A B) A = A and (A B) (A)
c
= B
by the positive separation of A and B. So
(A B) = ((A B) A) +((A B) (A)
c
= (A) +(B),
and is a metric measure.
4. Haar Measure
Let Q be a xed compact metric space.
Definition 4.0.6. We suppose that for subsets of Q the notion of congruence is dened to
satisfy the following conditions (here A

= B means A is congruent to B):

(i) A

= A.
(ii) A

= B B

= A.
(iii) A

= B, B

= C A

= C.
(iv) If A is an open set then so is any set congruent to A.
(v) If A is congruent to B and A can be covered by a sequence (A
n
) of open sets then B can
be covered by a sequence (B
n
) so that B
n

= A
n
for each n.
(vi) For any open set A the collection of sets congruent to A cover Q.
(vii) If (S
n
) is a sequence of open concentric balls with radii tending to zero, and if G
n

= S
n
and a
n
, b
n
G
n
with lim
n
a
n
and lim
n
b
n
existing then these limits coincide.
February 3, 2009
4. HAAR MEASURE 69
Example 4.0.7. If Q is a compact metrizable group with left invariant metric then A

= B
whenever B = xA for some x Q is a congruence.
Example 4.0.8. If Q is a compact metric space and G is a group of isometries of Q onto Q
that is transitive then A

= B if B = (A) for some G is a congruence.

Given two relatively compact open sets A, B, by Denition 4.0.6 (vi), the collection of sets
congruent to A covers B, a compact set. Hence there is a nite collection of sets congruent to A
that still cover B. This motivates Haars covering function h(B, A):
h(B, A) = the least number of sets congruent to A needed to cover B.
Proposition 4.0.9. Suppose A, B, and C are non-empty open subsets of Q. Then
(i) C B h(C, A) h(B, A).
(ii) h(B C, A) h(B, A) +h(C, A).
(iii) B

= C h(B, A) = h(C, A).
(iv) h(B, A) h(B, C)h(C, A).
(v) If d(A, B) = distance from A to B is positive (so

A

B = ) and (S
n
) is sequence of open
concentric balls with radii tending to zero, then there is a number N so that for n N
h(A B, S
n
) = h(A, S
n
) +h(B, S
n
).
Proof. (v) requires some serious and careful attention. Suppose (v) fails. Then there is (n
k
)
so that
h(A B, S
n
k
) < h(A, S
n
k
) +h(B, S
n
k
)
for each k. We can plainly suppose that the n
k
s are chosen so large that S
n
k
A ,= and S
n
k
B ,=
cannot both occur. It follows that there is a sequence (G
k
) such that G
k

= S
n
k
and G
k
A ,= and
G
k
B ,= . Why? Well x n
k
momentarily and imagine that any G that is congruent to S
n
k
could
meet at most one of A and B. If we cover A B by h(A B, S
n
k
) many sets congruent to S
n
k
,
then this cover (call it () would be the disjoint union of the collection / (respectively B) where /
(respectively B) consists of the members of ( that meet only / (respectively B). Consequently,
h(A B, S
n
k
) = [([ = [/[ +[B[ h(A, S
n
k
) +h(B, S
n
k
)
which is not an option. So we get a sequence (G
k
) of sets with G
k

= S
n
k
and G
k
A ,= , G
k
B ,= .
From each of the sets G
k
A pick a point a
k
and from each G
k
B pick a point b
k
. The sequences
(a
k
), (b
k
) lie inside relatively compact sets so there is a J P

(N) so that
a = lim
jJ
a
j
, b = lim
jJ
b
j
both exist. Of course, a

A, b

B. But now were in precisely the position to which (vii) of
denition 4.0.6 is applicable: a
k
, b
k
G
k
, G
k

= S
n
k
. Hence a = b. But A B = . OOPS! The
denial of (v) leads to unnecessary chaos.
Fix a non-empty open subset G of Q. Let (S
n
) be a sequence of open concentric balls with radii
tending to zero, each contained in G. For any open set A Q, dene
l
n
(A) =
h(A, S
n
)
h(G, S
n
)
.
Then
h(A, S
n
) h(A, G) h(G, S
n
),
February 3, 2009
70 5. BANACH AND MEASURE
and
h(G, S
n
) h(G, A) h(A, S
n
),
tell us that
1
h(G, A)
l
n
(A) h(A, G).
Therefore (l
n
(A)) is a bounded sequence of real numbers, each of whose terms exceeds the xed
positive number 1/h(G, A). Let LIM be a Banach limit, that is,
LIM B
l

and LIM satises

liminf x LIM(x) limsup x
for each and every x l

. Let
l(A) = LIM((l
n
(A))
for A G.
Proposition 4.0.10. If A and B are open sets then
(i) 0 < l(A) < , as long as A ,= .
(ii) A B l(A) l(B).
(iii) l(A B) l(A) +l(B).
(iv) A

= B l(A) = l(B).
(v) If d(A, B) > 0 then l(A B) = l(A) +l(B).
Proof. To see (v), note that by Proposition 4.0.9 (v), there exists an N such that for all
n N,
h(A B, S
n
) = h(A, S
n
) +h(B, S
n
).
From this we easily see that for all n N,
l
n
(A B) = l
n
(A) +l
n
(B),
and (v) follows.
Let X Q. Dene (X) as follows:
(X) = inf
_

n
l(A
n
) : X
_
n
A
n
, A
n
open
_
.
Here are the fundamental properties of .
Theorem 4.0.11. Let Q be a compact metric space. Then
(i) 0 (X).
(ii) If X is a non-empty open subset of Q then 0 < (X) < .
(iii) X Y Q (X) (Y ).
(iv) X
n
X
n
(X)

n
(X
n
).
(v) X

= Y (X) = (Y ).
(vi) d(X, Y ) > 0 (X Y ) = (X) +(Y ).
February 3, 2009
4. HAAR MEASURE 71
Proof. (i)-(iv) tells us that is a(n) (outer) measure; (v) assures us that respects congru-
ence and (vi) says that is a metric outer measure. It is a known consequence of s metric outer
measure character that every Borel set B Q is -measurable.
(i) deserves comment. If X is open then
(X) l(X) < .
If, in addition, X ,= then for any > 0 we can nd a sequence (A
n
) of open sets so that X
n
A
n
and
(X) +

n
l(A
n
).
Now if S is an open ball centered at a point in X and a subset of X, then only nitely many of the
A
n
s, say A
1
, A
2
, . . . A
N
are needed to cover S. Then
0 < l(S) l(S) l(A
1
. . . A
N
)

n
l(A
n
) < (X) +,
and 0 < (X) follows.
(v), too, deserves proof - after all, its (v) that ensures that is a metric outer measure. By
Proposition 4.0.10(iv), it suces to show that (X) +(Y ) (X Y ). If d(X, Y ) > 0 then there
are disjoint open sets U, V such that d(U, V ) > 0 with X U and Y V ; this is so thanks to
normality of metric spaces, if you please. Let > 0. Pick a sequence (A
n
) of relatively compact
open sets such that
X Y
_
n
A
n
,
and

n
l(A
n
) (X Y ) +.
Now each of the sets A
n
U, A
n
V are open and
d(A
n
U, A
n
V ) d(U, V ) > 0.
Hence by Proposition 4.0.10 (v)
l(A
n
U) (A
n
V )) = l(A
n
U) +l(A
n
V ) l(A
n
),
where the last inequality follows since A
n
U and A
n
V are disjoint open sets, whose union is a
subset of A
n
. Further
X
_
n
(A
n
U), Y
_
n
(A
n
V ).
So
(X)

n
l(A
n
U), (Y )

n
l(A
n
V ).
February 3, 2009
72 5. BANACH AND MEASURE
This in turns ensures that
(X) +(Y )

n
l(A
n
U) +

n
l(A
n
Y )

n
(l(A
n
U) +l(A
n
V ))

n
l(A
n
).
It follows that (X) +(Y ) (X Y ).
From this we know that is a metric outer measure on Q which assigns to any pair of congru-
ent subsets of Q the same value. Hence the collection of measurable sets is a eld of subsets
of Q which contains the Borel eld, and on this eld, is countably additive and assigns
congruent measurable sets the same measure.
5. Notes and Remarks
5.1. Saks Proof of C(K)

, K a Compact Metric Space. Soon after the appearance of

Saks monograph, Saks published an alternative proof of the theorem of Banach regarding positive
linear functionals on C(Q), Q a compact metric space. This proof relies on the theory of metric
outer measures to ensure that the resulting measure is a Borel measure.
We set our notation. Let Q be a compact metric space (with metric d), C(Q) is the Ba-
nach space of continuous real-valued functions dened on Q, and (to be consistent with Saks) is
a positive linear functional on C(Q). Recall that
Phi(x)[ (1)
whenever x B
C(Q)
, so is a member of C(Q)

with norm (1).

For any q Q, r > 0 denote by U
r
(q) and B
r
(q) the sets
U
r
(q) = y Q : d(q, y) < r, B
r
(q) = y Q : d(q, y) r.
Stage I For E Q, dene (E) by
(E) := inf(x) : x C(Q), x(q)
E
(x), for all x Q.
Heres whats so about :
(i) if A B then (A) (B);
(ii) if A, B Q then (A B) (A) +(B);
(iii) if d(A, B) > 0 (which is the same as A B = ) then (A B) = (A) +(B).
(iii) demands comment and proof, even. Let > 0. Pick x C(Q) so
x(q)
AB
(q)
for all q Q (where A, B Q satisfy d(A, B) > 0) and so that
(x) (A B) +.
February 3, 2009
5. NOTES AND REMARKS 73
Next chose h C(Q) so that 0 h(q) 1 for all q, with h(q) = 1 for q B, h(q) = 0 for q A.
Let x
A
= (1 h)x and x
B
= hx. Both x
A
, x
B
C(Q); also
x
A
(q) =
_
x(q) if q A
0 if q B

_

AB
(x) if q A
0 if q B
=
_
1 if q A
0 if q B
=
A
(q),
and
x
B
(q) =
_
0 if q A
x(q) if q B

B
(q)
for all x Q. It follows that
(A) +(B) (x
A
) + (x
B
)
= (x
A
+x
B
)
= (x)
(A B) +.
Stage II Let E Q and dene (E) by
(E) := inf

n
(G
n
) : G
n
is open, E
_
n
G
n
.
Then is an outer measure on Q, a metric outer measure, with the added property that for any
closed subset F of Q, (F) = (F). Well establish this last claim. Suppose G
n
is a sequence of
open sets so that
F
_
n
G
n
.
Since F is closed, Qs compactness is inherited by F, and so we can nd N so that
F G
1
G
N
.
It follows that
(F)
N

i=1
(G
i
)

n
(G
n
).
It follows that
(F) (F).
Now let > 0 be given. Pick x C(Q) so that x(q)
F
(q) for all q and (x) (F) + . Look
at the open set
G = q Q : x(q) > (1 +).
So F G since if q F then x(q)
F
(x) = 1. So
(F) (G) (Since G is open and F G)
(x) ( by denition of (G))
= (x)
((F) +).
Let 0. Then (F) (F) follows.
Stage III Like all metric outer measures, has among its measurable sets each and every Borel
February 3, 2009
74 5. BANACH AND MEASURE
subset of Q. In particular, each x C(Q) is measurable, and of course, bounded. Hence each
x C(Q) is integrable. Lets check (x) vis-a-vis
_
xd. Well show that
(x)
_
xd
for every x C(Q).
Take x C(Q). Let > 0. Realizing that (Q) = (1) we assume that x(q) > 0 for all
q Q - just add an appropriate constant to x and note that this has the exact same eect on
(x) and
_
xd, leaving their relationship unchanged. Choose > 0 so that if d(q, q

) then
[x(q) x(q

[ . Cover Q by open balls U

r
1
(q
1
), . . . U
r
n
(q
n
), of radii r
1
, . . . r
n
each < /2 centered
at q
1
, . . . q
n
respectively, with the added feature that
(y Q : d(y, q
i
) = r
i
) = 0.
This last feature can be assumed since (Q) < , and so for a xed q
0
Q only countably many
of the sets y Q : d(y, q
i
) = r can have positive measure. Now that the Us are in place, set
E
1
= U
r
1
(q
1
)
E
2
= U
r
2
(q
2
)E
1
E
3
= U
r
3
(q
3
)(E
1
E
2
)
.
.
.
E
n
= U
r
n
(q
n
)(E
1
E
n1
).
Each set E
1
, . . . E
n
is closed with diameter and
Q = E
1
E
n
.
Whats more, and this is crucial, the E
i
s overlap only in a set of measure 0! Let m
i
= minx(q) :
q E
i
, i = 1, 2 . . . n. Consider
a(q) =
n

i=1
m
i

E
i
(q);
notice that
x(q) a(q)
almost everywhere. Hence
_
xd
_
n

i=1
m
i

E
i
(q) d
=
n

i=1
m
i
(E
i
) =
n

i1
m
i
(E
i
).
For each k = 1, n, pick x
k
C(Q) so that for all q Q, x
k
(q)
E
k
(q) and
(E
k
) = (E
k
) (x
k
)

nm
k
.
Put
x = m
1
x
1
+ +m
k
x
k
+.
February 3, 2009
5. NOTES AND REMARKS 75
Since the oscillation of x on E
k
is no more than then
u(q) =
n

i=1
m
i
x
i
(x) + x(q).
It follows that
_
xd
n

i=1
m
i
(E
i
)

i=1
m
i
_
(x
i
)

nm
i
_
= (
n

i=1
m
i
x
i
)
= (u )
= (u) ()
(x) ((1) + 1).
Let 0 and be happy, dont worry; afterall,
_
xd (x)
for all x C(Q).
February 3, 2009
February 3, 2009
CHAPTER 6
The Arzela-Ascoli Theorem
77
February 3, 2009
February 3, 2009
CHAPTER 7
Von Neumanns Proof of the Existence and Uniqueness of an
Invariant Measure on a Compact Metric group
In this chapter, well show how to ascribe to each f C(G), a mean M(f), which is at one and
the same time, linear in f, non-negative when f is, and is a true average with the values at f and
any right translate of f, identical.
Let G be a compact metrizable topological group. Denote by T(G) the collection of non-empty
nite subsets of G and by C(G) the Banach space of all continuous real-valued functions dened on
G, equipped as usual with the supremum norm.
Throughout this section, if F
1
, F
2
T(G) then by F
1
F
2
, we mean all words a b, where a F
1
and b F
2
; in particular, if a
1
b
1
= a
2
b
2
but a
1
,= a
2
then we distinguish a
1
b
1
and a
2
b
2
.
Lemma 0.1.1. (i) If f C(G) then minf, max f, and Oscf = max f minf all exist.
(ii) If f C(G) and F T(G) then
OscRAve
F
f Oscf.
In fact,
minf minRAve
F
f max RAve
F
f max f.
(iii) If f C(G) and F
1
, F
2
T(G) then
RAve
F
1
RAve
F
2
f = RAve
F
1
F
2
f.
Proof. To see (ii), let F T(G) and f C(G). Dene
RAve
F
f (x) :=
1
[F[

aF
f (xa), x G.
Naturally RAve
F
f C(G).
79
February 3, 2009
80 7. VON NEUMANNS PROOF OF THE EXISTENCE AND UNIQUENESS OF AN INVARIANT MEASURE
To see (iii), if x G then
RAve
F
1
RAve
F
2
f(x) = RAve
F
1
1
[F
2
[

bF
2
f(xb)
=
1
[F
1
[

aF
1
S(xa)
[F
2
[
_
letting S(x) =

bF
2
f(xb)
_
=
1
[F
1
[ [F
2
[

aF
1
S(xa)
=
1
[F
1
F
2
[

aF
1

aF
2
f(xab)
=
1
[F
1
F
2
[

cF
1
F
2
f(xc)
= RAve
F
1
F
2
f(x).
Lemma 0.1.2. If f C(G) is not constant then there is an F T(G) such that
OscRAve
F
f < Oscf.
Proof. After all, fs not being constant ensures that there is an such that min f < <
max f. Set
U = [f < ] = x G : f(x) < .
Since minf < , U is a non-empty open set in G and G =

aG
Ua
1
. (If x G then for any
y U, x = y(y
1
x) U(y
1
x)

aG
Ua
1
.)
Now U is open (since f C(G)), and U ,= so Ua
1
is also a non-empty open set for each
a G. Therefore the Ua
1
s cover the compact G. There is F T(G) such that
G =
_
aF
Ua
1
.
Therefore for any x G there exists a
x
F such that x Ua
1
x
. i.e., for any x G there exists
a
x
F such that f(xa
x
) < . Thus
RAve
F
f(x) =
1
[F[

aF
f(xa)
=
1
[F[
_
_

aF,a=a
x
f(xa) +f(xa
x
)
_
_
<
1
[F[

aF,a=a
x
f(xa) +

([F[ 1) max f +
[F[
<
([F[ 1) max f + max f
[F[
= max f.
February 3, 2009
7. VON NEUMANNS PROOF OF THE EXISTENCE AND UNIQUENESS OF AN INVARIANT MEASURE 81
Therefore
OscRAve
F
f Oscf.

Lemma 0.1.3. Let f C(G) and dene / = RAve

F
f : F T(G). Then / is uniformly
bounded, equicontinuous family in C(G).
Proof. The key to this precious fact is that f is of course uniformly continuous. So given an
> 0 there is an open set V in G containing Gs identity such that if xy
1
V then [f(x)f(y)[ .
Notice that if a G and xy
1
V then (xa)(ya)
1
= xaa
1
y
1
= xy
1
V. So once xy
1
V ,
[f(xa) f(ya)[
for all a G. But now if F T(G) then whenever xy
1
V we have
[RAve
F
f(x) RAve
F
f(y)[ =
1
[F[

aF
f(xa)

aF
f(ya)

1
[F[

aF
[f(xa) f(ya)[

1
[F[
[F[ = .
Note that / is uniformly bounded since
[RAve
F
f(x)[ =
1
[F[

aF
f(xa)

1
[F[

aF
[f(xa)[

1
[F[
[F[ [[f[[ = [[f[[

.
We see that Lemma 0.1.3 takes on added signicance if we but recall the classical theory of Arzela
and Ascoli to the eect that / C(G) is relatively norm compact if and only if / is uniformly
bounded and equicontinuous.
With Lemmas 0.1.2 and 0.1.3 in hand, the plan of attack is clear. We want an averaging tech-
nique which will give a true average, assigning values in a uniformly distributed manner, If the
function f is constant then we will plainly want to assign that value of constancy to f. With the
aforementioned lemmas in hand, we handle non-constant functions thusly; if f is not constant, then
we can nd F
1
T(G) so that
OscRAve
F
1
< Oscf;
If RAve
F
1
(f) is constant then its value of constancy is the natural value to ascribe to f. If
RAve
F
1
(f) is not constant, then we appeal to Lemma 0.1.3 again to nd F
2
T(G) so that
RAve
F
2
RAve
F
1
OscRAve
F
1
(f).
Continuing in this vain, we see that in the worst case we can nd a sequence (F
n
) T(G) so that
for each n
OscRAve
F
n+1
RAve
F
n
(f) OscRAve
F
n
(f).
February 3, 2009
82 7. VON NEUMANNS PROOF OF THE EXISTENCE AND UNIQUENESS OF AN INVARIANT MEASURE
Appealing to Messrs. Arzela and Ascoli, we can pass to a sequence (F

n
) T(G) so (RAve
F

n
(f)) is
uniformly convergent.
The point is that because our averages were taken with respect to right translates, in the long
run, judicious choices of the F
n
s ought to produce an average that is right invariant. Remarkably
enough the wisdom needed has already been provided by Von Neumann.
Lemma 0.1.4. Let f C(G) and / = RAve
F
f : F T(G). Then
inf
gK
Oscg = 0.
Proof. Let
s = inf
gK
Oscg = infOscRAve
F
f : F T(G).
Therefore there exists (F
n
) in T(G) such that (RAve
F
n
) s. Thanks to Arzela and Ascoli we can,
by passing to subsequences if necessary, assume that
RAve
F
n
f g C(G),
uniformly. Its plain that on assuming the uniform convergence of (RAve
F
n
) that
minRAve
F
n
f ming and max RAve
F
n
f max g,
and so
OscRAve
F
n
f Oscg.
Thus Oscg = s. Heres the point: g is constant! Indeed if g were not constant there would be an
F
0
T(G) such that
s
0
= OscRAve
F
0
g < Oscg = s,
thanks to lemma 0.1.2. Since (RAve
F
n
f) is uniformly convergent, there exists N such that
[[RAve
F
N
f g[[

<
s s
0
3
.
i.e., for any x G,
[RAve
F
N
f(x) g(x)[
s s
0
3
.
But this is quickly seen to mean
[RAve
F
0
RAve
F
N
f(x) RAve
F
0
g(x)[
s s
0
3
,
for all x G as well. It follows that for all x G
[OscRAve
F
0
RAve
F
N
f OscRAve
F
0
g[ < 2
_
s s
0
3
_
.
i.e., for all x G,
[OscRAve
F
0
RAve
F
N
f(x) s
0
[ < 2
_
s s
0
3
_
.
But this in turn means that
OscRAve
F
0
RAve
F
N
f(x) < s
0
+ 2
_
s s
0
3
_
=
2
3
s +
1
3
s
0
< s.
But
OscRAve
F
0
RAve
F
N
f = OscRAve
F
0
F
N
f,
February 3, 2009
7. VON NEUMANNS PROOF OF THE EXISTENCE AND UNIQUENESS OF AN INVARIANT MEASURE 83
and
s = inf
FF(G)
OscRAve
F
f.
This should elicit an OOPS because
RAve
F
0
RAve
F
N
f = RAve
F
0
F
N
f /.
Therefore g is constant and s = 0. i.e.,
inf
gK
Oscg = 0.

We say the real number p is a right mean of f if for each > 0 there is an F T(G) such
that
[RAve
F
f(x) p[ <
for all x G. i.e.,
[[RAve
F
f p[[

< .
Theorem 0.1.5. Every f C(G) has a right mean.
Proof. By the techniques used in Lemma 0.1.4, there is a constant function h (say h(x) p)
and a sequence (F
n
) T(G) such that
lim
n
[[RAve
F
n
f h[[

= 0.
i.e.,
[[RAve
F
n
f p[[

0,
as n . Plainly p is a right mean of f.
Its plain that each f C(G) has a left mean as well, that is, there is a q R so that if
> 0 is given there exists an F T(G) so that

1
[F[

aF
f(ax) q

<
for all x G. For obvious reasons, we dene
LAve
F
f(x) =
1
[F[

aF
f(ax).
Theorem 0.1.6. Let f C(G). Let p be a right mean of f and q be a left mean of f. Then
p = q.
Proof. Let > 0. Find A, B T(G) so that
[[RAve
A
f p[[

2
, [[LAve
B
f q[[

2
.
February 3, 2009
84 7. VON NEUMANNS PROOF OF THE EXISTENCE AND UNIQUENESS OF AN INVARIANT MEASURE
Now
RAve
A
RAve
B
f(x) = RAve
A
1
[B[

bB
f(bx)
=
1
[A[
1
[B[

aA
S(xa) (where S(x) =

bB
f(bx))
=
1
[A[
1
[B[

aA

bB
f(bxa)
=
1
[B[
1
[A[

bB

aA
f(bxa)
=
1
[B[

bB
1
[A[

aA
f(bxa)
=
1
[B[

bB
RAve
A
f(bx)
= LAve
B
RAve
A
f.
Further,
RAve
A
(LAve
B
f q) = RAve
A
LAve
B
f q
and
LAve
B
(RAve
A
f p) = LAve
B
RAve
A
f p.
So for any x G,
[p q[ = [p RAve
A
LAve
B
f(x) + RAve
A
LAve
B
f(x) q[
[p RAve
A
LAve
B
f(x)[ +[RAve
A
LAve
B
f(x) q[
= [p LAve
B
RAve
A
f(x)[ +[RAve
A
LAve
B
f(x) q[
= [LAve
B
(p RAve
A
f(x))[ +[RAve
A
(LAve
B
f(x) q)[
[p RAve
A
f(x)[ +[LAve
B
f(x) q[ (since [LAve
B
f[ [f[ and RAve
B
f[ [f[)
<

2
+

2
= ,
and p = q. Go gure.
Corollary 0.1.7. For any f C(G) there is a unique number M(f) that is both a right and
left mean.
Theorem 0.1.8. The functional M on C(G)satises the following
(i) M is linear.
(ii) Mf 0 if f 0.
(iii) M(1) = 1.
(iv) M(
a
f) = M(f) = M(f
a
) for each a G, where
a
f(x) = f(ax) and f
a
(x) = f(xa).
(v) M(f) > 0 if f 0 but f ,= 0.
(vi) M(

f) = M(f) where

f(x) = f(x
1
) for each x G.
Proof. We start by showing
(0.1) M(RAve
F
f) = M(f)
February 3, 2009
7. VON NEUMANNS PROOF OF THE EXISTENCE AND UNIQUENESS OF AN INVARIANT MEASURE 85
for each f C(G) and each F T(G). Suppose that M(f) = p. If > 0 is given to us then we can
nd F
0
T(G) such that
[[LAveF
0
f p[[

.
i.e.,

1
[F
0
[

bF
0
f(bx) p

for all x G. It follows that for any x G and a F,
[RAve
F
LAve
F
0
f(x) p[ .
Since
RAve
F
LAve
F
0
f = LAve
F
0
RAve
F
f,
p is a left mean of RAve
F
f. Hence by our previous result,
M(RAve
F
f) = p,
and
M(RAve
F
f) = M(f).
To see that M is linear, let M(f) = p and M(h) = q. Pick H
0
T(F) so that
[[RAve
H
0
h q[[

.
i.e., for all x G,

1
[H
0
[

bH
0
h(xb) q

.
i.e., if E T(G) and x G then
[RAve
EH
0
h(x) q[ = [RAve
E
RAve
H
0
h(x) q[ < .
Now
p = M(f) = M(RAve
F
f)
for any F T. Therefore p is the right mean of RAve
H
0
f. Hence there exists F
0
T(G) so that
[[RAve
F
0
RAve
H
0
f p[[ .
i.e., for all x G,
[RAve
F
0
H
0
f(x) p[ = [RAve
F
0
RAve
H
0
f(x) p[ .
Since we already know that for all x G and each E T(G),
[RAve
EH
0
h(x) q[ ,
it follows that by taking E = F
0
we get for each x G,
[RAve
F
0
H
0
(f +h)(x) (p +q)[ 2.
Thus
M(f +h) = M(f) +M(h).
It follows from this and the easily established fact that M(kf) = kM(f) that M is linear, and we
have shown (i).
Parts (ii) and (iii) are clear. To see (iv), since
(0.2) RAve
F
f(xa) = RAve
aF
f(x),
February 3, 2009
86 7. VON NEUMANNS PROOF OF THE EXISTENCE AND UNIQUENESS OF AN INVARIANT MEASURE
M(f
a
) = M(RAve
F
f
a
(x)) (by (0.1))
= M(RAve
F
f(xa))
= M(RAve
aF
f(x)) (by (0.2))
= M(f) (by (0.1)).
Similarly,
(0.3) LAve
F
f(ax) = LAve
Fa
f(x),
and so
M(
a
f) = M(LAve
F
(
a
f(x))) (by (0.1) actually its equivalent with left averages)
= M(LAve
F
f(ax))
= M(LAve
Fa
f(x)) (by (0.3))
= M(f) (by (0.1) actually its equivalent with left averages),
and thus
M(
a
f) = M(f) = M(f
a
).
For (v), suppose that f C(G), f 0, f , 0. Then there is > 0 such that U = [f > ] is non-
empty and open; its easy to see that U
a
1 : a G is an open cover of the compact G. (If x G
then for any y G, x = y(y
1
x) U(y
1
x)

aG
Ua
1
.) It follows that for some a
1
, . . . , a
m
G
G = Ua
1
1
_
Ua
1
2
_
. . .
_
Ua
1
m
.
Lets check to see how this plays out.
If x G then x Ua
1
k
for some 1 k m. Hence, xa
k
U and thus f(xa
k
) > . It fol-
lows that
RAve
{a
1
,...,a
m
}
f(x) =
1
m
m

i=1
f(xa
i
) >

m
,
for all x G. Therefore
0 <

m
M(RAve
{a
1
,...a
m
}
f) = M(f).
Almost done; we have but to show that M(f) and M(

f) agree. To establish this, dene
N(f) = M(f inv),
where inv : G G is given by inv(x) = x
1
. N is a linear functional on C(G), N(f) 0 if f 0,
and N(1) = 1. Moreover
N(
a
f) = M(f
a
inv)
= M(
a
1

f) (since f
a
inv(x) = f
a
(x
1
) =

f(a
1
x) =
a
1 f(x))
= M(

f) (by (iv))
= N(f).
But by Corollary 0.1.7, there is only one invariant mean on C(G) so N(f) = M(f).
February 3, 2009
CHAPTER 8
The Fubini-Tonelli Theorem
1. Kakutanis Proof of the Uniqueness of Haar Measure
Let G be a compact topological group. We view G as a group of homeomorphisms of G onto
itself. A Borel measure is left Ginvariant if for any Borel set E G any x G,
(xE) = (E).
A Borel measure is right Ginvariant if for any Borel set E G any x G,
(Ex) = (E).
A Borel set E G is left Ginvariant where is any Borel measure on G if for each x G
(ExE) = 0.
A Borel set E G is right Ginvariant where is any Borel measure on G if for each x G
(EEx) = 0.
We say that G is left ergodic if given a left Ginvariant countably additive non-negative Borel
measure then any Borel set E G that is left Ginvariant satises either
(E) = 0 or (E
c
) = 0.
Similarly, G is right ergodic is dened analogously.
Well rst show that if G is left ergodic then the left invariant measure on G is unique. Then
we will show that G is left ergodic.
Assume then that G is left ergodic but
1
and
2
are both left invariant measures on G for which
there are Borel sets E
1
, E
2
G and a real such that

1
(E
1
) <
2
(E
1
), and
1
(E
2
) >
2
(E
2
)
(so that
1
,
2
are not constant multiples of each other). Look at
:= (E) =
1
(E)
2
(E),
a left invariant countably additive Borel measure on G.
The Hahn Decomposition Theorem splits G into the disjoint union
G = P N
of Borel sets P, N in such a way that if E is a Borel subset of P then (E) 0 and if E is a Borel
subset of N then (E) 0. Moreover, P and N are essentially unique in this regard.
87
February 3, 2009
88 8. THE FUBINI-TONELLI THEOREM
Now is left invariant so P is left invariant and so is N. Why is this so? Let E be a Borel
subset of P, and let y G. Then for any x G,
(xE) = (E) 0.
But if F is a Borel subset of yP then y
1
F is a Borel subset of P = y
1
yP and so
(F) = (y
1
F) 0.
It follows that for any Borel subset F of yP, (F) 0 and yP is also a positive set for . Thus
(P) (yP) 0.
Since P is essentially unique as a positive set for , it follows that
(P) = (yP).
In a similar fashion we see that N is left invariant.
Look at the countably additive, non-negative, left invariant measure [[, the variation of ,
[[(E) = (E P) +(E N).
Ergodicity of G says
[[(P) = 0 or [[(N) = 0,
since each of P and N is G[[invariant. But (E
1
) > 0 and (E
2
) < 0. Therefore
[[(P) (E
1
P) (E
1
) > 0, and [[(N) (E
2
N) (E
2
) > 0,
Oops! The uniqueness of left-invariant measures on G follows.
We now show that G is left ergodic. Let be a left invariant countably additive non-negative
(real-valued) Borel measure on G and let E be a left invariant Borel subset of G.
Look at on GG where is an arbitrary but xed right invariant regular Borel measure on
G. Consider
E
, a Borel function on G to be sure. Set
(x, y) =
E
(yx).
Since
E
is a Borel measurable function is a Borel measurable function on GG. For any y G,
[
E
(x) (x, y)[ =
Ey
1
E
(x).
But our assumption on E is that E is left invariant; hence
(EyE) = 0
for each y G; in particular,
_
[
E
(x) (x, y)[ d(x) =
_

Ey
1
E
(x) d(x) = (Ey
1
E) = 0,
for each y G. It follows that
_
G
_
G
[
E
(x) (x, y)[ d(x) d(y) = 0.
Monsieur Fubini steps in to say that for almost all x G, we have
_
[
E
(x) (x, y)[ d(y) = 0.
February 3, 2009
1. KAKUTANIS PROOF OF THE UNIQUENESS OF HAAR MEASURE 89
He further stipulates that there exists M (G, ) M a Borel set, (M) = 0 so if x , M there exists
C
x
(G, ), C
x
a Borel set, (C
x
) = 0 so for any y , C
x
we have
E
(x) =
E
(yx). To see this, take
x , M. Then y , C
x
means yx , C
x
x; on letting z = yx we have for z , C
x
x that
E
(x) =
E
(z).
Now look at x

M. Then there exists C

x
(G, ), C
x
a Borel set, (C
x
) = 0 so if y , C
x
then

E
(x

) =
E
(yx

). Again, y , C
x
is the same as yx

C
x
x

so if z = yx

we have z , C
x
x

,
implying
E
(x

) =
E
(z).
Since
(C
x
x) = (C
x
) = 0 = (C
x
x

) = (C
x
x

),
C
x
x, C
x
x

both have the same measure as C

x
and C
x
, respectively; that is,
(C
x
x) = 0 = (C
x
x

).
So there exists z G((C
x
x) (C
x
x

)). For such a z,

E
(x) =
E
(z) =
E
(x

).
In other words
E
is constant on M
c
. Check out the possibilities:

E
= 0 or
E
= 1.
Either way is okay!
February 3, 2009
February 3, 2009
CHAPTER 9
Homogeneous Spaces
Let G be a compact topological group and K be a compact Hausdor space. We say that G
acts transitively on K if there is a continuous map GKK : (g, k) g(k) such that
(i) e(k) = k for all k K (e is the identity of G);
(ii) (g
1
g
2
)(k) = g
1
(g
2
(k)) for all g
1
, g
2
G, h K;
(iii) given k
1
, k
2
K there is a g G so that g(k
1
) = k
2
.
It is noteworthy that each g G may be viewed as a homeomorphism of K onto itself; after all, the
map kg(k) is continuous and has kg
1
(k) as an inverse.
Condition (iii) says, in particular, that the space K is homogeneous;i.e., we can move points of
K around K via homeomorphisms (members of G, in fact) of K onto itself.
If is the unique translation invariant Borel probability on G then induces a Ginvariant Borel
probability on K. This is an important construction, one worth understanding in general as well as
in special cases. WE HAVE NOT USED THE WORD PROBABILITY ANYWHERE
BEFORE THIS PARAGRAPH.
Suppose H is a closed subgroup of the compact topological group G. Consider the set G/H
with the so-called quotient topology, that is, the strongest topology that makes the natural map
q
H
: GG/H (taking g G to gH G/H) continuous; so U G/H is open precisely when q

H
(U)
is open in G. In other words, a typical open set in G/H is of the form xH : x V when V is open
in G. Because H is supposed to be closed, this topology is Hausdor; because q
H
is continuous and
surjective, G/H is compact.
More is so. G acts transitively on G/H. The map (g, g

H)gg

H ts the bill in the denition.

In fact, any transitive action of a compact group on a compact space is of the sort just described.
To be sure we need to tell when seemingly dierent spaces are the same under Gs action. Let
G act transitively on each of the compact Hausdor spaces K
1
, K
2
. We say that K
1
and K
2
are
isomorphic under Gs action if there is a homeomorphism : K
1
K
2
such that
(g(k
1
)) = g((k
1
))
for each k
1
K
1
.
Theorem 0.0.9 (Weil). Let the compact group G act transitively on the compact Hausdor
space K. Then there is a closed subgroup H of G such that K and G/H are isomorphic under Gs
action.
91
February 3, 2009
92 9. HOMOGENEOUS SPACES
Proof. Fix k
0
K. Look at
H = g G : g(k
0
) = k
0
.
H is called the isotopy subgroup. It is plain that H is a closed subgroup of G. A natural candidate
for the isomorphism of G/H and K is at hand: : G/HK given by
(gH) = g(k
0
).
For g
1
, g
2
G, g
1
(k
0
) = g
2
(k
0
) precisely when
g
1
1
(g
2
(k
0
)) = g
1
1
(g
1
(k
0
)) = e(k
0
) = k
0
,
or g
1
1
g
2
H, which is tantamount to g
1
H = g
2
H. This assures us that is well-dened and
injective.
The transitivity of Gs action ensures s surjectivity. To see that is also continuous, x g G
and let V be an open set in K containing g(k
0
). By the continuity of the map
(g, k)g(k)
on GK, there is an open set U in G which contains g so that u(k
0
) V for all u U. But q
H
(U) is
open in G/Hs quotient topology and q
H
(U)

(V ). This shows that is a continuous bijection

between the compact Hausdor spaces G/H and K; as such is a homeomorphism.
Further if g
1
, g
2
G then
g
1
((g
2
H)) = g
1
(g
2
(k
0
)) = (g
1
g
2
)(k
0
) = (g
1
(g
2
H)).
Thus G/H and K are isomorphic under Gs action.
Note that because of this isomorphism theorem we can consider any G/H where H is the iso-
topy subgroup associated with any k
0
K.
Now were ready for the main course.
Theorem 0.0.10 (Weil). Suppose the compact group G acts transitively on the compact Haus-
dor space K. Then there is a unique Ginvariant regular Borel probability measure on K.
Proof. We identify K with the isotopy subgroup G/H as in our previous theorem. Let
q
H
: GG/H
be the natural quotient map. Suppose is the normalized Haar measure on G and dene
G/H
on
G/H by

G/H
(B) = (q

H
(B))
for any Borel set B G/H.
If g G and B is Borel subset of G/H then
g(q

H
(B)) = gx : xH B
= gx : gxH gB
= q

H
(gB).
February 3, 2009
9. HOMOGENEOUS SPACES 93
Therefore

G/H
(gB) = (q

H
(gB))
= (g(q

H
(B))
= (q

H
(B)) =
G/H
(B),
and
G/H
is Ginvariant.
Uniqueness is a touchier issue, as is always the case it seems. We take a close look at how members
of rca(B
0
(G)) act on C(G). Take C(G) and g G. DEFINE rca SOMEWHERE? Dene

g
C(G) by

g
(x) = (gx).
Denote by
H
the Haar measure (normalized so
H
= 1) on H. The map GC(G) that takes g to

g
is uniformly continuous (this is an easy modication of Theorem 3.0.9) so that

(g) =
_
H

g
(h)d
H
(h), g G
denes a member

of C(G).
Suppose g
1
H = g
2
H. Then g
1
1
g
2
H,

(g
1
) =
_
H

g
1
(h) d
H
(h)
=
_
H

g
1
(g
1
1
g
2
h) d
H
(h) (by
H
s invariance and g
1
1
g
2
H)
=
_
H
(g
2
H) d
H
(h)
=
_
H

g
2
(h) d
H
(h) =

(g
2
).
Therefore

is constant on the left cosets of H so we can lift

to a continuous function

on G/H:

(gH) =

(g).
To summarize: if C(G) then we dene

C(G) and from this we get

C(G/H). Remarkably,
each member of C(G/H) comes about from this procedure. In fact, if f C(G/H) then f q
H

C(G) and for any g G

(f q
H
)(gH) =

(f q
H
)(g)
=
_
H
(f q
H
)
g
(h) d
H
(h)
=
_
H
(f q
H
)(gh)d
H
(h)
=
_
H
f(ghH)d
H
(h)
=
_
H
f(gH)d
H
(h)
= f(gH)
H
(H) = f(gH).
February 3, 2009
94 9. HOMOGENEOUS SPACES
In other words, f =

(f q
H
).
Now we look at Gs action. Take any Ginvariant regular Borel probability measure on G/H.
For C(G) dene
() =
_
G/H

(gH) d(gH).
Then is a probability measure in C(G)

. Moreover, is translation invariant. Indeed if x G

(
x
) =
_
G/H

x
(gH) d(gH)
=
_
G/H

x
(g) d(gH)
=
_
G/H
(xg) d(gH)
=
_
G/H

(xgH) d(gH)
=
_
G/H

(gH) d(gH) = ().

So is nothing else but normalized Haar measure on G. WE SHOULD TALK ABOUT WHAT
IT MEANS TO BE TRANSLATION INVARIANT IN TERMS OF MEMBERS OF
C(K).
If
1
and
2
are Ginvariant regular Borel probabilities on G/H and if x =

(x q
H
) C(G/H)
then

1
(x) = (x q
H
) =
2
(x);
in other words,
1
and
2
are the same. WHY?
The worth of an abstract construction lies, at least in part, it its applicability to concrete cases. Our
rst application is classical and was well-known before Weils general theorem. It is, nonetheless,
interesting and important.
Our setting: O(n), the orthogonal group of order n is our compact group; S
n1
, the unit sphere in
R
n
is our compact Hausdor space. The action of O(n) on S
n1
is given, naturally by
(u, x)u(x).
It is easy to verify that O(n) acts on S
n1
in a suitable fashion! Transitivity follows by letting
x, x

S
n1
; choose orthonormal bases x
1
, x
2
, . . . , x
n
and x

1
, x

2
, . . . , x

n
for R
n
, and let u : R
n
R
n
be the member of O(n) that takes x to x

, x
j
to x

j
for j = 2, . . . , n.
Acknowledging the descriptions of members of O(n) as rotations of R
n
, a direct application of
Weils theorem says: There is a unique rotation-invariant regular Borel probability measure on S
n1
.
Geometry is replete with examples of compact Hausdor spaces that are homogeneous spaces on
which various compact groups act transitively.
February 3, 2009
9. HOMOGENEOUS SPACES 95
Here are a few more.
Again our group is O(n). This time our underlying compact space is

(n) = (x, y) S
n1
S
n1
: x y,
where x y means x is perpendicular to y. Note that (x, y)

(n) precisely when for any real
valued numbers a, b:
[[ax +by[[
2
= a
2
+b
2
.
It is easy to see from this that

(n) is a closed subset of S
n1
S
n1
, hence, is compact. The
action of O(n) is natural enough, too: (u, (x, y))(ux, uy). It is quick and reasonably painless to
see that O(n) acts transitively on

(n).
One more. Let 1 m n. Denote by

(m)
(n) the set
(m)

(n) =
_
_
_
(x
1
, . . . , x
m
) S
n1
S
n1
. .
m times
: x
1
, . . . , x
m
is orthonormal
_
_
_
.
Note that (x
1
, . . . , x
m
)

(m)
(n) precisely when regardless of the real numbers a
1
, . . . a
m
, we have

j=1
a
j
x
j

2
=
m

j=1
a
2
j
.
This in mind,

m
(n) is a compact set of (S
n1
)
m
is easy to see; moreover, the action of O(n) on

(m)
(n) is given by
(u, (x
1
, . . . , x
m
))(ux
1
, ux
2
, . . . ux
m
)
is a transitive one, establishing, with a modicum of tender love and care, that O(n) acts transitively
on

(m)
(n).
Next let (
m
(n) denote the mdimensional Grassmanian manifold, that is, (
m
(n) is the space
of all mdimensional linear subspaces of R
n
. There is a natural surjection of

(m)
(n) onto (
m
(n)
that takes (x
1
, . . . , x
m
)

(m)
(n) to the linear span of x
1
, . . . , x
m
(
m
(n). If we equip (
m
(n)
with the natural quotient topology the result is a compact Hausdor space. Clearly O(n) acts
transitively on (
m
(n). The map reecting the action of O(n) on (
m
is plain: if x
1
, . . . , x
m
is an
orthonormal set in R
n
then
(u, spanx
1
, . . . , x
m
) = spanux
1
, . . . , ux
m
.
Here we interject that the geometry imparted above on (
m
(n) is such that if E = spanx
1
, . . . , x
m

and E

= spanx

1
, . . . , x

m
are members of (
m
(n) and if each x
k
is close to x

k
in R
n
then E is
close to E

in (
m
(n).
In this way we nd that there is a unique rotation invariant probability Borel measure on the
ndimensional Grassmanian manifold (
m
(n).
February 3, 2009
February 3, 2009
CHAPTER 10
Metrics in Compact Groups
Let G be a locally compact metrizable group with left invariant metric. Then G has a neigh-
borhood basis of open balls with compact closure. Can G have a metric in which all of the balls
have compact closure? If so then G must be second countable; afterall G =
n
B
n
where the B
n
s
are centered at a xed point of G and have a radius n. Since each B
n
has B
n
compact, and since
G is separable and metrizable, its second countable.
Heres a theorem of R. Struble. [?]
Theorem 0.0.11. A locally compact group metrizable topological group has a left invariant
metric (that generates its topology) in which all its open balls have compact closure if and only if G
is second countable.
Though the rst Lemmas content is a consequence of the Birkho-Kakutani thereom, the proof
rendered here (and due to Struble) is too clever and enlightening not to be included.
Lemma 0.0.12. Let G be a locally compact group with left Haar measure. Let (V
n
) be a decreasing
sequence of open sets that form a neighborhood basis of the identity e in G where V
n
compact for
each n. Then
(x, y) = sup
n
(xV
n
yV
n
)
denes a left invariant metric on G which is compatible with the topology of G.
Proof. Its clear that (x, y) is well-dened and that (x, y) = (y, x). Moreover (x, y) 0
and (x, y) < regardless of x, y G since each V
n
is a Borel set with compact closure. Further,
(zx, zy) and (x, y) coincide because is left invariant.
If x ,= y then since G is Hausdor there must be an m so that xV
m
yV
m
= ; but now
(x, y) (xV
m
yV
m
) = 2(V
m
) > 0.
On noticing that for any n and any x, y, z G,
xV
m
yV
n
(xV
n
zV
n
) (zV
n
yV
n
),
we see that for any x, y, z G,
(xV
n
yV
n
) ((xV
n
zV
n
) (zV
n
yV
n
))
(xV
n
zV
n
) +(zV
n
yV
n
)
(x, z) +(z, y),
and with this
(x, y) (x, z) +(z, y).
97
February 3, 2009
98 10. METRICS IN COMPACT GROUPS
In sum, is a left invariant metric on G. If Gs topology is discrete then (e) > 0 so V
m
= e
for some m; hence if x ,= y,
(x, y) (xV
m
yV
m
) = (x, y) = 2(e) > 0,
and the topoology induced by is discrete. If Gs topology is not discrete then (V
n
) (
n
V
n
) =
(e) = 0. If V is any open set containing e then there is an m N so that V
m
V
1
m
V.
Claim 1: x V whenever (x, e) < (V
m
). To see this, let (x, e) < (V
m
). Then
(xV
m
V
m
) (x, e) < (V
m
),
a positive number. Were xV
m
V
m
= then
(xV
m
V
m
) = 2(V
m
) < (V
m
),
oops! So xV
m
V
m
,= and thus there are v
1
, v
2
V
m
so xv
1
= v
2
xV
m
V
m
and
x = v
2
v
1
2
V
m
V
1
m
V.
This is so whenever (x, e) < (V
m
), and our claim is justied.
Lets look at all of the points x such that (x, e) < r, where r Q, r > 0. There must be an
m N so that (V
n
) <
r
4
, whenever n m. Each of the functions
f
k
(x) = (xV
k
V
k
)
is continuous and satises f
k
(e) = (V
k
V
k
) = () = 0. But now we know there is an l N so
that if x V
l
then f
1
(x), . . . , f
m1
(x) < r.
Claim 2: if x V
l
then (x, e) < r. Why is this so? Well if x V
l
then by choice of l N,
we have
(xV
1
V
1
), . . . , (xV
m1
V
m1
) < r.
What about (xV
k
V
k
) for k m? In this case,
(xV
k
V
k
) 2(V
k
) < 2
r
4
< r.
It follows that (x, e) = sup
n
(xV
m
V
n
) < r.
Our two claims taken in tandem show that generates Gs topology about e. Since is left invariant
and since Gs topology is too this is enough to say that generates Gs topology everywhere.
Lemma 0.0.13. Let G be a locally compact, second countable (hence metrizable, separable) group.
Then there exists a family U
r
: r N, r > 0 such that
(i) for each r, each U
r
is open and U
r
is compact,
(ii) U
r
= U
1
r
(iii) U
r
U
s
U
r+s
(so if r < s then U
r
U
r
U
rs
U
s
),
(iv) U
r
: r > 0 is a base for the open sets about e,
(v)
r>0
U
r
= G.
February 3, 2009
10. METRICS IN COMPACT GROUPS 99
Once Lemma 0.0.13 is established, were ready for business. Indeed, let U
r
: r > 0 be the family
of open sets about e generated from Lemma 0.0.13. For x, y G, set
d(x, y) = infr : y
1
x U
r
.
Since G =
r>0
U
r
, for an pair x, y G, we have y
1
x U
r
for some r > 0. It follows
that d(x, y) 0.
e U
r
for each r > 0 so d(x, x) = 0.
If y
1
x ,= e then there is an r
0
> 0 so that y
1
x , U
r
0
(Part (iv) of Lemma 0.0.13) tells
us this). But whenever 0 < r
0
< r we have (by Part (iii) of Lemma 0.0.13)
U
r
0
U
r
0
U
rr
0
U
r
,
so d(x, y) r
0
> 0.
U
r
= U
1
r
so y
1
x U
r
precisely when x
1
y U
r
; consequently, d(x, y) = d(y, x).
Suppose x, y, z G with y
1
x U
r
, z
1
y U
s
. Then
z
1
x = z
1
yy
1
x U
s
U
r
U
r+s
,
so d(x, z) r +s. This is so whenever y
1
x U
r
so d(x, z) d(x, y) +s; again this is so
whenever z
1
y U
s
so d(x, z) d(x, y) +d(y, z).
Finally, if x, y, z G then
d(zx, zy) = infr : (zy)
1
zx U
r
= infr : y
1
x U
r
= d(x, y).
To summarize: d is a left invariant metric on G.
Since d(x, e) < r means x = e
1
x U
r
, the open d-ball of radius r centered at e is contained
in U
r
. Also this same d-ball contains U
r
for any 0 < r

< r since if x U
r
then
e
1
x = x U
r
U
r
U
rr

2
U
r+r

2
,
and so d(x, e)
r+r

2
< r. Therefore if 0 < r

< r then
U
r
x : d(x, e) < r U
r
.
Thus by Part (iv) of Lemma 0.0.13 the metric d is compatible with the topology of G and by Part
(i) of Lemma 0.0.13, all d-balls are bounded. Joe: you wrote the following instead but I had troubles
reading your handwriting: Thus the open d-balls of radius r about e are ??? with the collection
U
r
: r > 0, so the closure of each open d-ball is compact
Proof. (Lemma 0.0.13) Let be the left invariant metric resulting from Lemma 0.0.12. We
can assume that each of the open balls
B
r
= x G : (x, e), r
has compact closure for 0 < r 2; afterall, there is an r
0
so that for r < r
0
, B
r
0
is compact by Gs
locally compact nature so recalibrate to make r
0
= 2 if necessary.
For 0 < r < 2 we let U
r
= B
r
. This assures us clearly of (iv) and since well keep these U
r
s, (iv)
is assumed henceforth. Also (i), (ii), and (iii) hold when r+s < 2 by s left invariant metric nature.
February 3, 2009
100 10. METRICS IN COMPACT GROUPS
G is locally compact and satises the second countability axiom so G admits a countable open
base
W
2
n : n N
for its topology, where we can (and do) assume that W
2
n is compact for each n. We dene
U
2
= B
2
W
2
.
Its easy to verify that (i) and (ii) hold for 0 < r < 2 and if r +s < 2 then (iii) holds as well.
Well now inch our way from from (i), (ii), and (iii), (r +s < 2) holding for 0 < r 2 to 0 < r 4.
First we have to dene U
r
for 2 < r < 2
2
. Let 0 < r < 2
2
. Set
U
r
=
_
U
t
1
U
t
m
where the union extends over all t
1
, . . . t
m
so that each t
i
satises 0 < t
i
2 and t
1
+ +t
m
= r.
If 2 < r < 2
2
and t
1
+ t
m
= r where each t
i
> 0 then there must be k, l N so that 1 k < l < m
and t
1
+ t
k
2, t
k+1
+ +t
l
2, and t
l+1
+ t
m
2. Why is this so? Well let k be the least
j
1
so that t
1
+ t
j
1
2, and let l be the least j
2
so that t
j
1
+1
+ t
j
2
2. Then

m
j
2
+1
t
j
2
because otherwise, t
j
k
+1
+ t
m
> 2 and t
1
+ t
j
1
+1
2 too where j
1
+ 1 < j
2
+ 1.
It follows that
U
t
1
U
t
2
U
t
m
(U
t
1
U
t
k
)(U
t
k+1
U
t
l
)(U
t
l+1
U
t
m
)
U
t
1
++t
k
U
t
k+1
++t
l
U
t
l+1
++t
m
by (iv)
U
2
U
2
U
2
,
so U
r
U
2
U
2
U
2
whenever 0 < r < 2
2
. U
2
is compact so U
2
U
2
U
2
U
2
U
2
U
2
is too and U
r
is compact for 0 < r < 2
2
. Since
(U
t
1
U
t
m
)
1
= U
1
t
m
U
1
t
1
= U
t
m
U
t
1
,
we see that