You are on page 1of 52

Math 891

Core Analysis I
Fall 2012

Lecture Notes
by
Benjamin C. Wallace

Instructor
Prof. Serban Belinschi

Queens University
Dept. of Mathematics and Statistics

Introduction

The textbook for this course will be Walter Rudins Real and Complex Analysis.

Review

Definition. A metric space is a pair (X, ), where X 6= is a set and : X X [0, ) is a


metric, i.e. it satisfies
1. nonnegative definiteness, i.e. (x, y) = 0 if and only if x = y,
2. symmetry, i.e. (x, y) = (y, x), and
3. the triangle inequality, i.e. (x, y) (x, z) + (z, y),
for all x, y, z X.
Definition. An open ball in a metric space X is a set of the form {y X : (x, y) < r}, for
some x X and r > 0.
Example 1. X = Rn is a metric space with the metric

p ((x1 , . . . , xn ), (y1 , . . . , yn )) =

n
X

1/p
|xj yj |p

j=1

for any p 1. In the limit p , we get the metric ((x1 , . . . , xn ), (y1 , . . . , yn )) = max |xi
1in

yi | (prove this as an exercise).


Definition. A topological space is a pair (X, ), where is a topology on X, i.e. a collection of
subsets of X that contains and X and is closed under finite intersections and arbitrary unions
n
\
[
(i.e. if X1 , . . . , Xn , then
Xi and if X for in some index set I, then
X ).
i=1

The elements of are known as open sets. A set whose complement is open is said to be closed.
A neighbourhood of a point x X is an open set containing x.
Note. Sometimes, in the context of functional analysis, a neighbourhood of a point x is simply
a set containing an open set containing x. We will not follow this convention here.
Example 2. Every metric space is a topological space under the topology generated by open
balls of the metric (i.e. obtained by taking arbitrary unions of open balls). For instance, if
(x, y) = xy , the Dirac -function, then = P(X). Note that any two distinct points of a
metric space admit disjoint neighbourhoods.

Definition. A topological space in which any two distinct points are contained within disjoint
neighbourhoods is said to be Hausdorff.
Example 3. Note that all metric spaces are Hausdorff. If |X| > 1 and = {, X}, then X can
be seen not to be Hausdorff. Hence, cannot be induces by any metric.
Definition. If X and Y are topological spaces and x X, then a function f : X Y is said
to be continuous at x if for any neighbourhood W of f (x), there exists a neighbourhood V of x
such that f (V ) W . We say that f is continuous if it is continuous at each point in X.
It is a good exercise to show that f is continuous if and only if for all open sets W in Y , the
preimage f 1 (W ) is open in X.

Measure and Integration

3.1

Measurable Sets

Definition. Given a set X, a collection M of subsets of X is called a -algebra if


1. X M,
2. M is closed under complement (i.e. if A M, then AC M),
3. M is closed under countable union (i.e. if An M for n N, then

An M).

In this case, elements of M are called measurable sets and (X, M) is a measurable space.
Example 4.
1. M = {0, X} is a -algebra on X.
2. M = P(X) is a -algebra on X.
Remark.
1. It is immediate that if M is a -algebra, then M.
2. By De Morgans law, a -algebra is closed under countable intersection.
3. It is also easily seen that a -algebra if closed under set differences (i.e. if A, B M, then
A \ B M).
Definition. If X is a measurable space and Y is a topological space, then a function f : X Y
is said to be measurable if f 1 (V ) is measurable for each open set V Y .

Note. Oftentimes, in the definition above, Y is a measurable space instead of a topological


space. We will not follow this convention in this course.
The following proposition follows immediately from the definitions.
Proposition 3.1. If f : X Y is measurable and g : Y Z is continuous, then g f : X Z
is measurable1 .
Theorem 3.2. Let X be a measurable space and Y be a topological space. If u1 , . . . , un : X R
are measurable2 and : Rn Y is continuous, then
h :X Y
x 7 (u1 (x), . . . , un (x))
is measurable.
Proof. Define
f :X Rn
x 7 (u(x), v(x))
. Since h = f , by the previous proposition we need only show that f is measurable. Let V
1
be an open set in Rn . For each s = (s1 , . . . , sn ) V Qn , let s = d (s, V C ). Since V is open,
2
[
each s > 0. Moreover, letting Rs = {p Rn : d (p, s) < }, we have that V =
Rs , and
sV Qn

so f

(V ) =

(Rs ). Since this is a countable union, we need only show that f 1 (Rs )

sV Qn

is measurable. But by our use of the -norm, we have that each Rs = I1 In , where the
Ik are intervals in R. Therefore,
f 1 (Rs ) = {x X : (u1 (x), . . . , un (x)) I1 In }
= {x X : u1 (x) I1 } . . . {x X : un (x) In }
1
= u1
1 (I1 ) . . . un (In ).

Each of the sets in the last line is measurable by hypothesis, so we are done.
Remark. Let X be measurable and f : X C, so we can write f = u + iv, where u, v :
p
X R. Then if f is measurable, so are u, v, and |f |, where |f |(x) = |f (x)| = u2 (x) + v 2 (x).
Conversely, if u and v are measurable, then so is f (and therefore |f |). This follows easily from
the previous two propositions.
1

Note that it is implicit here that X is a measurable space and Y and Z are topological spaces.
In dealing with Euclidean spaces Rn , unless otherwise indicated, we will be using the usual topology induced
by the Euclidean metric.
2

Remark. Assume X is a measurable space and f, g : X R(C) are measurable. Then by the
previous proposition, f + g and f g are also measurable. Thus, the space L(X) of measurable
functions X R(C) has the structure of an algebra, i.e. a vector space V equipped with an
associative multiplication rule V V V that distributes over vector addition. These kinds of
structures are very important in functional analysis.
Theorem 3.3. If X is a set and F P(X), then there exists a unique minimal -algebra on
X containing F.
\
Proof. The desired -algebra is easily seen to be given by
M, where the intersection is over
MF

all -algebra M containing F.


Definition. We call the minimal -algebra containing a class F of sets the -algebra generated
by F and denote it (F).
Definition. If X is a topological space, the -algebra on X generated by the open sets of X is
called the Borel -algebra on X and its elements are called Borel sets. We may refer to functions
measurable with respect to the Borel -algebra Borel measurable functions.
We typically assume a given topological space is a measurable space under its Borel -algebra.
Theorem 3.4. Assume that X is a measurable space with -algebra M and that Y is a topological
space. Let f : X Y and let = {E Y : f 1 (E) M}. Then the following hold.
1. is a -algebra on Y .
2. If f is measurable and E is a Borel set in Y , then f 1 (E) M. Hence, contains all
the Borel sets of Y .
3. If Y = [, ] and f 1 ((, ]) M for all Q, then f is measurable3 .
4. If f is measurable, Z is a topological space, and g : Y Z is Borel measurable, then g f
is also measurable.
Proof.
1. This is straightforward to verify.
2. If f is measurable, then contains all open sets in Y . Since is a -algebra, and by
minimality of the Borel -algebra, it follows that must contain the Borel sets of Y .
3

As usual, the topology in Y is the one generated by intervals of the form (, ), (, ], [, ), and [, ].

3. For any R, we may write [, ) = [, ]C =

(q, ] to see that

>qQ

f 1 ([, )) =

\

C
M,
f 1 ((q, ])

since this intersection is countable. Similarly, f 1 ((, ]) M, for any R. Thus,


f 1 ((, )) = f 1 ([, ) (, ]) = f 1 ([, )) f 1 ((, ]) M.
[
Since for any open set E in Y , we can write E =
(q q , q + q ), where q = d(q, E C )
qQE

and, moreover, we can approximate these q by rationals as we did above, we get that
f 1 (E) M.
4. We leave this as an exercise.

Definition. F -sets are countable unions of closed sets. G -sets are countable intersections of
open sets.
In particular, F - and G -sets are Borel.
Definition. If Y is partially ordered set that admits sup and inf, then we define the limit
superior of an by
lim sup an = inf sup ak = lim sup ak
n

n0 kn

n kn

and the limit inferior of an by


lim inf an = sup inf ak = lim inf ak .
n

n0 kn

n kn

Example 5.
1. The limit superior/inferior are always defined in [, ].
2. Let F(X, [, ]) = {f : X [, ]} and define the ordering f g if and only if
f (x) g(x) for all x X. Let H F(X, [, ]) be bounded from above. Then sup H
is the function h given by h(x) = sup{f (x) : f H}. The infimum is given similarly. We
shall frequently use this ordering for sets of real-valued functions.
3. If X is measurable, let L(X, [, ]) = {f : X [, ] : f is measurable} and use
the same ordering as above. We then have the following theorem.
Theorem 3.5. If fn is a sequence in L(X, [, ]), then sup fn , inf fn , lim sup fn , and lim inf fn
n

are all measurable, i.e. they are in L(X, [, ]).


5

Proof (Sketch). Let h = sup fn , so h(x) = sup fn (x). Then


n

h1 ((, ]) = {x X : h(x) > }


= {x X : fn (x) > , for some n}
[
fn1 ((, ])
=
|
{z
}
n
M

M.
By the previous theorem, this is enough to show that h is measurable. The proof for the infimum
is analogous; the case of lim sup and lim inf follow directly.
Corollary 3.6.

1. If fn L(X, [, ]) converge to g pointwise, then g L(X, [, ]).

2. If f and g are measurable, then so are max(f, g), min(f, g), f + = max(f, 0), f =
min(f, 0), and |f | = f + + f .
Note that f = f + f .
Proposition 3.7. If f = g h and h, g 0, then f + g and f h.
Definition. If X is a measurable space, then s : X C is called simple if s(x) =

n
X

j Aj (x),

j=1

where the Aj are measurable and partition X, A is the characteristic function of A, and j C.
This is equivalent to saying that s is a measurable function with finite range.
Note that sometimes, C is replaced with some other topological space in the above definition.
Throughout this course, the range of simple functions will always be a subset of [0, ).
Theorem 3.8. If X is a measurable space and f : X [0, ] is measurable, then there exists
a sequence sn of simple functions that converges pointwise to f . Moreover, we can assume that
0 si si+1 f for all i.


j+1
j
n
1
n, n n
and let A2n =
Proof. For n N and j = 0, . . . , 2 1, let Aj = f
2n
2
f 1 ([n, ]). Then the Aj (j = 1, . . . , 2n ) form a measurable partition of [0, ] and sn (x) =
2n
X
j
n
defines the required sequence.
2n Aj (x)
j=0

3.2

Positive Measures

Definition. A (positive) measure on a measurable space X is a function : M [0, ] that is


countably
i.e. if An M is a countable sequence of mutually disjoint, measurable sets,
[additive,
 X
then
An =
(An ). A measure space is a measurable space equipped with a (positive)
measure.
6

Note that any measurable space can trivially be made into a measure space by the measure
0.
Remark. Let be a measure on M and suppose (E) (0, ) for some E M. Then the
following statements are easily verified to be true.
1. () = 0.

2. is finitely additive, i.e.

n
[

Aj =

j=1

n
X

(Aj ), whenever the Aj are mutually disjoint.

j=1

3. is monotonic, i.e. if A B, then (A) (B).


!
4. is continuous from below, i.e. if Bi Bi+1 , then

Bn

= lim (Bn ).

5. is continuous from above, i.e. if Bi Bi+1 , then


(Bm ) < for some m.

\


Bn = lim (Bn ), provided that
n

In all that follows, we shall use the convention


(
, 0 < a
a=a=
.
0,
a=0

3.3

Integration of Positive Functions

We are moving towards the construction of the Lebesgue integral. First, we consider the case of
measurable functions with nonnegative range. We approximate the integral of such a function
by the (obviously defined) integrals of approximating (from below) simple functions.
Definition. Let X be a measure space with measure , let f : X [0, ] be a measurable
function, and let E X be a measurable set. Define
Z

Z
f d = sup
s d : 0 s f is simple ,
E

where if s =

Z
j Aj (x) is simple, then

s d =

Remark. The following statements are easily verified.


Z
Z
1. If 0 f g, then
f d
g d.
E

Z
2. If A B, then

Z
g d

g d.
B

j (Aj E).

Z
3. If c [0, ), then

f d.

cf d = c
X

Z
4. If (E) = 0, then

f d = 0.
E

Z
5. If f |E 0, then

f d = 0.
E

Proposition 3.9. If s and t are simple functions on X, then

s + t d =

s d +

Proof. Write s =

n
X

j Aj and t =

j=1

m
X

t d.
X

j Bj . Let Ejk = Aj Bk . These sets form a measurable

j=1

partition of X. Moreover, s + t =

XX
j

Z
On the other hand,

s d =
X

XX
(j + k )(Ejk ).

s + t d =

(j + k )Ejk , so

Z
j (Aj ) and similarly for

t d. The rest follows by


X

comparison.
The proof of the following proposition is left as an exercise.
Z
Proposition 3.10. If f : X [0, ] is measurable, then (E) =

f d defines a positive
E

measure on X.

Theorem 3.11 (The Lebesgue monotone convergence theorem). If fn is an Zincrease sequence


of
Z
measurable functions and if f (x) = lim fn (x), then f is measurable and lim
n

fn d =
X

f d.
X

Proof. Since
of a set of measurable functions is measurable, f is measurable.
Z the supremum
Z
Moreover,
fn d
f d, so
X

Z
fn d

lim
n

f d.
X

Z
n

Z
fn d c

For the converse, we will show that for all 0 < c < 1 we have lim
X

f d. It will
X

then follow that this holds when c = 1. So fix 0 < c < 1, take a simple function 0 s f on
X, and define
En = {x X : fn (x) cs(x)} = (fn cs)1 ([0, ]).
Then it is clear that the En form an increasing sequence of measurable sets. Moreover, if we
[
suppose x
/
En , so that for all n,
n

fn (x) < cs(x),

then we get that


f (x) = lim fn (x) cs(x) < s(x) f (x),
n

a contradiction. Hence,

En = X.

Now since the sequence fn increases to f , there exists an n (depending on x and c) such that
fn (x) cf (x) cs(x). For such n,
Z
Z
Z
fn d
fn d c
s d.
X

En

En

Z
But the function defined by (E) =

s d (for E measurable) is a measure, so


E

!
(X) =

En

= lim (En ),
n

and therefore
Z
c

Z
s d lim

s d = c(X) = c lim (En ) = c lim


n

i.e.

En

fn d,
X

Z
s d lim

fn d.
X

Since s was arbitrary, we get


Z
Z
Z
c f d = c sup
s d lim fn d,
0sf

concluding the proof.


Theorem 3.12. If fn : X [0, ] is measurable, then f =
Z
X

Z
fn is measurable and

f d =
X

n=0

fn d.

n=0 X

Proof. This follows from the monotone convergence theorem and linearity of the integral of
simple functions.
XX
XX
Corollary 3.13. If aij [0, ) for i, j N, then
aij =
aij .
i

Proof. Apply the previous theorem with fi (j) = aij and X the space of mappings N [0, )
with the counting measure.
Theorem 3.14 (Fatous lemma). If fk : X [0, ] is measurable for each k, then
Z
Z
lim inf fk d lim inf
fk d.
X

Proof. Let gk (x) = inf fi (x) for each k. Then gk fk , so


ik

Z
gk d

fk d.

Moreover, the gk form an increasing sequence of nonnegative measurable functions with pointwise
limit lim inf fn . Thus, by the monotone convergence theorem
n

Z
lim inf fn d = lim
n

Z
gk d lim inf

gk d = lim inf

fk d.

Example 6. Let X = [1, 1] with = m, the Lebesgue measure, and define f2k (x) =
(
Z
1, x [0, 1]
fk d = 1.
and f2k+1 (x) = 1 f2k (x). Then lim inf fk = 0, but lim inf
X
0, x [1, 0)

3.4

Integration of Complex Functions

We now turn to the integration of functions whose values are not restricted to lie in the subset
[0, ] of C. Let us introduce the space


Z
1
L () = f : X C measurable :
|f | d < .
X

Note that |f | above denotes the complex modulus of f .


f +f
f f
, v = Im(f ) =
, and
2
2
+ +

let u , u , v , and v be the positive and negative parts of u and v, as defined earlier (i.e.
u+ = max(u, 0), u = min(u, 0), etc.). Then we define the Lebesgue integral of f over a
measurable set E X as
Z
Z
Z
Z
Z
+

+
f d = u d u d + i v d i v d,
Definition. Let f : X C be measurable, u = Re(f ) =

provided f L1 ().
Remark. If f L1 (), then
Z
X

Z
Z



|f | d < .
f d f d
X

Theorem 3.15 (Lebesgue dominated convergence theorem). Let (X, ) be a measure space and
let fn be a sequence of (generally speaking, complex-valued) measurable functions on X. Assume
lim fn (x) = f (x) exists for each x X; thus, f is measurable. If there exists g L1 () such
n

that |fn (x)| g(x) for all x X, then f L1 () and


Z
lim |fn f | d = 0.
n

10

Z
n

f d.

fn d =

It follows that lim

Proof. Since |fn (x)| g(x), it follows that |f (x)| g(x). Thus,
Z
Z
g d < ,
|f | d
X

since g L1 (). Hence, f L1 ().


Similarly, we have that |fn (x) f (x)| |fn (x)| + |f (x)| 2g(x) (for all x X), thus
fn f L1 (). Now
0

Z
2g d =

z
}|
{
(2g lim inf |fn f |) d
n

ZX

lim inf (2g |fn f |) d


n
Z

Z
lim inf
2g d
|fn f | d
n
X
 X

Z
Z
= 2g d + lim inf
|fn f | d
n
X
Z
Z
=
2g d lim sup
|fn f | d.
=

(by Fatous lemma)

That is,
Z

2g d

2g d lim sup

so

|fn f | d,
X

Z
|fn f | 0.

lim sup
n

Z
But since |fn f | 0, it follows that lim sup
n

|fn f | d = 0 and likewise for the limit inferior.


X

Therefore,
Z

|fn f | d = 0.

lim
n

This of course implies that


Z
lim
n

3.5

Z
fn d =

f d.
X

Sets of Measure Zero

If (X, ) is a measure space, then a property P (x) dependent on x X is said to hold almost
everywhere or (or a.e. ()) if
({x X : P (x) does not hold}) N
11

for some measurable N with (N ) = 0. For instance, we say that f = g a.e. () whenever
the set of points in X for which f (x) 6= g(x) has measure 0. Note that if g is in the set of
measurable functions that agree with f almost everywhere (with respect to Lebesgue measure),
then in general not a single value of g is fixed. That is, any single value of g can be changed
without altering the fact that it agrees with f almost everywhere.
Note that we did not define the notion of a property holding almost everywhere by saying
that the set on which the property does not hold has measure 0; after all, this set may not be
measurable. However, if it is contained in a set of measure 0, our intuition tells us that it should
be in the same sense negligible. To make things more concrete, we may wish to set all subsets of
a set of measure 0 to have measure 0 themselves. However, we must be careful to make sure that
the inclusion of these previously non-measurable sets into our -algebra preserves the fact that
it is a -algebra; we must also ensure that the extension of our measure to this new -algebra
is still a measure. The following theorem tells us that we can always do this.
Theorem 3.16. Let (X, M, ) be a measure space and let
M = {E X : A, B M, A E B, (A) = (B), (B \ A) = 0}.
Extend to M by setting (E) = (A) whenever E and A are as in the definition of M above.
Then M is a -algebra and is a (well-defined) measure on M .
Proof. First, let us show that is well-defined. Choose A1 , B1 , A2 , B2 M corresponding to a
given E X. So (for i = 1, 2) (Ai ) = (Bi ), and therefore (Bi \ Ai ) = 0. Moreover,
Ai E B i A1 \ A2 E \ A2 B 2 \ A2
(A1 \ A2 ) (B2 \ A2 ) = 0
(A1 \ A2 ) = 0
and similarly, (A2 \ A1 ) = 0. Hence, (A1 ) = (A2 ).
Next, let us show that M is a -algebra. First, It is clear that X M . Second, if E M
and A E B, then B c E c Ac . Since Ac \ B c = B \ A, it follows that (Ac ) = (B c ).
Thus, E c M . Finally, suppose that En M is countable sequence of sets in M with
[
An En Bn , where An , Bn M and (An ) = (Bn ). Letting B =
Bn M and
n
[
A=
An M , we get
n
[
[
B \ A = (Bn \ A) (Bn \ An ).
n

Hence, (B \ A)

(Bn \ An ) = 0. It follows that E =

[
n

En is in M .

It is left as an exercise to show that is countably additive on M .

12

Note. In general, M 6= 2X .
Remark.
Z
1. If f : X [0, ] and

f d = 0, then f = 0 a.e. on E.
E

Z
2. If

f d = 0 for all measurable E and f L1 (), then f = 0 a.e. on X.

Z
Z



3. If f d =
|f | d, then there exists C such that f = |f | a.e. ; we call the
X
X
argument of f .
4. Suppose (X) < and f L1 (). Let S = S C be such that
Z
1
f d S
(E) E
for all E M with (E) > 0. Then f (x) S for almost every x X.

The Riesz Representation Theorem

Let us recall a few basic definitions.


Definition. If K is a field (usually R or C), then V is a vector space over K if V is an abelian
group (written additively) with an operation : K V V , called scalar multiplication that is
associative with respect to multiplication in the field K and distributes over addition in V and
such that addition in K distributes over scalar multiplication, 1v = v (for v V and 1 K),
and 0K v = 0V (for v V , 0K K, and 0V V ).
If a vector space V (over K) is endowed with an operation V V V (called multiplication)
that is associative with respect to itself and scalar multiplication distributive over addition, then
V is said to be an algebra (over K). If the multiplication is, moreover, commutative, then V is
a commutative algebra.
Definition. If V and W are vector spaces over K, then : V W is linear if (v + w) =
(v) + (w) and (v) = (v) for all v V , w W , and K. If, moreover, V and W are
algebras and (vw) = (v)(w) for all such v, w, then is called an algebra homomorphism.
Definition. A topological space is said to be locally compact if every point in it has a relatively
compact neighbourhood (i.e. a neighbourhood whose closure is compact).
Definition. Recall that a sequence xn in a metric space (X, d) is Cauchy if for every  > 0,
there is an N N such that d(xm , xn ) <  whenever m, n N . A metric space X is said to be
complete if for every Cauchy sequence xn , there is x X such that lim d(xn , x) = 0.
n

13

Example 7.
1. For any n 1, Cn is a vector space over C.
2. If K is a topological space, then C(K) = {f : K C : f continuous} is a vector space
over C. If, moreover, K is compact, then kf k = max |f (x)| is a norm on C(K) (because
xK

a continuous function on a compact space attains its maximum), called the maximum
norm or the infinity norm. Under the metric induced by this norm, C(K) is a complete
metric space. In addition, C(K) is a commutative algebra under pointwise multiplication
of functions.
3. Let K be a locally compact Hausdorff topological space and let C(K) be as above. If
f C(K), then we define supp(f ) = {x K : f (x) 6= 0}, the support of f , where the
bar denotes closure. Let Cc (K) = {f C(K) : supp(f ) compact}. Then Cc (K) is a
vector space with the maximum norm. Note that under the topology induced by this
norm, Cc (K) is complete if and only if K is compact. For if K is not compact, we can
take a sequence of functions in Cc (K) whose supports are growing without bound. If K
is compact, Cc (K) = C(K).
4. Let C0 (K) = Cc (K), where the bar here denotes completion. This is the class of functions
that vanish at infinity. When K is compact and Hausdorff, we shall denote C0 (K) by
C(K) from now on and consider it a normed space with the maximum norm.
5. If D is open in C, then A(D) = {f : D C : f analytic} can be made into a metric space,
but the corresponding metric is not induced by a norm.
Theorem 4.1. A closed subset of a compact set is compact.
Corollary 4.2. A subset of a relatively compact set is relatively compact.
Theorem 4.3. Suppose X is a Hausdorff space, K X is compact, and p K c . Then there
are disjoint open sets U and W with p U and K W .
Corollary 4.4.

1. Compact subsets of Hausdorff spaces are closed.

2. If F is closed and K is compact in a Hausdorff space, then F K is compact.


Theorem 4.5. If K is a collection of compact subsets of a Hausdorff space and if

K = ,

then some finite subcollection of K also has empty intersection.


Theorem 4.6. Let X be a locally compact Hausdorff space and suppose K X is compact,
V X is open, and K V . Then there is a relatively compact open set U such that K U
U V.
14

Definition. We call a function real-valued4 function f on a topological space lower semicontinuous if f 1 ((, )) is open for all R and upper semicontinuous if f 1 ((, )) is open
for all R.
A real-valued function f is continuous if and only if it is both upper and lower semicontinuous.
Theorem 4.7. Suppose f : X Y is continuous. If K X is compact, then so is f (K).
Corollary 4.8. The range of any f Cc (X) is a compact subset of the complex plane.
Definition. Let f Cc (X) satisfy 0 f 1. If K is a compact set with f |K = 1, then we
write K f . If V is an open set with supp(f ) V , we write f V . If both these conditions
hold, we write K f V .
A function K f V lies between the characteristic functions of K and of V . The
following highly useful theorem says we can always find such a function.
Theorem 4.9 (Urysohns lemma). Let X be a locally compact Hausdorff space. If =
6 KX
is compact and V K is open, then there is an f Cc (X) such that K f V .
Proof. We can find a relatively compact open set V0 with K V0 V 0 V . Since K is
compact and V0 is open, we can then find a relatively compact open set V1 with
K V1 V 1 V0 V 0 V.
Continuing this process, we get a countable collection of relatively compact open sets Vr indexed
by rationals r [0, 1] Q such that V s Vr whenever r < s. For each r (0, 1) Q, let us
define
fr = rVr
f = sup fr
r

gr = V s + sV c

g = inf gs .
s

Then supp(f ) V , 0 f 1, and f |K = 1. To show that f is the desired function, we must


prove that it is continuous. Now

,
>r

1
fr ((, )) = Vr , 0 < r ,

X, < 0
4

This definition also extends to extended-real-valued functions.

15

so fr is lower semicontinuous and so is f . Similarly, it can be shown that g is upper semicontinuous. So to prove continuity of f , we will show that f = g.
Now let r, s Q (0, 1). If x
/ Vr , then
rVr (x) = 0 (V s + sV c )(x).
s

If x Vr , then rVr (x) = r and there are two cases to consider. If x V s , then we get
Vr (x) = r < 1 = V s + sV c .
s

If x
/ V s , then we must have r < s (otherwise, V r Vs ), so
rVr = r < s = V s + sV s .
In all cases, we have
rVr (x) V s + sV c .
s

Thus, 0 fr gs 1 for all r, s Q (0, 1). It follows that 0 f g 1.


To show that f = g, suppose otherwise; so there exists x X for which f (x) < g(x). So we
can take r, s Q [0, 1] satisfying
f (x) < r < s < g(x).
It follows that fr (x) < r and s < gs (x). But these two statements imply together that x V s \Vr .
This is a contradiction, since r < s implies that V s Vr .
Corollary 4.10 (Partition of unity). If V1 , . . . , Vn are nonempty open sets in a locally compact
Hausdorff
! space X and 6= K V1 . . . Vn is compact, then there exist hi Vi satisfying
n

X

hi = 1.

i=1

Proof. By Theorem 4.6, all points x K have a relatively compact neighbourhood Wx Vi for
[
some i. Then
Wx is a covering of K, so by compactness we can find x1 , . . . , xN K such
xK

that K Wx1 . . . WxN . For each i, let Hi be the union of those W xj lying in Vi . Then the Hi
are compact and contained in Vi , so by Urysohns lemma, there exist gi such that Hi gi Vi .
Now define h1 = g1 and for each i = 2, . . . , N , define
hi = (1 g1 ) . . . (1 gi1 )gi .
Then it is easy to see by induction that
h1 + + hj = 1 (1 g1 ) . . . (1 gj ) [0, 1]
for each j = 1, . . . , n. Thus,
h1 + + hj = 0
16

whenever g1 , . . . , gj are 0 and


h1 + + hj = 1
whenever at least one g1 , . . . , gj is 1. But for each x K H1 . . . Hn , there is always such a
gi , so
!
n

X

hj = 1.

i=1

Definition. A measure on a space X is said to be outer regular if (E) = inf{(V ) : E


V open} and inner regular if (E) = sup{(K) : E K compact}, for all measurable subsets
E X. If is both inner and outer regular, then we call regular.
Theorem 4.11 (Riesz representation theorem). Let X be a locally compact Hausdorff space. If
: Cc (X) C is a positive linear functional, then there is a -algebra M containing the Borel
sets of X and a unique positive measure on M such that
Z
1. f =
f d,
X

2. (K) < for any compact K X,


3. is outer regular, i.e. (E) = inf{(V ) : E V open},
4. if E is open and [[or?]] (E) < , then (E) = sup{(K) : K E compact} (so is
inner regular on open sets of finite measure), and
5. M is a complete -algebra.
Proof. We begin by proving uniqueness, so assume that existence holds. Assume
Z
Z
f = f d1 = f d2
for some measures 1 and 2 satisfying the above conditions. By properties (3) and (4) above,
we need only show that 1 (K) = 2 (K) for all compact K. Now for such K, by (3), there exists
an open set V K such that 2 (V ) < 2 (K) +  (for any  > 0). Moreover, by Urysohns
lemma, we can find K f V . Thus,
Z
Z
Z
Z
1 (K) =
K d1
f d1 = f =
f d2
V d = 2 (V ) < 2 (K) + ,
X

i.e. 1 (K) < 2 (K) + . Similarly, we can get 2 (K) < 1 (K) + . Thus, 1 = 2 .
To prove existence, let (V ) = sup{f : f V } for all V open. Then
f1 f2 f2 f1 0
(f2 f1 ) = f2 f1 0,
17

so it follows that V1 V2 implies (V1 ) (V2 ), so is monotonic (on open sets). To extend
to all subsets E X, let
(E) = inf{(V ) : E V open}.
It is easy to verify that is then monotonic for all subsets of X. However, will not in general
be a measure on the power set, so we define the classes of sets
MF = {E X : (E) = sup{(K) : E K compact} < }
and
M = {E X : E K MF , for all K compact}.
We will show that is a measure on M. We proceed in nine steps.
Step 1: show that is countably subadditive on M. Let V1 , V2 be open and let g V1 V2 .
By the partition of unity, we can find hi Vi such that h1 + h2 = 1 on supp(g) (for i = 1, 2).
Then ghi Vi and g = g(h1 + h2 ) = gh1 + gh2 , so
g = gh1 + gh2 (V1 ) + (V2 ).
Thus,
(V1 V2 ) = sup{g : g V1 V2 } (V1 ) + (V2 ).
Now let En M for n N. If (En ) = for some n, then countable additivity holds for these
En (by definition of ). Otherwise, let  > 0. Then for all n, there is Vn En open such that
[
[

Vn
En = E and choose f V . Then supp(f ) V is
(Vn ) (En ) + n+1 . Let V =
2
n
n
n0
n0
[
[
compact, so there is n0 such that supp(f )
Vi . That is, f
Vi . Thus, by definition of
i=1

i=1

, finite subadditivty of on open sets (shown above), and definition of the Vi ,


f (V1 . . . Vn0 ) (V1 ) + + (Vn0 ) < (E1 ) + + (En0 ) + .
Thus,
(E) (V )

(En ) + ,

n=1

where  > 0 is arbitrary.


Step 2: show that MF contains all compact sets K and that
(K) = inf{f : K f }.
To do this, let K f , (0, 1), and V = f 1 ((, )), so that K V . Since g V implies
g f , we have
1
(K) (V ) = sup{g : g V } f.

18

Since was arbitrary, (K) f . But since f , too, was arbitrary, we get (K) inf{f :
K f } < . This proves (2) and shows that K MF .
Now by definition of , for all  > 0, there is an open set V such that (V ) (K) +  and
V K. For such V , we can find K f V . For such f , we get
(K) f (V ) (K) + .
Thus,
(K) inf{f : K f } (K) + 
for all  > 0, so
(K) = inf{f : K f }.
Step 3: show that every open set of finite measure is in MF . Let E be such a set and let
< (E). By definition of , there is f E with < f . Now for any W supp(f ) (so
f W ), we have f (W ), so f (K) for some compact K (e.g. W ). For such a K,
< f (K).
So for any < (E), there is a compact K with < (K). Thus, (E) = sup{(K) : K
E compact}.
Step 4: show that MF is closed under countable union when such a union has finite measure
and that is countably subadditive on MF . We begin by letting K1 , K2 be disjoint compact
sets. Then by Urysohns lemma, we can find f Cc (X) with K1 f X \ K1 , so in particular
f |K2 = 0. Now let  > 0. Then by Step 2, we can find g Cc (X) with K1 K2 g and
g < (K1 K2 ) + . Then K1 f g and K2 (1 f )g. Therefore,
(K1 ) + (K2 ) (f g) + ((1 f )g) = g < (K1 K2 ) + .
Since  was arbitrary, we get
(K1 ) + (K2 ) (K1 K2 ).
But by Step 1 (countable subadditivity on M) and Step 2, this becomes
! an equality.
[
Next, let En MF be a sequence of disjoint sets. If
En = , then countably
n

additivity holds; so assume otherwise [[where is this used?]]. For any  > 0 and each j, we can

19

find Hj Ej compact with (Ej ) (Hj ) +


2j+1

. Then

Ej (H1 . . . Hn )

j=1

n
X
j=1
n
X

(Hj )

(as was just shown)

(Ej ) .

j=1

!
[

(En ). Again by Steps 1 and 2, this


[
En MF [[why?]].
becomes an equality. Moreover, as H1 . . . Hn is compact,

Since  was arbitrary, this gives us

En

[[incomplete]]

4.1

Lebesgue Measure and Regularity of Borel Measures

Let X = Rk and f Cc (Rk ), so supp(f ) = f 1 (C \ {0}) is compact. Thus, f 1 (C \ {0})


k
Y
[M, M ] for some integer M > 0. Now for any integer n > 0, we can cut [0, 1] in to 2n equal
j=1

pieces, thereby cutting [0, 1]k into 2nk pieces, each with edge length 1/2n . Repeat this process
for all hypercubes of edge length 1 in Rk and denote the set of centers of these sub-hypercubes
by c(n) . Define : Cc (Rk ) C by
X
X
n f =
f (q)2nk =
f (q)2nk max |f (q)|(2M )k .
qc(n)

qc(n) |q||(M,...,M )|

qRk

We define the Lebesgue measure


m to be the unique measure obtained from the Riesz repreZ
sentation theorem such that
f dm = lim sup n (f ) (the right-hand side is just the Riemann
X

integral of f ).
Remark. On Cc (Rk ), the Lebesgue integral with respect to the Lebesgue measure is identical
to the Riemann integral.
Recall that the measure as obtained in the Riesz representation theorem was not quite inner
regular. Under certain strengthenings of the hypotheses of this theorem, the resulting measure
can be assured to be regular.
Definition. Call a set E -compact if E =
if E =

Kn for some compact sets Kn . Call E -finite

n=1

Cn for some measurable sets Cn of finite measure.

n=1

20

Theorem 4.12. Let X be a locally compact -compact Hausdorff space and let M be a -algebra
on X containing the Borel sets. If is a positive measure satisfying the conclusions of the Riesz
representation theorem, then the following are true:
1. if E M and  > 0, then there exist a closed set F and an open set V such that F E V
and (V \ F ) < ;
2. is (Borel) regular;
3. there is an F -set A and a G -set B such that A E B and (B \ A) = 0.
Proof.

[[Incomplete]]

Corollary 4.13. If X is a locally compact Hausdorff space with each of its open sets -compact
and if is a positive Borel measure such that (K) < for all K compact, then is regular.
Theorem 4.14 (Lusins theorem). Let X be a locally compact Hausdorff space and be a
measure on X satisfying the conclusions of the Riesz representation theorem. If f : X C is
measurable, A X is measurable, (A) < , and f |Ac = 0, then for all  > 0, there exists
g Cc (X) such that ({x X : f (x) 6= g(x)}) <  and sup |g(x)| sup |f (x)|.
Proof.

[[Incomplete]]

Corollary 4.15. If f is as in Lusins theorem, then there is a sequence gn Cc (X) such that
lim gn (x) = f (x) for -almost every x X.
n

Theorem 4.16 (Vitali-Caratheodory theorem). If f L1 () is real-valued and  > 0, then


there exists an upper semicontinuousZ function u : X R and a lower semicontinuous function
v : X R such that u f v and

(v u) d < .

Banach, Hilbert, and Lp Spaces

Definition. Let X be a complex vector space. A map k k : X [0, ) satisfying


1. kx + yk kxk + kyk (the triangle inequality) and
2. kxk = ||kxk (homogeneity)
(for all x, y X and C) is called a seminorm. If, in addition, k k is nonnegative definite
(i.e. kxk = 0 if and only if x = 0), then it is called a norm. A vector space with a norm is called
a normed (vector) space.

21

Remark. If k k is a seminorm on X, then {x X : kxk = 0} is a vector subspace of X. It is


easy to verify that
X/X0 = {
x : x X},
where x
= {y X : x y X0 }, is a vector space under the obvious operations; moreover, the
map
k k : X/X0 [0, )
x
7 kxk
is well-defined and a norm on X/X0 .
Example 8.

1/p
n
X
1. The vector space Cn is normed under kxkp =
|xj |p
for any p [1, ).
j=1

2. If X is compact, then C(X) is normed under kgk = max |g(x)|.


xX

Definition. A bilinear map (, ) : X X C on a (complex) vector space X is said to be


sesquilinear if
1. (x, y) = (y, x),
2. (x + x0 , y) = (x, y) + (x0 , y), and
3. (x, y) = (x, y)
for all x, y X. If (x, x) 0 and (x, x) = 0 if and only if x = 0, then (, ) is said to be
nonnegative definite. If (, ) is both sesquilinear and nonnegative definite, then it is called an
inner (or scalar ) product. A vector space with an inner product is an inner product space.
Note that an inner product (, ) induces a norm via kxk2 = (x, x). Thus, every inner
product space is a normed space. In turn, a norm k k on a vector space X defines a metric
d(x, y) = kx yk which in turn generates a topology on X.
Definition. A complete normed vector space is called a Banach space. A complete inner product
space is called a Hilbert space.
Clearly, every Hilbert space is a Banach space.
Example 9.
1. C([0, 1]) is a Banach space under kf k = max |f (t)|.
0t1

22

2. P([0, 1]) = {

n
X

aj t : aj C, n N} is normed under k k . However, by the Weierstrass

j=0

approximation theorem, P([0, 1]) = C([0, 1]), so P([0, 1]) is not complete.
n
X
3. P([0, 1]) = {
aj t : aj C, n N} is normed under
j=0

kf k,1 = max |f (t)| + max |f 0 (t)|


0t1

0t1

and under the topology induced by this norm,


P([0, 1]) = C 1 ([0, 1]) = {f C([0, 1]) : f is C 1 }.
We can define similar norms k k,m such that the closure above becomes C m ([0, 1]), the
space of C m functions on [0, 1].
4. C ([0, 1]) = {f C([0, 1]) : f is infinitely differentiable} is not a Banach space.
Z
5. One can define the inner product (f, g) =
f g dm on C([0, 1]). This induces the norm
[0,1]

k k2 , a special case of k kp defined above. C([0, 1]) is not complete under this norm (we
can approximate, say, step functions by continuous functions with this norm).

5.1

Linear Operators

Let X and Y be normed spaces over C.


Definition. A linear operator : X Y is said to be bounded if kk = sup kxk < (note
kxk1

the three different uses of the notation k k here). We call a contraction if kk 1 and an
isometry if kxk = kxk for all x X.
In particular, if is a contraction, then kk = 1 (however, the converse is not true in
general).
Theorem 5.1. The following are equivalent:
1. is continuous;
2. is continuous at a point;
3. is bounded.
Proof. Obviously, (1) implies (2).
To show that (3) implies (1), let x0 X and let  > 0. Then for any x 6= x0 ,




x

x
0

kx x0 k = k(x x0 )k =
kx x0 k kx x0 k sup kykkx x0 k = kkkx x0 k,
kyk1
23


.
kk
To show that (2) implies (3), suppose is continuous at x0 X. Then by definition, there
exists 1 > 0 such that when kx x0 k < 1 ,
so kx x0 k <  whenever kx x0 k <

1 > kx x0 k = k(x x0 )k.


Thus, sup kyk 1, so sup kyk
kyk1

kyk1

1
.
1

Example 10.
1. Let X = Y = C([0, 1]) and fix g C([0, 1]). If
Mg : X X
f 7 f g
then
kMg k = sup kf gk
kf k1

sup

( max |f (t)g(t)|)

sup

( max |f (t)| max |g(t)|)

max |f (t)|1 0t1


max |f (t)|1 0t1

0t1

max |g(t)|
0t1

= kgk .
In fact, taking f = 1, we get kMg k = kgk .
2. Let X = Y = P([0, 1]) and let f = f 0 . Then for fn (t) = tn , we have kfn k = 1, but
kfn k = kntn1 k = n, so is unbounded.
Example 11. Let (X, ) be a measure space, let Lp (X, ) and kf kp be as before (for p [1, ))
and define
L (X, ) = {f : X C : N X, (N ) = 0, sup |f (x)| < }
xX\N

with kf k =

inf

sup |f (x)|. Then k kp is merely a seminorm for all p [1, ]. To

N X(N )=0 xX\N

make this a norm, we need some results on convexity.

5.2

Convex Functions and Inequalities

Definition. Let X be a vector space. A subset X is convex if for all x, y and [0, 1]
we have x + (1 )y . If is convex, then a function : R is said to be convex if for
all x, y and [0, 1],
(x + (1 )y) (x) + (1 )(y).
24

In particular, : (a, b) R is convex if and only if for all a < s < t < u < b,
(u) (t)
(t) (s)

.
ts
ut

(1)

Thus, the derivative of a convex function is increasing, i.e. 00 > 0 whenever it exists.
Remark.
1. Letting 1 in the definition of a convex function, it is easy to see that any convex
function is continuous.
2. If : (a, b) R is convex, then 0 exists everywhere outside an at most countable set.
Show this as an exercise using (1).
Theorem 5.2 (Jensens inequality). If is a probability measure on , f : (a, b) is
measurable, and : (a, b) R is convex, then
Z
 Z

f d
f d.

Note. In the context of probability theory, f (and f ) is called a random variable and
integration of f with respect to a probability measure gives the expectation of f . Thus, the
statement of Jensens inequality becomes that a convex function evaluated at the expected
value of a random variable f is no more than the expected value of the random variable f .
Z
Proof. Let t = f d (a, b), let a < s < t < u < b, and let
= sup
s<t

(u) (t)
(t) (s)

,
ts
ut

where this inequality holds for all u (t, b) by the above mentioned equivalent definition of a
convex function. Thus, if t < f (x), then

(f (x)) (t)
.
f (x) t

On the other hand, if f (x) < t, then by definition of ,


(t) (f (x))
.
t f (x)
In either case, we have
(t) (f (x)) + (t f (x)).
Thus,

Z


f d

Z
= (t) =

Z
(t) d(x)

(f (x)) d(x) +

t d(x) f (x) d(x)


.
| {z } |
{z
}
t

25

Definition. If A V , where V is a vector space, then

n
n
X

X
co(A) =
j xj : xj A, j [0, 1],
j = 1, n 1

j=1

j=1

is called the convex hull of A.


It can be shown that
\

co(A) =

C,

CAC convex

so co(A) is the smallest convex set containing A.


Example 12.
1. Let be a probability measure on . The following are special cases of Jensens inequality:
Z
 Z
exp
f d ef d
Z
Z
ln f d ln f d
(for f : (0, ))
Z
p Z
f d f p d
(for f : [0, ] and p 1)
Z
Z


f d |f | d


Z
1
1
R

d
(for f : (0, ))
f
f d
2. Let = {p1 , . . . , pn } and (pj ) = j > 0 such that

j = 1. Then

Z
f d = 1 f (p1 ) + + n f (pn ),
so by the first inequality in the previous example,
(ef (p1 ) )1 . . . (ef (pn ) )n 1 ef (p1 ) + + n ef (pn ) .
Thus, for all y1 , . . . , yn > 0, we can say that
y11 . . . ynn 1 y1 + + n yn .
In particular, when j =

1
for each j, we get
n

y1 . . . yn

26

y1 + + yn
.
n

3. With as above, the last inequality in the first of these examples gives us
1
1
n

+ +
.
1 f (p1 ) + + n f (pn )
f (p1 )
f (pn )
In particular, when j =

1
, we get
n

n
1

f (p1 ) + + f (pn )
n

1
1
+ +
f (p1 )
f (pn )


.

Thus, for all y1 , . . . , yn > 0,


y1 + + yn

n
+ +

1
y1

1
yn

Theorem 5.3 (H
older and Minkowski inequalities). Let f, g : X C be measurable. Then we
have H
olders inequality
Z

Z
|f g| d
for p, q (1, ) such that
Z

1/p Z

1/q

|f | d

|g| d

1 1
+ = 1, and we have Minkowskis inequality
p q

|f + g|p d

1/p

Z

|f |p d

1/p

Z
+

|g|p d

1/p

for p 1.
Z
Proof. Without loss of generality, we assume f, g 0 and
prove Holders inequality. Let
F (x) = R

f (x)
f p d

and
G(x) = R
so

g(x)
g q d

F d =

1/p

1/q ,

Gq d = 1.

We can also write


1

s(x)

t(x)

F (x) = e p
and
G(x) = e q

27

|f | d,

|g|q d > 0. We first

and we can assume s, t > (the case where either of these is is trivial). With = {p1 , p2 },
1
1
where p1 = p and p2 = q, and where 1 = and 2 = , one of the previous examples gives us
p
q
F (x)G(x) = e

t(x)
s(x)
+ q
p

1 s(x) 1 t(x) 1
1
e
+ e
= F (x)p + G(x)q .
p
q
p
q

Thus,
Z
F G d

1 1
+ = 1,
p q

from which H
olders inequality follows.
Minkowskis inequality follows from Holders inequality. Note that
(f + g)p = f (f + g)p1 + g(f + g)p1 ,
p
, we get
p1
Z
Z
Z
(f + g)p d = f (f + g)p1 d + g(f + g)p1 d

so letting q =

Z

f d
Z

1/p Z

((f + g)
1/p

f d

Z
+

p1 q

1/q

Z
+

g d

1/p ! Z

g d

1/p Z

(f + g) d

p1 q

((f + g)

1/q

p1
p,

which concludes the proof.

5.3

Lp Space

The Holder and Minkowski inequalities can be restated as


kf gk1 kf kp kgkq
and
kf + gkp kf kp + kgkp .
Thus, kf kp is a seminorm on Lp (). To make it into a norm, we need only factor out the
elements of seminorm 0. Thus, we redefine Lp via
Lp () := Lp ()/{f Lp () : kf kp = 0}.
Theorem 5.4. If fn Lp () is a Cauchy sequence in k kp , then there is a subsequence fni
such that lim fni (x) = f (x) exists -almost everywhere (for p [1, )).
i

28

Proof. Since fn is Cauchy, there is a subsequence fni such that kfni+1 fni kp <

gk (x) =

k
X

1
. Define
2i

|fni+1 (x) fni (x)|

i=1

and
g(x) =

|fni+1 (x) fni (x)|.

i=1

By Fatous lemma,
Z

lim inf gk (x) d lim inf

gk (x)p d.

By Minkowski,
kgk kpp

|gk | d

k
X

kfni+1

i=1

fni kpp

k
X
1
< 1.
<
2i
i=1

Thus,
kgkpp

Z
=

lim inf
k

gkp d

Z
lim inf
k

gkp d = lim inf kgk kpp 1,


k

so g Lp (). Moreover, kgkp 1 implies g < almost everywhere, so g converges almost


everywhere. Thus,
f (x) = fn1 (x) +

X
(fni+1 (x) fni (x)) = lim fni (x)
i

i=1

converges -almost everywhere.


Theorem 5.5. (Lp , k kp ) is complete, hence a Banach space.
Proof. Let fn be Cauchy in k kp , so there is a subsequence fni whose limit f exists almost
everywhere. Moreover, since fn is Cauchy, for any  > 0 there exists N N such that m, n N
implies kfm fn kp < . Thus,
Z
p
kfn f kp = |fn f |p d
Z
= lim |fn fni |p d
i
Z
lim inf |fn fni |p d.
i

Taking n N we get

Z
lim inf
i

|fn fni |p d < p ,

so kfn f kp 0 as n .

29

5.4

Banach Spaces

Remark. The set of simple functions with finite measure support is dense in Lp ().
Theorem 5.6. If X is a locally compact Hausdorff space, satisfies the conclusions of the Riesz
representation theorem, and p [1, ), then Cc (X) is dense in Lp ().
Proof. If s is a simple function of finite measure support and  > 0, then by Lusins theorem we
can find g Cc (X) such that |g| max |s(x)| and ({x : g(x) 6= s(x)}) < . For such g,
xX

Z
kgskp =

|g s|p d

1/p

!1/p

|g s|p d

((2 max |s(x)|)p )1/p = 1/p (2 max |s(x)|),


xX

g(x)6=s(x)

xX

so Cc (X) is dense in the space of simple functions of finite measure support. By the remark,
Cc (X) is dense in Lp ().
Theorem 5.7. (L (), k k ) is a Banach space.
Proof (Sketch). Let fn L () be a Cauchy sequence (in k k ), let Ak = {x : |fk (x)| >
kfk k }, and let
Bm,n = {x : |fm (x) fn (x)| > kfm fn k }.
Then (Ak ) = (Bm,n ) = 0, so
!

[
k

Ak

!!

Bm,n

= 0.

m,n

Moreover, outside the set being measured above, fn converges uniformly to a bounded function
f . Just define f = 0 on E.
Remark. The space Cc (X) need not be dense in L (X, ). For instance, 1 L (R, m), but
any convergent (in k k ) sequence in Cc (R) tends to a function f C(R) with lim f (x) = 0.
|x|

Similarly, in [0,1/2] L ([0, 1], m) cannot be approximated in k k by functions in


Cc ([0, 1]) = C([0, 1]). [[verify these]]
Theorem 5.8 (Baires theorem). If (X, ) is a complete metric space, then the intersection of
any countable family of dense open sets in X is dense in X.
Proof. Consider a countable collection Vn of dense open sets. Let x X and W 3 x open. Since
V1 is dense, the intersection W V1 is nonempty. Thus, since this intersection is open, it contains
the closure B(x1 , r1 ) of the ball of radius r1 about x1 for some x1 W V1 and 0 < r1 < 1.
1
Similarly, since Vn is dense, we have B(x2 , r2 ) B(x1 , r1 ) V2 for 0 < r2 < . Proceeding
2
1
inductively, we get for each n a radius 0 < rn < such that B(xn , rn ) B(xn1 , rn1 ) Vn .
n
30

That is, we get a decreasing sequence B(xn , rn ) of open balls. Hence, for all m > n, we have
xm B(xn , rn ), so
2
(xn , xm ) < 2rn < 0.
n
That is, xn forms a Cauchy sequence. Thus, xn is convergent by completeness of X and so
lim xn B(xn , rn ) Vn for all n. Therefore,
n

!
lim xn
n

Vn

W.

Since W is an arbitrary open set, this shows that the intersection of Vn is dense.
Definition. A set E is called nowhere dense in X if E has empty interior. A countable union
of nowhere dense subset is said to be of Baire first category. All other sets are of (Baire) second
category.
Theorem 5.9 (Banach-Steinhaus). If X is a Banach space, Y is a normed space, and
is a collection of bounded linear operators X Y , then either there is an M > 0 such that
k k < M for all , or sup k xk = for all x in a G dense subset of X.

Proof. Let Vn = {x X : sup k xk > n}. Since the map x 7 k xk is continuous, the map

: x 7 sup k xk must be lower semicontinuous; thus, Vn = 1 ((n, )) is open. Now if

there is an N such that VN is not dense in X, then we can find x0 X and r > 0 such that
x + x0
/ VN for all kxk r. That is, for such x,
(x + x0 ) = sup k (x + x0 )k N.

Thus, for all and all kxk r,


k xk = k (x0 + x) x0 k k (x0 + x)k + k x0 k N + N = 2N.
Hence, for all ,
2N sup k xk = sup rk xk = rk k
kxkr

kxk1

or
k k

2N
,
r

which implies that


sup k k

so

2N
is the required uniform bound M .
r

31

2N
,
r

If, on the other hand, all the Vn are dense in X then by Baires theorem so is their intersection.
Moreover, this intersection is a G by definition and
\
x
Vn x Vn , n sup k xk > n, n sup k xk = .

Theorem 5.10 (Open mapping). If U and V are the unit balls in Banach spaces X and Y ,
respectively, then for all surjective bounded linear functionals : X Y , there is a > 0 such
that U V .
Proof.

[[Incomplete]]

Corollary 5.11. If we add to the hypotheses of the open mapping theorem the requirement that
be injective (hence bijective), then 1 : Y X is a bounded linear functional.
Proof. By the open mapping theorem, there is a > 0 such that U V , so 1 (V ) U .
Thus,
sup k1 xk 1
kxk

1
and so k1 k .

Example 13. Let X = Y = `2 (N), which is L2 () where is the counting measure on N. More
precisely,

X
`2 (N) = {f : N C :
|f (n)|2 < }.
n=0

1
n , where we denote n = {n} (extend via
n+1
linearity). Then is clearly injective, kk = 1 (show this as an exercise), and (`2 (N)) {n :
n N}, so since (`2 (N)) is a vector space, it must contain the vector space generated by this
set, i.e.
(`2 (N)) hn : n Ni.
Define : `2 (N) `2 (N) by (n+1 ) =

Now the generated space (on the right-hand side) consists only of finite sequences, so it is not
necessarily the case that (`2 (N)) contains all of `2 (N).
In fact, it is not at all the case. We have 1 n = (n + 1)n , so k1 k n + 1 for all n N,
i.e. 1 is unbounded. Therefore, cannot be surjective. However, one can show as an exercise
that (`2 (N)) is dense in `2 (N).
Example 14. Let S : `2 (N) `2 (N) be defined via n 7 n+1 . Then for all f `2 (N),
2

kSf k =

(f (n 1)n (n)) =

n=0

(f (n)n (n))2 = kf k2 ,

n=0

where we use the convention f (1) = 0. However, S is not surjective, hence not invertible.
32

Before proving the next theorem, we need two ingredients. The first is the Hausdorff maximality principle, which is an equivalent statement of the axiom of choice. Let us state it as a
theorem.
Theorem 5.12 (Hausdorff maximality principle). Every nonempty partially ordered set contains
a maximal totally ordered subset.
Note that this maximal set need not be unique.
Next, we have the following lemma, which allows us to reduce the proof of the following
theorem to the case of real functionals.
Lemma 5.13. Let V be a complex vector space.
1. If f : V C is a linear functional and u = Re(f ), then f (x) = u(x) iu(ix).
2. If u : V R is (R-)linear, then f (x) = u(x) iu(ix) is C-linear.
3. If V is normed, f : V C is linear, and u = Re(f ), then kf k = kuk.
Proof.
1. Let v = Im(f ). Then
u(ix) + iv(ix) = f (ix) = if (x) = iu(x) v(x),
so v(x) = u(ix).
2. This is easy to verify.
3. Since |u(x)| |u(x) iu(ix)| = |f (x)|, we have kuk kf k. Conversely, we can write
|f (x)| = f (x) = f (x) for || = 1. Thus,
f (x) = u(x) kukkxk = kukkxk,
[[how do we get the first equality?]] so kf k kuk.

Theorem 5.14 (Hahn-Banach). If X is a normed space, M X is a linear subspace, and


f : M C is a bounded linear functional then there is a linear map F : X C such that
F |M = f and kF k = kf k.
Proof. By the proposition above, we need only consider the case where X is a real vector space
and f is real-valued (when X is complex and f is complex-valued just let u = Re(f ) and use
what follows to extend u to a function U on X with kuk = kU k; then to extend f to F on X,
let F (x) = U (x) iU (ix)). The case f = 0 is trivial, so assume otherwise. We can thus assume
33

without loss of generality that kf k = 1 (otherwise, let f 0 =

f
and extend f 0 to F 0 on x with
kf k

kF 0 k = kf 0 k = 1; then just let F = kf kF 0 ).


If X = M , were done; otherwise, take x0 X \ M . We begin by extending f to a functional
f1 on
span(M {x0 }) = {x + x0 : x M, R}
via f1 (x + x0 ) = f (x) + for some fixed R. In order for kf1 k = kf k = 1, we need only
show that
|f1 (x + x0 )| = |f (x) + | kx + x0 k
for all x M and R (so that kf1 k 1 by setting the right-hand side of the above line to 1
and the reverse inequality follows because f1 is an extension of f ). Replacing x by x in the
above inequality gives us
|f (x) + | kx + x0 k | f (x)| kx0 xk
f (x) kx0 xk f (x) + kx0 xk.
But since kf k = 1,
f (x) f (y) = f (x y) kx yk kx x0 k + kx0 yk,
so
f (x) kx0 xk f (y) kx0 yk
for all x, y M . Thus, there exists
[ sup (f (x) kx0 xk), inf (f (y) + kx0 yk)]
yM

xM

as required; that is, we can choose such that kf1 k = kf k.


What we have shown is that the set
= {(M 0 , f 0 ) : M M 0 a real subspace, f 0 : M 0 R linear, f 0 |M = f, kf 0 k = kf k = 1}
is nonempty. Thus, defining the partial order
(M 0 , f 0 ) (M 00 , f 00 ) M 0 M 00 and f 00 |M 0 = f 0 ,
we can apply the Hausdorff maximality principle to obtain a maximal totally ordered subset
[
. It is easy to verify that M0 =
M 0 is a subspace of X. Now define F0 : M0 R
(M 0 ,f 0 )

by F0 (x) = f 0 (x), where (M 0 , f 0 ) is such that x M 0 ; one may verify that F0 is a welldefined linear functional of norm 1. Moreover, we must have M0 = X because otherwise we
could take x0 X \ M0 and extend F0 to span(M {x0 }) by the method used previously; but
this would contradict the maximality of . Thus, the F we seek is simply F0 .
34

Note. The extension obtained via the Hahn-Banach theorem is not unique. For instance,
consider the space L1 ([0, 1]). There is aZnatural way of seeing L1 ([0, 1/2]) as a subspace, so
(x)gf (x) dx, where gf L ([0, 1/2]). Then

define f : L1 ([0, 1/2]) C by f () =

[0,1/2]

kf k = kgf k [[how?]]. We can extend f to F on L1 ([0, 1]) by extending gf to [0, 1] such that
the restriction of this extension to [0, 1/2] has -norm no more than kgf k on [0, 1/2]. [[need
to look at this]]
Remark.
1. If X is normed and 0 6= x0 X, then there exists f : X C such that kf k = 1 and
f (x0 ) = kx0 k. Indeed, just define f0 on the span of (x0 ) by f0 (x0 ) = kx0 k and extend
it using the Hahn-Banach theorem.
2. The dual space X of all bounded linear functionals X C separates elements of X in
the sense that for all x1 6= x2 X, there exists f X such that f (x1 ) 6= f (x2 ). To see
this, use the previous remark to construct a bounded linear functional f of norm 1 such
that f (x1 x2 ) = kx1 x2 k.
3. If x X,
kxk = sup{|f (x)| : f X , kf k 1}.
The supremum is achieved by f X with f (x) = kxk.
Example 15. There are complete metric linear spaces whose dual space is 0. They have no
convex neighbourhood of the origin. The following are some examples.
1. Lp (X, ) for 0 < p < 1.
2. Z
For X with (X) = 1, let L0 () be the space of measurable f : X C with (f, g) =
|f g|
d. [[incomplete]]
1 + |f g|

5.5

Hilbert Spaces

Proposition 5.15 (Cauchy-Schwarz inequality). If H is a Hilbert space and x, y H, then


|(x, y)| kxkkyk.
Proof. The case y = 0 is trivial, so assume this is not the case. Writing (y, x) = rei , we can let
= ei so that || = 1 and (y, x) = r = |(y, x)|. Then
0 (x cy, x cy)
= (x, x) c(y, x) c(x, y) + c2 (y, y)
= kxk2 c|(y, x)| c|(y, x)| + c2 (y, y)
= kxk2 2c|(y, x)| + c2 kyk2 ,
35

for any c R. In particular, setting c =

|(x, y)|
, we get
kyk2

0 kxk2

|(x, y)|2
kyk2

from which the result follows.


Proposition 5.16. If E H is a closed, convex set, then there exists a unique x E such that
kxk = min(kyk : y E).
Proof. It is easy to see that
kx + yk2 + kx yk2 = 2(kxk2 + kyk2 )
for any x, y H (this is known as the parallelogram identity and states that the sum of the
squares of sides of the parallelogram spanned by x and y equals the sum of the squares of its
diagonals). Let x, y E and set = inf(kyk : y E). Then


x + y 2 1
1
1
2
2
2
2
2
2

kx yk = (kxk + kyk )
2 2 (kxk + kyk ) ,
4
2
where the inequality follows because by the fact that

x+y
E. Thus,
2

kx yk2 2(kxk2 + kyk2 ) 4 2 .


So if kxk = kyk = , then kx yk2 = 0, so x = y; this shows uniqueness.
Now let yn E be a sequence with kyn k as n . Then substituting x = yn and
y = ym in the above gives
kyn ym k2 2(kyn k2 + kym k2 ) 4 2 0
as m, n . That is, the sequence yn is Cauchy, so has a limit y E = E. By our assumption
on this sequence, kyk = .
Recall the following notation.
1. We write x y whenever (x, y) = 0.
2. If M H is a nonempty set, we write x M if x y for all y M .
3. If =
6 M H is a subset [[subspace?]], then M = {x H : x M } is a closed subspace
of H.
Theorem 5.17. Let M H be a closed subspace.
1. For all x H, there exist unique P x, Qx H, where P x M , Qx M , and P x+Qx = x.
36

2. P x and Qx are nearest to x in M and M , respectively.


3. P and Q are bounded linear operators.
4. kxk2 = kP xk2 + kQxk2 .
Theorem 5.18. If H is a Hilbert space and f : H C is a bounded linear functional, then
there is a unique element xf H such that f (x) = (x, xf ) and kf k = kxf k.
Proof. The case f = 0 is trivial, so assume this is not the case. Then letting M = Ker(f ),
we have M c 6= . We must thus have for some x H \ M that the decomposition x = y + z
(y M, z M ) yields z 6= 0; otherwise, we would have M = H, implying that f = 0. That
is, M is non-trivial. Since it is a linear space, we can find z M with kzk = 1. Now for any
x H, let u = f (x)z f (z)x. Then
f (u) = f (x)f (z) f (z)f (x) = 0,
so u M , which implies that
0 = (u, z) = f (x)(z, z) f (z)(x, z) = f (x) f (z)(x, z).
That is,
f (x) = f (z)(x, z) = (x, f (z)z) = (x, xf ),
where xf = f (z)z.
This theorem is also known as the Riesz representation theorem (for Hilbert spaces).
Remark (Polarization relation). A straightforward verification shows that
(x, y) =

3
X
1
j=0

ij (x + ij y, x + ij y) =

3
X
ij
j=0

kx + ij yk2 .

[[verify it]]
5.5.1

Orthonormal Bases

Definition. A subset {x : A} (we may simply denote this by x ) of a Hilbert space H is


called an orthonormal system or ONS if (x , x ) = . An orthonormal basis or ONB is an
ONS x such that span(x ) (i.e. the set of all finite linear combinations of the x ) is dense in
H.
Remark. Elements of an ONS x are linearly independent. If c1 x1 + + cn xn = 0, then
for each j = 1, . . . , n,
cj = (c1 x1 + + cn xn , xj ) = (0, xj ) = 0.
37

Definition. If x ( A) is an ONS, then for each x H, we define x


: A C by x
() =
(x, x ). We may view as a mapping from H to the set of functions A C.
Remark.
1. If x span(x : A), then x =

(x, x )x =

2. If : A C has finite support, then y =

x
()x , where this sum is finite.

()x satisfies () = y().

Note. Let F A be finite and x H. We have


!
X\
(),
x
()x () = x
F

so

!
x

x
()x

for all F . Thus, for all s span(x : F ),


!
!
X
X
x
x
()x
x
()x s .
F

Therefore,

2


X
X


kx sk2 = x
x
()x +
x
()x s


F
F

2
2


X

X




= x
x
()x +
x
()x s .




F

It follows that

(2)





X


x

()x

kx sk


F

for all s span(x : F ) and this holds with equality if and only if s =

X
F

Remark. Taking s = 0 in (2), we get



2

X
X


|
x()|2 .
x
()x =
kxk2


F

This is known as Bessels inequality.

38

x
() = s.

We wish to generalize the above to the case of (arbitrary) infinite index sets. First, let us
say what we mean by a sum over such a set. We define
X
X
c = sup
c ,
A

where the supremum is over all finite subsets F A. In fact, this is precisely the Lebesgue
integral of the map 7 c with respect to counting measure on A. Note that if such a sum is
finite, then the set of for which c 6= 0 is at most countable.
Theorem 5.19. If x is an ONS, then
X

|
x()| kxk2

for all x H.
Remark. The mapping
U : H `2 (A)
x 7 x

is a contraction:
!1/2
kU x U yk = k
x yk =

|
x( y()|

!1/2
=

|x[
y()|

kx yk.

If x span(x : A), then kU xk = kxk, i.e. U is an isometry on span(x : A).


Theorem 5.20. Let x be an ONS. The following are equivalent:
1. {x : A} is a maximal ONS;
2. span(x : A) is dense in H (i.e. x is an ONB);
X
3.
|
x()|2 = kxk2 for all x H;
A

4.

x
()
y () = (x, y) for all x, y H.

Proof.

[[Incomplete!]]

The last part of this theorem is Parsevals identity.


Corollary 5.21. In any Hilbert space, one can find an orthonormal basis.
Proof.

[[Incomplete]]
39

5.5.2

Trigonometric Series

Let T = {z C : |z| = 1} = D be the 1-torus, where D = {z C : |z| < 1} is the (open) unit
disc. Note that for any z T, there is a unique t [, ) such that z = eit . In this way, we
get a probability measure
Z
1
1
(t) dt
m(a) =
2 i log A
on T. There is also a natural correspondence between 2-period functions f : R C and
functions F : T C via F (eit ) = f (t) (check that continuity is preserved as an exercise).
Let us define
un (t) = eint = cos(nt) + i sin(nt)
for t < and n Z. Note the following basic properties of this family of functions:
1. un = un (closure under conjugation);
2. u0 = 1;
3. if s 6= t, then u1 (s) 6= u1 (t).
Thus, span(un : n Z) is an algebra containing 1 that is closed under conjugation and separates
points of T. Note the following theorem.
Theorem 5.22 (Stone-Weierstrass). Let X be a compact Hausdorff space and let S be a set of
continuous functions X C. If span(S) contains 1, is closed under conjugation, and separates
points of X, then span(S) is dense in C(X) (in k k ).
In particular, we get the following.
Theorem 5.23. span(un : n Z) = C(T) (where the closure is in k k ).
By the correspondence between T and (, ], we can write


Z
1
|f (t)|2 dt = kf k22 < .
L2 (T) = f : (, ] C :
2
Recall that for all p [1, ), C(T) = Lp (T) (in k kp ). Now in a probability space, we have
k kp k k , so
span(un : n Z) = Lp (T)
in k kp for 1 p < (the case p = 2 is of particular interest).
Even more,
Z
Z
1
1
kun k22
|un (t)|2 dt =
dt = 1
2
2
and

Z
Z

1
1
1
1
int imt
i(nm)t
i(nm)t
(un , um ) =
e e
dt =
e
dt =
e
= 0,
2
2
2 i(n m)

for n 6= m, so {un : n Z} is an orthonormal basis in L2 (T).


40

Definition. Polynomials in the un are called trigonometric polynomials.


Definition. We define the Hardy space


Z
1
it 2
2
n
|f (re )| dt < ,
H (T) = span(z : n 0) = f : T C : f analytic, lim
r1 2pi
where the closure is with respect to k k2 .
For any f L2 (T), we have f : Z C such that f (t) =

f(n)eint for t (, ] (or

nZ

f (z) =

f(n)z n for z T). Now f H 2 (T) if and only if f(n) = 0 for all n < 0. So the

nZ

embedding H 2 (T) , L2 (T) is in natural correspondence with l2 (N) , l2 (Z).


Remark. There are infinite sets K C for which the set of polynomials on K is dense in C(K).
The unit circle T is not such a set.
Define Mz : L2 (T) L2 (T) by (Mz f )(z0 ) = z0 f (z0 ) for z0 T (where z denotes the mapping
z0 7 z0 ). This operator corresponds to a map l2 (Z) l2 (Z). Which one? It is easy to see that
!
X
X
Mz
f(n)z n =
f(n)z n+1 .
nZ

nZ

So the corresponding map on l2 (Z) is the (left) shit (by one unit). To make the details of this
correspondence more!precise, let U : L2 (T) l2 (Z) and W : l2 (Z) l2 (Z) be given by U (f ) = f
X
X
and W
f(n)en =
f(n 1)en , where en = z n . Then the following diagram commutes:
nZ

nZ

L2 (T)
Mz

L2 (T)

/ l2 (Z) .

/ l2 (Z)

Upon restricting to H 2 (T), we get the following diagram instead:


H 2 (T)
Mz

H 2 (T)

/ l2 (N) ,


/ l2 (N)

where S = W |l2 (N) . Note that while W is invertible, S is not.


Remark. The space H 2 (T) is isomorphic (as a Hilbert space) to l2 (N), which is in turn isomorphic to l2 (Z) (via any bijection between N and Z [[check this]]), which itself is isomorphic
to L2 (T). So we can move W to H 2 (T) and S to L2 (T) via
V 1 W V : H 2 (T) H 2 (T)
41

and
V SV 1 : L2 (T) L2 (T).
Functional Calculus
Note that
(Mz )2 f = Mz (Mz f )(z) = z 2 f (z) = Mz 2 f (z).
Similarly, we can see that for a polynomial P C[x], say P (x) = a0 + a1 + + an xn , we have
P (Mz ) = a0 f + a1 Mz f + + an (Mz )n f
= a0 f + a1 zf + + an z n f
= (a0 + a1 z + + an z n )f
= MP (z) f.
Now for C(T), we know there exists a sequence of trigonometric polynomials Pn (z, z) (where
z = eint T) converging in k k to (z). We claim that for such Pn we have MPn M in
k k . To see this, write
Z
2
2k(MPn M )f k2 =
|(Pn )f |2 dz
T
Z
2
|f |2 dz
sup |Pn (z) (z)|
zT
T
Z
2
= 2kPn k |f |2 dz
T

= (2)2 kPn k2 kf k22 .


Then the last line converges to 0 as n .
Remark. If we replace L2 by H 2 above, then we must require that Pn uniformly on D.
5.5.3

The Poisson Integral

Recall the maximum modulus principle for analytic functions, which states that if f : D C is
analytic, then sup |f (z)| = sup |f (z)| for all r (0, 1). In particular, if f C(D) is analytic
zD

zD\rD

on D, then sup |f (z)| = max |f (z)|.


|z|=1

zD

Remark. If f : D C is continuous and f (V ) is open for all open V D (i.e. f is an open


mapping), then sup |f (z)| = max |f (z)|, so this property is not restricted to analytic functions.
zD

|z|=1

Let V be a vector space of functions in C(D) satisfying this property, i.e. if f V , then
max |f (z)| = max |f (z)|.
zT

zD

42

Then for f, g V , if f |T = g|T , then (f g)|T = 0, so we must have f g = 0, i.e. f = g. That


is, functions in V are determined by their values on T. We will find an explicit formula for f (z)
for z D when f |T is known.
For any z C, the map z : f 7 f (z) C is a bounded linear functional (by assumption, f
achieves its maximum in D on T, so this map actually has norm kz k = 1). Thus, by the Riesz
representation theorem, we can write
Z
1
f (t) dz (t)
z (f ) =
2 T
for some measure z .
Next we assume the functions gn (z) = z n (or gn (t) = eint ) on D are members of V (the gn
do satisfy the required property of elements of V by the maximum modulus principle). Then
we have
Z
n in
n
gn (t) dz (t)
r e = z = z (gn ) =
T

for n 0. On the other hand, define the Poisson kernel Pr ( t) =

r|n| ein(t) , which is

n=

dominated by the geometric series


Z

|n|

r . So by the dominated convergence theorem,


!
Z

X
|m| im(t)
eint dt
Pr ( t)gn (t) dt =
r e

m=

r|m|

m im

|r| e

ei(nm)t dt

ei(nt+mmt) dt

m=

|r|m eim (2mn )

= 2r|n| ein
= 2z (gn )
for all n. It follows that for any trigonometric polynomial Q,
Z
1
Q(eit )Pr ( t) dt = z (Q).
2
Thus, for any uniform limit f of trigonometric polynomials,
Z
1
f (z) = z (f ) =
f (eit )Pr ( t) dt.
2
Remark. Notice that
Pr ( t) = 1 +

X
n=1

n in int

r e

n in int

n=1

=1+

X
n=1

43

n in int

r e

X
n=1

rn ein eint .

Since the last two summations are easily seen to be complex conjugates of each other, Pr ( t)
is the real part of



X
1

1
1+2
(zeit )m = 1 + 2
zeit
n=1

2
1
1 zeit
1 + zeit
=
1 zeit
eit + z
= it
e z
1 r2 + 2ir sin( t)
=
.
|1 zeit |2
=

[[easy to way get last equality?]] That is,


Pr ( t) =

1 r2
.
1 2r cos( t) + r2

The following theorem summarizes what we have shown.


Theorem 5.24. Let V be a subspace of C(D) containing gn (t) = z n for all n and satisfying the
following property: if f V , then
max |f (z)| = max |f (z)|.
zT

zD

Then
1
f (z) =
2

1 r2
f (eit ) dt
1 2r cos( t) + r2

for any f V and z = rei D.


Note. The above formula will not work for r = 1 even though its purpose is to give the values
of f for r < 1 based on the values of f for r = 1.
The above should be compared with the Cauchy integral formula
Z
f ()
1
f (z) =
d,
2i D z
where |z| < 1.

Fourier Series in L1 (T)

Note that for f L1 (T), the function f (t)eint is clearly absolutely integrable (since |eint | = 1),
hence integrable. We define
Z
1
f(n) =
f (t)eint dt.
2
44


Remark. Let  > 0. Then there exists g C(T) with kf gk1 <
and a trigonometric
2

polynomial P C(T) with kg P k < . Then
2
Z

1

int


|f (n)| =
(f g + g P + P )e
dt
2
Z
Z
Z
1
1
1

|f g| dt +
|g P | dt +
P (t)eint dt
2
2
2

1 

<
+ +0
2 2 2

=
2
for n > deg(P ) [[really?]], which is required to get the line with the strict inequality. Therefore,
lim |f(n)| = 0.

|n|

It can be shown that the converse is not true in general (i.e. given a sequence an C with
an 0 there may not be an f L1 (T) with f(n) = an ).

Now if f L (T) we know that

f(n)eint = f (t) in k k2 [[we do?]]. Let us sketch a

n=

proof that this does not hold in k k for f C(T) [[check everything below]]. For each eix T,
define s(, x) : C(T) C by
Z
1
sn (f, x) =
f (t x)Dn (t) dt,
2
where Dn (t) =

n
X

eijt . Clearly, ksn (, x)k 2n + 1. Moreover,

j=n

1
|sn (f, x)|
kf k
2

|Dn (t)| dt.

We claim that kDn k1 as n . Indeed, Dn (t) =


Z

Z
|Dn (t)| dt > 2



sin(n + t/2)

dt


t/2

(n+1/2)

=4
>4

0
n
X

sin(n + t/2)
, so
sin(t/2)

1
k

k=1
n
X

k=1

|sin(s)|
ds
s

(where s = n + t/2)

| sin(s)| ds
(k1)

1
k

([[check]]).

45

Now for each n,


1
kDn k1 =
2
where sgn(x) =

(
1
1

x>0
x<0

|Dn (t)| dt =

1
Dn (t) sgn(Dn (t)) dt,
2

. We can approximate sgn( Dn (t)) by continuous functions to get


ksn (, x)k = kDn k1 .

Since sn (, x) C(T) , it follows by Banach-Steinhaus that


sup |sn (f, x)| =
n

for all f in a dense G subset of C(T). Thus,

einx f(n) (not f (x)).

Integration on Product Spaces

Let (X, S) and (Y, T ) be measurable spaces. We will construct a -algebra that behaves nicely
under projection.
Definition. If A S and B T , we will refer to A B X Y as an elementary rectangle.
Any finite disjoint union of elementary rectangles is called an elementary set. We let E denote
the collection of all elementary sets and define S T = (E).
Note.

1. The in S T is merely notational and does not denote a Cartesian product.

2. The -algebra S T is sometimes denoted S T . The reason for this is that actually
L (X Y, S T , ) = L (X, S, )L (Y, T , ),
where will be introduced below and is defined appropriately.
Definition. A monotone class on a set X is a family M of subsets of X such that if Ai M
[
is an increasing (respectively, decreasing) family of sets in M, then
Ai M (respectively,
i
\
Ai M).
i

Definition. If E X Y and (x0 , y0 ) X Y , we denote Ex0 = {y Y : (x0 , y) E} and


E y0 = {x X : (x, y0 ) E}.
Theorem 7.1. If E S T , then Ex T for all x X (and by symmetry, E y S for all
y Y ).

46

Proof. Let = {E S T : Ex T , x X} be the collection of sets for which this holds.


Since X Y , it is nonempty. Now suppose E . It is easy to see that (Ex )c = (E c )x , so
since T is a -algebra,
E c . If we have Ei for i N, then it can similarly be seen that
!
[
[
[
(Ei )x =
Ei . Thus, is a -algebra. It is easy to see that
Ei , so again we get
i

contains the elementary sets A B E, so S T . But the reverse inclusion follows by


the definition of , so = S T as required.
Theorem 7.2. The -algebra S T is the smallest monotone class containing E.
Proof. Let M be the smallest monotone class containing E (that such M exists follows from the
fact that an arbitrary intersection of monotone classes is itself a monotone class, as is easy to
verify). Then since -algebras are monotone classes, M S T . It is easy to see that
(A1 B1 ) (A2 B2 ) = (A1 A2 ) (B1 B2 )
and
(A1 B1 ) \ (A2 B2 ) = ((A1 \ A2 ) B1 ) ((A1 A2 ) (B1 \ B2 ))
for any A1 , A2 S and B1 , B2 T . Thus, E is closed under finite intersections and set
differences. Thus, writing
(A1 B1 ) (A2 B2 )
as
((A1 \A2 )B1 )((A1 B1 )(A2 B2 ))((A1 A2 )B1 )((A2 \A1 )B2 )((A2 A1 )(B2 \B1 )),
we see that E is also closed under finite unions.
Now for any P X Y , define (P ) = {Q X Y : P \ Q, Q\, P Q M}. Notice that
1. Q (P ) if and only if P (Q) (by symmetry) and
2. (P ) is a monotone class (this follows from the fact that M is a monotone class).
Now for P E, we have Q (P ) for all Q E by the properties of E shown above. That
is, E (P ). It follows by minimality of M and the fact that (P ) is a monotone class that
M (P ). Next, take Q M (P ), so that P (Q). That is, E (Q), so as before,
M (Q).
Now by definition of (P ) and (Q), we conclude that if P, Q M, then P \ Q, P Q M.
On the other hand, if we have Rn M for n N, then we know that Sk = R1 . . . Rk M
for all k N. Since Sk is increasing in k, it follows by the fact that M is a monotone class that
[
[
Rn =
Sk M. Therefore, M is a -algebra. So by minimality of S T as a -algebra,
n

we have M S T . But as we showed, S T is a monotone class, so by minimality of M as a


monotone class, M = S T .
47

Definition. If W is a set and f : X Y W , we define


fx : Y W
y 7 f (x, y)
and
fy : X W
x 7 f (x, y)
for any x X and y Y .
Theorem 7.3. If f : X Y W is measurable (so W has a topology now), then
1. fx is T -measurable for all x X and
2. fy is S-measurable for all y Y .
Proof. If V W is open, then Q = f 1 (V ) S T is open by measurability of f . Since
Qx = fx1 (V ), measurability of fx follows from the previous theorem (and similarly for fy ).
Theorem 7.4. Let (X, S, ) and (Y, T , ) be -finite measure spaces and fix Q S T . If
(x) = (Qx ) and (y) = (Qy ), then and are measurable and
Z
Z
(x) d(x) =
(y) d(y).
X

Proof. Let us show that = {Q S T : the above holds} is a monotone class containing E;
it then follows from Theorem 7.2 that S T , which is all we need to show. Begin by taking
Q to be an elementary rectangle A B S T . Then
(x) = A (x)(B)
and
(y) = B (y)(A),
so these are both measurable. Moreover,
Z
Z
(x) d(x) = (A)(B) = (y) d(y).
It easily follows that this holds when Q is an elementary set in E. Therefore, E .
Next, we will show is a monotone class. Start by taking an increasing sequence of Qi
. Then it is easy to see that (Qi )x is also increasing in i, so i (x) i+1 (x) for all x by
monotonicity. Thus, we can apply the monotone convergence theorem to get
Z
Z
Z
Z
(x) d(x) i (x) d(x) = i (y) d(y) (y) d(y),
48

so

Qi .

Before we proceed to showing the analogous result for decreasing sequences of sets, let us
show that is closed under countable disjoint unions. Begin by taking Q1 , Q2 disjoint.
Then
(x) = ((Q1 Q2 )x ) = ((Q1 )x (Q2 )x ) = ((Q1 )x ) + ((Q2 )x ) = 1 (x) + 2 (x)
and a similar equality holds for . It thus follows by closure under increasing unions (shown
above) that this holds for countable disjoint collections.
Finally, take a decreasing sequence Qi . Let us first assume that Q1 A B for some
A S and B T with (A), (B) < . Then ((Qj )x ) A (x)(B) for any j, so by the
dominated converge theorem,
Z
Z
j (x) d(x) (x) d(x)
as j , where (x) =

\

Qj

 
x

. A similar result holds for . It follows that

Qj .

Now for any Q, let Qmn = Q (Xm Yn ), where Xm and Yn are finite measure partitions
of X and Y , respectively (these exist by -finiteness). Let M = {Q S T : Qmn , m, n}.
Then by what was shown above, M is a monotone class containing E, so M S T . Therefore,
M=S T.
Definition. We define the product measure on S T by
Z
Z
( )(Q) =
(x) d(x) =
(y) d(y)
X

for any Q S T , where (x) = (Qx ) and (y) = (Qy )


Corollary 7.5 (Fubinis theorem). Let f : X Y C be S T -measurable.
Z
Z
1. If 0 f , then the functions (x) =
fx (y) d(y) and (y) =
f y (x) d(x) are
measurable and
Z

2. If (x) =

Z
f (x, y)d( )(x, y) =

d =
XY

(y) d(y).

(3)

|f |x (y) d(y) L1 (), then f L1 ( ) and (3) holds.

Y
1

3. If f L ( ), then fx L1 () for almost every x (and by symmetry, fy L1 () for


almost every y).
Proof.
1. We have shown this to hold above in the case of f = Q . It follows that this holds
by linearity for f a simple function; we thus get the result by the monotone convergence
theorem.
49

2. Just apply the previous part to |f |.


3. Assume without loss of generality that f takes values in R (otherwise, take real and
imaginary parts). Writing f = f + f , it is easy to see that fx = fx+ fx and |f |x =
fx+ + fx . Thus,
Z
1
f L ( )
|f | d( ) <
XY
Z
Z
f d( ) <
f + d( ) +

XY

XY
1

L ( ).

Let

Z
1 (x) =

fx+ d

and

Z
2 (x) =

fx d

and define 1 (y), 2 (y) similarly. Then by the first part,


Z
Z
1 d =
f + d( ) < ,
X

XY

so 1 L1 (). Similarly, 2 L1 (). Hence, 1 , 2 < almost everywhere, which means


that fx L1 () for almost every x X.

Convolution

Let G be a locally compact topological group.


Theorem 8.1. There exists a unique (up to a multiplicative non-zero constant) non-zero positive
Borel measure m on G that is invariant under translation, i.e. m(xV ) = m(V ) for all x G
and V Borel measurable. Moreover, m(K) < if and only if K is compact.
Definition. The unique measure in the above theorem is called the Haar measure on G.
Example 16. The following are some familiar instances of the Haar measure.
1. Lebesgue measure on Rk .
2. Counting measure on Zk (or any discrete group).
3. (Normalized) Lebesgue measure (i.e. arc length) on Tn .

50

Definition.
If f, g L1 (G, m), we define the convolution f g of f and g by (f g)(x) =
Z
f (x y)g(y) dm(y).
G

Theorem 8.2. f g = g f if and only if G is abelian.


From now on, let G be abelian.
Theorem 8.3. If f, g L1 (G, m), then (f g)(x) L1 (G) and kf gk1 kf k1 kgk1 .
Proof. Let F (x, y) = f (x y)g(y). Since f, g L1 and since the group operation the map
(x, y) 7 y are continuous, F is measurable. Thus, by Fubinis theorem,
Z
Z
Z
|F (x, y)|d(m m)(x, y) =
|g(y)|
|f (x y)| dm(x)dm(y)
GG
G
G
Z
Z
|g(y)|
|f (x)|dm(x)dm(y)
(by translation invariance)
=
G

= kf k1 kgk1 .
The result easily follows from this equality.
[[compare convolution with Poisson integral]]

51

You might also like