An Introduction To Functional Analysis

An Introduction to Functional Analysis
Laurent W. Marcoux
Department of Pure Mathematics
University of Waterloo
Waterloo, Ontario Canada N2L 3G1
December 14, 2010
Preface to the Second Edition - December 14, 2010
This set of notes has now undergone its second incarnation. I have corrected as
many typos as I have found so far, and in future instalments I will continue to add
comments and to modify the appendices where appropriate. The course number for
the Functional Analysis course at Waterloo has now changed to PMath 753, in case
anyone is checking.
The comment in the preface to the rst edition regarding caution and buzz
saws is still `a propos. Nevertheless, I maintain that this set of notes is worth at least
twice the price
1
that Im charging for them.
For the sake of reference: excluding the material in the appendices, and allowing
for the students to study the last section on topology themselves, one should be able
to cover the material in these notes in one term, which at Waterloo consists of 36
fty-minute lectures.
My thanks to Xiao Jiang and Ian Hincks for catching a number of typos that I
missed in the second revision.
Preface to the First Edition - December 1, 2008
The following is a set of class notes for the PMath 453/653 course I taught at the
University of Waterloo in 2008. As mentioned on the front page, they are a work in
progress, and - this being the rst edition - they are replete with typos. A student
should approach these notes with the same caution he or she would approach buzz
saws; they can be very useful, but you should be thinking the whole time you have
them in your hands. Enjoy.
I would like to thank Paul Skoufranis for having pointed out to me an embar-
rassing number of typos. I am glad to report that he still has both hands and all of
his ngers.
1
If you were charged a single penny for the electronic version of these notes, you were robbed.
You can get them for free from my website.
i
ii
The reviews are in!
From the moment I picked your book up until I laid it down I was
convulsed with laughter. Someday I intend reading it.
Groucho Marx
This is not a novel to be tossed aside lightly. It should be thrown with
great force.
Dorothy Parker
The covers of this book are too far apart.
Ambrose Bierce
I read part of it all the way through.
Samuel Goldwyn
Reading this book is like waiting for the rst shoe to drop.
Ralph Novak
Thank you for sending me a copy of your book. Ill waste no time
reading it.
Moses Hadas
Sometimes you just have to stop writing. Even before you begin.
Stanislaw J. Lec
Contents
i
1. Normed Linear Spaces 1
2. An introduction to operators 17
3. Hilbert space 31
4. Topological Vector Spaces 44
5. Seminorms and locally convex spaces 55
6. The Hahn-Banach theorem 70
7. Weak topologies and dual spaces 89
8. Local compactness and extremal points 104
9. The chapter of named theorems 111
10. Operator Theory 116
11. Appendix topological background 127
Bibliography 133
Index 135
iii
1. NORMED LINEAR SPACES 1
1. Normed Linear Spaces
I dont like country music, but I dont mean to denigrate those who do.
And for the people who like country music, denigrate means put down .
Bob Newhart
1.1. It is expected that the student of this course will have already seen the
notions of a normed linear space and of a Banach space. We shall review the
denitions of these spaces, as well as some of their fundamental properties. In both
cases, the underlying structure is that of a vector space. For our purposes, these
vector spaces will be over the eld K, where K = R or K = C.
1.2. Denition. Let X be a vector space over K. A seminorm on X is a map
: X R
satisfying
(i) (x) 0 for all x X;
(ii) (x) = [[ (x) for all x X, K; and
(iii) (x +y) (x) +(y) for all x, y X.
If satises the extra condition:
(iv) (x) = 0 if and only if x = 0,
then we say that is a norm, and we usually denote () by | |. In this case, we
say that (X, | |) (or, with a mild abuse of nomenclature, X) is a normed linear
space.
1.3. A norm on X is a generalisation of the absolute value function on K. Of
course, equipped with the absolute value function on K, one immediately denes a
metric d : KK R by setting d(x, y) = [x y[.
In exactly the same way, the norm | | on a normed linear space X induces a
metric
d : X X R
(x, y) |x y|.
The norm topology on (X, | |) is the topology induced by this metric. For each
x X, a neighbourhood base for this topology is given by
B
x
= D
(x) : > 0,
where D
(x) = y X : d(y, x) < . We say that the normed linear space (X, | |)
(or informally X) is complete if the corresponding metric space (X, d) is complete.
2 L.W. Marcoux Functional Analysis
1.4. Example. Dene
c
K
00
(N) = (x
n
)
n=1
: x
n
K, n 1, x
n
= 0 for all but nitely many n 1.
For x = (x
n
)
n
c
K
00
(N), set |x|
= sup
n1
[x
n
[. Then (c
K
00
(N), | |
) is a
normed linear space. It is not, however, complete.
The space
c
K
0
(N) = (x
n
)
n=1
: x
n
K, n 1, lim
n
x
n
= 0,
equipped with the same norm |x|
= sup
n1
[x
n
[ does dene a complete normed
linear space.
1.5. Example. Consider
T
K
([0, 1]) = p = p
0
+p
1
z +p
2
z
2
+ +p
n
z
n
: n 1, p
i
K, 0 i n.
Then
|p|
= sup[p(z)[ : z [0, 1]
denes a norm on T
K
([0, 1]).
If we select x
0
[0, 1] arbitrarily, then it is straightforward to check that (f) :=
[f(x
0
)[ denes a seminorm on T
K
([0, 1]) which is not a norm.
1.6. Example. Let n 1 be an integer. If 1 p < is a real number, then
|(x
1
, x
2
, ..., x
n
)|
p
=
_
n
k=1
[x
k
[
p
_
1/p
denes a norm on K
n
, called the p-norm. We also write
p
n
for (K, | |
p
), when the
underlying eld K is understood. We also dene
|(x
1
, x
2
, ..., x
n
)|
= max([x
1
[, [x
2
[, ..., [x
n
[).
Again (K
n
, | |
) is a normed linear space. Again, we abbreviate this to
n
when
K is understood.
1.7. Example. For 1 p < , we dene
p
K
(N) = (x
n
)
n=1
: x
n
K, n 1 and
n=1
[x
n
[
p
< .
For (x
n
)
n=1

p
K
(N), we set
|(x
n
)
n
|
p
=
_

n=1
[x
n
[
p
_
1/p
.
Then | |
p
denes a norm, again called the p-norm. on
p
K
(N).
As above, we may also dene
K
(N) = (x
n
)
n=1
: x
n
K, n 1, sup
n
[x
n
[ < .
The -norm on
K
(N) is given by
|(x
n
)
n
|
= sup
n
[x
n
[.
In most contexts, the underlying eld K is understood, and we shall write only
p
(N), or even
p
, 1 p .
The last two examples have one especially nice property not shared by c
K
00
(N)
and T
K
([0, 1]), namely: they are complete.
1.8. Denition. A Banach space is a complete normed linear space.
1.9. Example. Let (([0, 1], K) = f : [0, 1] K : f is continuous, equipped
with the uniform norm
|f|
= max[f(z)[ : z [0, 1].

Then ((([0, 1], K), | |
) is a Banach space.
1.10. Denition. Let H be an inner product space over K; that is, there exists
a map
, ) K
which, for all x, x
1
, x
2
, y H and K, satises:
(i) x
1
+x
2
, y) = x
1
, y) +x
2
, y);
(ii) x, y) = y, x);
(iii) x, y) = x, y);
(iv) x, x) 0, with equality holding if and only if x = 0.
(Of course, when K = R, the complex conjugation in (ii) is superuous.) Recall
that the canonical norm on H induced by the inner product is given by
|x| = x, x)
1/2
.
If H is complete with respect to the corresponding metric, then we say that H is a
Hilbert space. Thus every Hilbert space is a Banach space.
1.11. Example. Recall that
2
K
(N) is a Hilbert space with inner product
(x
n
)
n
, (y
n
)
n
) =
n=1
x
n
y
n
.
More generally, let (X, ) be a measure space. Then H = L
2
(X, ) is a Hilbert
space with
f, g) =
_
X
fgd.
1.12. It is easy to see that if X is a normed linear space, then the vector space
operations
: X X X
(x, y) x +y
and
: KX X
(, x) x
of addition and scalar multiplication are continuous (from the respective product
topologies on X X and on K X to the norm topology on X). The proof is left
as an exercise for the reader. In particular, therefore, if 0 ,= K, y X, then
y
: X X dened by
y
(x) = x + y and
: X X dened by
(x) = x are
homeomorphisms.
As a simple corollary to this fact, a set G X is open (resp. closed) if and only
if G + y is open (resp. closed) for all y X, and G is open (resp. closed) for all
0 ,= K. We shall return to this in a later section.
1.13. New Banach spaces from old. We now exhibit a few constructions
which allow us to produce new Banach spaces from simpler building blocks.
Let (X
n
, | |
n
)
n=1
denote a countable family of Banach spaces. Let X =

n
X
n
.
(a) For each 1 p < , dene
n=1
p
X
n
= (x
n
)
n
X : |(x
n
)
n
|
p
=
_

n=1
|x
n
|
p
n
_
1/p
< .
Then

n=1
p
X
n
is a Banach space, referred to as the
p
-direct sum of
the (X
n
)
n
.
(b) With p = ,
n=1
X
n
= (x
n
)
n
X : |(x
n
)
n
|
= sup
n1
|x
n
|
n
< .
Again,

n=1
X
n
is a Banach space - namely the
-direct sum of the

(X
n
)
n
.
(c) We also dene
c
0
(X) = (x
n
)
n
X : x
n
X
n
, n 1 and lim
n
|x
n
|
n
= 0.
The norm on c
0
(X) is |(x
n
)
n
|
= sup
n1
|x
n
|
n
, and equipped with this
norm, c
0
(X) is easily seen to be a closed subspace of

n=1
X
n
.
1.14. Denition. Let X be a vector space equipped with two norms | | and
[[[ [[[. We say that these norms are equivalent if there exist constants
1
,
2
> 0
so that
1
|x| [[[x[[[
2
|x| for all x X.
We remark that when this is the case,
1
2
[[[x[[[ |x|
1
1
[[[x[[[,
resolving the apparent lack of symmetry in the denition of equivalence of norms.
1.15. Example. Fix n 1 an integer, and let X = C
n
. For x = (x
1
, x
2
, ..., x
n
)
X,
|x|
1
=
n
k=1
[x
k
[
n
k=1
_
max
j
[x
j
[
_
=
n
k=1
|x|
= n|x|
.
Moreover,
|x|
= max
j
[x
j
[
n
k=1
[x
k
[ = |x|
1
,
so that
|x|
|x|
1
n|x|
.
This proves that | |
1
and | |
are equivalent norms on X. As we shall later see,

all norms on a nite dimensional vector space are equivalent.
1.16. Example. Let X = (([0, 1], C), and consider the norms
|f|
= sup[f(x)[ : x [0, 1]
and
|f|
1
=
_
1
0
[f(x)[dx
on X. If, for each n 1, we set f
n
to be the function f
n
(x) = x
n
, then |f
n
|
= 1,
while |f
n
|
1
=
_
1
0
x
n
dx =
1
n+1
. Clearly | |
1
and | |
are inequivalent norms on

X.
1.17. Proposition. Two norms || and [[[[[[ on a vector space X are equivalent
if and only if they generate the same metric topologies.
Proof. Suppose rst that | | and [[[ [[[ are equivalent, say
1
|x| [[[x[[[
2
|x|
for all x X, where
1
,
2
> 0 are constants. If x X and (x
n
)
n
is a sequence in
X, then it immediately follows that
lim
n
|x
n
x| = 0 if and only if lim
n
[[[x
n
x[[[ = 0.
That is, the two notions of convergence coincide, and thus the topologies are equal.
Conversely, suppose that the metric topologies
and
||||||
, induced by | |
and [[[ [[[ respectively, coincide. Then G = x X : |x| < 1 is an open nbhd of 0
in (X, [[[ [[[), and so there exists > 0 so that H = x X : [[[x[[[ < G. That
is, [[[x[[[ < implies |x| < 1. In particular, therefore, [[[x[[[ /2 implies |x| 1,
so that that [[y[[ (2/)[[[y[[[ for all y X. By symmetry, there exists a constant
2
> 0 so that [[[y[[[
2
|y| for all y X.
Thus | | and [[[ [[[ are equivalent norms.
2
1.18. Corollary. Equivalence of norms is an equivalence relation for norms on
a vector space X.
1.19. Denition. Let (X, | |) be a normed linear space. A series

n=1
x
n
in
X is said to be absolutely summable if

n=1
|x
n
| < .
The following result provides a very practical tool when trying to decide whether
or not a given normed linear space is complete. We remark that the second half of
the proof uses the standard fact that if (y
n
)
n
is a Cauchy sequence in a metric space
(Y, d), and if (y
n
)
n
admits a convergent subsequence with limit y
0
, then the original
sequence (y
n
)
n
converges to y
0
as well.
1.20. Proposition. Let (X, | |) be a normed linear space. The following
statements are equivalent:
(a) X is complete, and hence X is a Banach space.
(b) Every absolutely summable series in X is summable.
Proof.
(a) implies (b): Suppose that X is complete, and that

x
n
is absolutely
summable. For each k 1, let y
k
=

k
n=1
x
n
. Given > 0, we can nd
N > 0 so that m N implies

n=m
|x
n
| < . If k m N, then
|y
k
y
m
| = |
k
n=m+1
x
n
|
n=m+1
|x
n
|
n=m+1
|x
n
|
< ,
so that (y
k
)
k
is Cauchy in X. Since X is complete, y = lim
k
y
k
=
lim
k
k
n=1
x
n
=

n=1
x
n
exists, i.e.

n=1
x
n
is summable.
(b) implies (a): Next suppose that every absolutely summable series in X is
summable, and let (y
j
)
j
be a Cauchy sequence in X. For each n 1 there
exists N
n
> 0 so that k, m N
n
implies |y
k
y
m
| < 1/2
n+1
. Let x
1
= y
N
1
and for n 2, let x
n
= y
Nn
y
N
n1
. Then |x
n
| < 1/2
n
for all n 2, so
that
n=1
|x
n
| |x
1
| +
n=2
1
2
n
|x
1
| +
1
2
< .
By hypothesis, y =

n=1
x
n
= lim
k
k
n=1
x
n
exists. But
k
n=1
x
n
= y
N
k
, so that lim
k
y
N
k
= y X. Recalling that (y
j
)
j
was
Cauchy, we conclude from the remark preceding the Proposition that (y
j
)
j
also converges to y. Since every Cauchy sequence in X converges, X is
complete.
2
1.21. Theorem. Let (X, | |) be a normed linear space, and let M X be a
linear manifold. Then
p(x +M) := inf|x +m| : m M
denes a seminorm on the quotient space X/M.
This formula denes a norm on X/M if and only if M is closed.
Proof. First observe that the function p is well-dened; for if x+M= y +M, then
x y M and so
p(y) = inf|y +m| : m M
= inf|y +m+ (x y)| = |x +m| : m M
= p(x).
Clearly p(x + M) 0 for all x + M X/M. If 0 ,= k K, then m M if and
only if
1
k
m M and so
p(k(x +M)) = p(kx +M)
= inf|kx +m| : m M
= inf|k(x +
1
k
m)| : m M
= [k[ inf|x +m
0
| : m
0
M
= [k[p(x +M).
If k = 0, then p(0 +M) = 0, since m = 0 M.
Finally,
p((x +M) + (y +M)) = p(x +y +M)
= inf|(x +y) +m| : m M
= inf|(x +m
1
) + (y +m
2
)| : m
1
, m
2
M
inf|x +m
1
| +|y +m
2
| : m
1
, m
2
M
= p(x +M) +p(y +M).
In the case where M is closed in X, suppose that p(x +M) = 0 for some x X.
Then
inf|x +m| : m M = 0,
so there exist m
n
M, n 1, so that x = lim
n
m
n
. Since Mis closed, x M
and so x +M= x + (x) +M= 0 +M, proving that p is a norm.
The converse statement is left as an exercise.
2
1.22. Let X be a normed linear space and Mbe a linear manifold in X. We shall
denote the canonical quotient map from X to X/M by q (or q
M
if the need to be
specic arises). When Mis closed in X, we shall denote the norm from Theorem 1.21
once again by | | (or | |
X/M
), so that
|q(x)| = |x +M| = inf|x +m| : m M.
It is clear that |q(x)| |x| for all x X, and so q is continuous. Indeed, given
> 0, we can take = to get |x y| < implies |q(x) q(y)| |x y| < .
We shall see below that q is also an open map - i.e. it takes open sets to open sets.
1.23. Theorem. Let X be a normed linear space and M be a closed subspace
of X.
(a) If X is complete, then so are M and X/M.
(b) If M and X/M are complete, then so is X.
Proof.
(a) Suppose that X is complete. We rst show that M is complete.
Let (m
n
)
n=1
be a Cauchy sequence in M. Then it is Cauchy in X and
X is complete, so that x = lim
n
m
n
X. Since Mis closed in X, x M.
Thus M is complete.
Note that this argument shows that any closed subset of a complete
metric space is complete.
Next we show that X/M is also complete.
Let

n
q(x
n
) be an absolutely summable series in X/M. For each
n 1, choose m
n
M so that |x
n
+m
n
| |q(x
n
)| +
1
2
n
. Then
n
|x
n
+m
n
|
n
_
|q(x
n
)| +
1
2
n
_
< ,
so

n
(x
n
+m
n
) is summable in X since X is complete. Set
x
0
:=
n
(x
n
+m
n
).
By the continuity of q,
q(x
0
) = q(
n
(x
n
+m
n
))
=
n
q(x
n
+m
n
)
=
n
q(x
n
).
Thus every absolutely summable series in X/M is summable, and so by
Proposition 1.20, X/M is complete.
(b) Suppose next that M and X/M are both complete.
Let (x
n
)
n=1
be a Cauchy sequence in X. Then (q(x
n
))
n=1
is Cauchy in
X/M and thus q(y) = lim
n
q(x
n
) exists, by the completeness of X/M.
For n 1, choose m
n
M so that
|y (x
n
+m
n
)| < |q(y) q(x
n
)| +
1
2
n
.
Since (x
n
+ m
n
)
n=1
converges to y in X, it follows that it is a Cauchy
sequence. Since both (x
n
)
n=1
and (x
n
+m
n
)
n=1
are Cauchy, it follows that
(m
n
)
n=1
is also Cauchy a fact that follows easily from the observation
that
|m
j
m
i
| |(x
j
+m
j
) (x
i
+m
i
)| +|x
j
x
i
|.
But M is complete and so m := lim
n
m
n
M. This yields
y m = lim
n
(x
n
+m
n
) m = lim
n
x
n
,
so that (x
n
)
n=1
converges to y m in X. That is, X is complete.
2
1.24. Proposition. Let X be a normed linear space and Mbe a closed subspace
of X. Let q : X X/M denote the canonical quotient map.
(a) A subset W X/M is open if and only if q
1
(W) is open in X.
(b) The map q is an open map - i.e., if G X is open, then q(G) is open in
X/M.
Proof.
(a) If W X/M is open, then q
1
(W) is open in X because q is continuous.
Suppose next that W X/M and that q
1
(W) is open in X. Let
q(x) W. Then x q
1
(W), and so we can nd > 0 so that V
(x)
q
1
(W). If |q(y) q(x)| < , then |y x +m| < for some m M, and
thus q(y) = q(y + m) q(V
(x)) W. That is, V
(q(x)) W, and W is
open.
(b) Let G X be an open set. Observe that q
1
(q(G)) = G+M=
mM
G+m
is open, being the union of open sets. By (a), q(G) is open.
2
1.25. Let M be a nite-dimensional linear manifold in a normed linear space
X. Then M is closed in X. The proof of this is left as an assignment exercise.
1.26. Proposition. Let X be a normed linear space. If M and Z are closed
subspaces of X and dim Z < , then M+Z is closed in X.
Proof. Let q : X X/M denote the canonical quotient map. Since Z is a nite
dimensional vector space, so is q(Z). By the exercise preceding this Proposition,
q(Z) is closed in X/M. Since q is continuous, M+Z = q
1
(q(Z)) is closed in X.
2
Appendix to Section 1.
1.27. This course assumes that the reader has taken at least enough Real Anal-
ysis to have seen that (
p
K
(N), | |
p
) is a normed linear space for each 1 p .
Having said that, let us review Holders Inequality as well as Minkowskis Inequality
in this setting, since Holders Inequality is also useful in studying dual spaces in the
next Section. The reader will recall that Minkowskis Inequality is the statement
that the p-norm is subadditive; that is, that the p-norm satises condition (iii) of
Denition 1.2. We remark that both inequalities hold for more general L
p
-spaces.
Our decision to concentrate on
p
-spaces instead of their more general counterparts
is an attempt to accommodate the background of the students who took this course,
as opposed to a conscious eort to avoid L
p
-spaces.
Before proving Holders Inequality, we pause to prove the following Lemma.
1.28. Lemma. Let a and b be positive real numbers and suppose that
1 < p, q < satisfy
1
p
+
1
q
= 1. Then
a
1
p
b
1
q
a
p
+
b
q
.
Proof. Let 0 < t < 1 and consider the function
f(x) = x
t
tx +t 1,
dened on (0, ). Then
f
(x) = tx
t1
t = t(x
t1
1).
Thus f(1) = 0 = f
(1). Since f
(x) > 0 for x (0, 1) and f
(x) < 0 for x (1, ),

it follows that
f(x) < f(1) = 0 for all x ,= 1.
That is, x
t
(1 t) +tx for all x > 0, with equality holding if and only if x = 1.
Letting x = a/b, t = 1/p yields
a
1
p
b
1
q
1
= a
1
p
b
1
p
= (
a
b
)
1
p
(1
1
p
) +
1
p
(
a
b
)
=
1
p
(
a
b
) +
1
q
.
Multiplying both sides of the equation by b yields the desired inequality.
2
1.29. Theorem. Holders Inequality
Let 1 p, q , and suppose that
1
p
+
1
q
= 1. Let x = (x
n
)
n

p
and
y = (y
n
)
n

q
. If z = (z
n
)
n
, where z
n
= x
n
y
n
for all n 1, then z
1
and
|z|
1
|x|
p
|y|
q
.
Proof. The cases where p = 1 or p = are routine and are left to the reader.
First let us suppose that |x|
p
= |y|
q
= 1. Applying the previous Lemma to
our sequences x and y yields, for each n 1,
[x
n
y
n
[ = ([x
n
[
p
)
1
p
([y
n
[
q
)
1
q
1
p
[x
n
[
p
+
1
q
[y
n
[
q
,
so that
n
[z
n
[ =
n
[x
n
y
n
[
1
p
n
[x
n
[
p
+
1
q
n
[y
n
[
q
=
1
p
|x|
p
p
+
1
q
|y|
q
q
= 1.
In general, if x
p
and y
q
, let u = x/(|x|
p
), v = v/(|y|
q
) so that
|u|
p
= 1 = |v|
q
and so
1
|x|
p
|y|
q
n
[x
n
y
n
[ =
n
[u
n
v
n
[
1.
Thus
|z|
1
|x|
p
|y|
q
.
2
Holders Inequality is the key to proving Minkowskis Inequality.
1.30. Theorem. Minkowskis Inequality.
Let 1 p , and suppose that x = (x
n
)
n
and y = (y
n
)
n
are in
p
. Then
x +y = (x
n
+y
n
)
n

p
and
|x +y|
p
|x|
p
+|y|
p
.
Proof. Again, the cases where p = 1 and where p = are left to the reader.
Suppose therefore that 1 < p < . Observe that if a, b > 0, then
_
a +b
2
_
p
a
p
+b
p
,
so that (a +b)
p
2
p
(a
p
+b
p
). It follows that
n
[x
n
+y
n
[
p
2
p
n
([x
n
[
p
+[y
n
[
p
) < ,
which proves that x +y
p
.
By Holders Inequality,
n
[x
n
+y
n
[
p1
[x
n
[ |x|
p
|([x
n
+y
n
[
p1
)
n
|
q
,
and similarly
n
[x
n
+y
n
[
p1
[y
n
[ |y|
p
|([x
n
+y
n
[
p1
)
n
|
q
.
Now
|([x
n
+y
n
[
p1
)
n
|
q
=
_
n
[x
n
+y
n
[
(p1)q
_1
q
=
_
n
[x
n
+y
n
[
(pqq)
_1
q
=
_
n
[x
n
+y
n
[
p
_1
q
= |(x
n
+y
n
)
n
|
p/q
p
.
Hence
|x +y|
p
p
=
n
[x
n
+y
n
[ [x
n
+y
n
[
p1
n
([x
n
[ +[y
n
[) [x
n
+y
n
[
p1
(|x|
p
+|y|
p
) |([x
n
+y
n
[
p1
)
n
|
q
= (|x|
p
+|y|
p
) |(x
n
+y
n
)
n
|
p/q
,
from which we get
|x +y|
p
= |x +y|
pp/q
p
|x|
p
+|y|
p
.
2
Let us now examine a couple of examples of useful Banach spaces whose de-
nitions require a somewhat better background in Analysis than we are assuming in
the main body of the text.
1.31. Example. Let x = (x
n
)
n
be a sequence of complex (or real) numbers.
The total variation of x is dened by
V (x) :=
n=1
[x
n+1
x
n
[.
If V (x) < , we say that x has bounded variation. The space
bv := (x
n
)
n
: x
n
K, n 1, V (x) <
is called the space of sequences of bounded variation. We may dene a norm
on bv as follows: for x bv, we set
|(x
n
)
n
|
bv
:= [x
1
[ +V (x) = [x
1
[ +
n=1
[x
n+1
x
n
[.
It can be shown that bv is complete under this norm, and hence that bv is a
Banach space.
If we let bv
0
= (x
n
)
n
bv : lim
n
x
n
= 0, then
|(x
n
)
n
|
bv
0
:= V ((x
n
)
n
)
denes a norm on bv
0
, and again, bv
0
is a Banach space with respect to this norm.
1.32. Example. The geometric theory of real Banach spaces is an active and
exciting area. For a period of time, the following question was open [Lin71]: does
every innite-dimensional Banach space contain a subspace which is linearly home-
omorphic to one of the spaces
p
, 1 p < or c
0
? In 1974, B.S. Tsirelson [Tsi74]
provided a counterexample to this conjecture. In this example, we shall discuss
the broad outline of the construction of the Tsirelson space, omitting the proofs of
certain technical details.
We begin by considering the space c
0
of Example 1.4. For each n 1, let e
n
c
0
denote the sequence (0, 0, ..., 0, 1, 0, 0, ...), with the unique 1 occurring in the n
th
coordinate. Given x = (x
n
)
n
c
0
, we may write x =

n=1
x
n
e
n
. Let us also dene
the map P
n
: c
0
c
0
via P
n
(x
k
)
k
:= (0, 0, ..., 0, x
n+1
, x
n+2
, x
n+3
, ...).
Given a nite set v
1
, v
2
, ..., v
r
of vectors in c
0
, we shall say that they are block-
disjoint for consecutively supported written v
1
< v
2
< < v
r
if there exist
1
,
2
, ...,
r
,
1
,
2
, ...,
r
N with
1

1
<
2

2
< <
r

r
so that supp(v
j
) [
j
,
j
], 1 j r. Here, for x = (x
n
)
n
c
0
,
supp(x) := j N : x
j
,= 0.
We shall write (v
1
, v
2
, ..., v
r
) for

r
j=1
v
j
when v
1
< v
2
< < v
r
.
For a subset B c
0
, we consider the following set of conditions which B may or
may not possess:
(a) x B implies that |x|
1; i.e. B is contained in the unit ball of c

0
.
(b) e
n
n=1
B.
(c) If x =

n=1
x
n
e
n
B, y = (y
n
)
n
c
0
and [y
n
[ [x
n
[ for all n 1, then
y B. (This is a hereditary property.)
(d) If v
1
< v
2
< < v
r
lie in B, then
1
2
P
r
((v
1
, v
2
, ..., v
r
)) B.
(e) For every x B there exists n N for which 2P
n
(x) B.
Our rst goal is to construct a set K which has all ve of these properties.
Let L
1
= re
j
: 1 r 1, j 1 and for n 1, set
L
n+1
= L
n

1
2
P
r
((v
1
, v
2
, ..., v
r
)) : r 1, v
1
< v
2
< < v
r
L
n
.
Let K denote the pointwise closure of
n1
L
n
. It can be shown that K c
0
. We
set T = co(K) denote the closed convex hull of K (with the closure taking place in
c
0
).
The Tsirelson space T is then dened as spanT. The norm on T is given by
the Minkowski functional which we shall encounter later when studying locally
convex spaces. It is given by |x|
T
= infr (0, ) : x rT, where rT = ry :
y T. As we shall later see, the denition of this norm ensures that T is precisely
the unit ball of T.
Although we shall not prove it here, (T, | |
T
) is a Banach space which does not
contain any copy of c
0
or
p
, 1 p < .
1.33. Example. Another Banach space of interest to those who study the
geometry of said spaces is James space.
For a sequence (x
n
)
n
of real numbers, consider the following condition, which
we shall call condition J: for all k 1,
sup
n
1
<n
2
<<n
k
_
(x
n
1
x
n
2
)
2
+ (x
n
2
x
n
3
)
2
+ + (x
n
k1
x
n
k
)
2
< .
The James space is dened to be:
J = (x
n
)
n
c
0
: (x
n
)
n
satises condition J.
The norm on J is dened via:
|(x
n
)
n
|
J
:= sup
n
1
<n
2
<<n
k
_
(x
n
1
x
n
2
)
2
+ (x
n
2
x
n
3
)
2
+ + (x
n
k1
x
n
k
)
2
1
2
.
It can be shown that J is a Banach space when equipped with this norm.
1.34. Example. Let X be a locally compact topological space and let B denote
the algebra of Borel subsets of X. Let be a positive measure on X, so that
: B R
satises
(a) () = 0;
(b) (B) 0 for all B B;
(c) if B
n
n
is a sequence of disjoint, measurable subsets from B, then
(
n
B
n
) =
n
(B
n
).
The measure is said to be nite if (X) < , and it is said to be regular if
(i) (K) < for all compact subsets K B;
(ii) (B) = sup(K) : K B, K compact for all B B; and
(iii) (B) = inf(G) : B G, G open for all B B.
A complex-valued, Borel measure on X is a function
: B C
satisfying:
(a) () = 0, and
(b) if B
n
n
is a sequence of disjoint, measurable subsets from B, then
(
n
B
n
) =
n
(B
n
).
Let be a complex-valued Borel measure on X. For each B B, a measurable
partition of B is a nite collection E
1
, E
2
, ..., E
k
of disjoint, measurable sets
whose union is B. We dene the variation [[ of to be the function dened as
follows: for B B,
[[(B) := sup
k
j=1
[(E
j
)[ : E
j
k
j=1
is a measurable partition of B.
It is routine to verify that [[ is then a nite, positive Borel measure on X. We
say that is regular if [[ is.
It is clear that every complex-linear combination of nite, positive, regular Borel
measures yields a complex-valued, regular Borel measure on X. A standard result
from measure theory known as the Hahn-Jordan Decomposition Theorem
states that the converse holds, namely: every complex-valued, regular Borel measure
can be written as a complex-linear combination of (four) nite, positive, regular
Borel measures.
Let M
C
(X) denote the complex vector space of all complex-valued, regular Borel
measures on X. Then the map
| | : M
C
(X) [0, )
[[(X)
denes a norm on M
C
(X), and M
C
(X) is complete with respect to this norm.
1.35. In Theorem 1.23, we showed that if X is a Banach space and Mis a closed
subspace, then X/Mis complete. Our proof there was based upon Proposition 1.20.
This result also admits a direct proof in terms of Cauchy sequences:
Theorem. Let X be a Banach space and suppose that M is a closed subspace of
X. Then X/M is complete.
Proof.
Let (q(x
n
))
n=1
be a Cauchy sequence in X/M. For each n 1, there exists
k
n
> 1 so that i, j k
n
implies |q(x
i
) q(x
j
)| < 2
n
. Without loss of generality,
we may assume that k
n
> k
n1
for all n 2.
Set z
n
:= x
kn
, n 1 and let m
1
= 0. For n > 1, choose m
n
M so that
|(z
n1
+m
n1
) (z
n
+m
n
)| < 2
(n1)
.
That this is possible follows from the denition of the quotient norm along with the
inequality of the second paragraph. If we now dene y
n
:= z
n
+ m
n
, n 1, then
q(y
n
) = q(z
n
) = q(x
kn
), and for n
2
> n
1
,
|y
n
1
y
n
2
|
n
2
n
1
j=1
|y
n
1
+j
y
n
1
+j1
|
n
2
n
1
j=1
(
1
2
)
(n
1
+j1)
j=0
(
1
2
)
(n
1
+j)
= (
1
2
)
n
1
1
,
from which it follows that (y
n
)
n=1
is Cauchy in X. Since X is complete, y :=
lim
n
y
n
X, and since the quotient norm is contractive, q(y) = lim
n
q(y
n
) =
lim
n
q(x
kn
). Since (q(x
n
))
n=1
is Cauchy, q(y) = lim
n
q(x
n
), which proves
that every Cauchy sequence in X/M converges - i.e. that X/M is complete.
2
I have the body of an eighteen year old. I keep it in the fridge.

Spike Milligan
2. AN INTRODUCTION TO OPERATORS 17
2. An introduction to operators
Some people are afraid of heights. Not me, Im afraid of widths.
Steven Wright
2.1. The study of mathematics is the study of mathematical objects and the
relationships between them. These relationships are often measured by functions
from one object to another. Of course, when both objects belong to the same
category (be it the category of vector spaces, groups, rings, etc), it is to be expected
that the most important maps between these objects will be morphisms from that
category. In this Section we shall concern ourselves with bounded linear operators
between normed linear spaces. These bounded linear maps, as we shall soon discover,
coincide with those linear maps which are continuous in the norm topology. Since
normed linear spaces are vector spaces equipped with a norm topology, the bounded
linear operators are the natural morphisms between them.
It should be pointed out that Banach spaces can be quite complicated to analyze.
For this reason, many people working in this area often study the structure and
geometry of these spaces without necessarily emphasizing the study of the linear
maps between them. In the next Section we shall examine the notion of a Hilbert
space. These are amongst the best-behaved Banach spaces, and their structure
is relatively well understood. For this reason, fewer people study Hilbert spaces
alone; Hilbert space theory tends to focus on the theory of the bounded linear maps
between them, as well as algebras of such bounded linear operators.
We would also be remiss if we failed to point out that not everyone on the planet
restricts themselves to bounded (i.e. continuous) linear operators. Dierentiation
has the grave misfortune of being an unbounded linear operator, but nevertheless
it is hard to avoid if one wishes to study the world around one - or around ones
friends, acquaintances, enemies, and every other one. Indeed, in applied mathemat-
ics and physics, it is often the case that the unbounded linear operators are the more
interesting examples. Having said that, we shall leave it to the disciples of those
schools to wax poetic on these topics.
2.2. Denition. Let X and Y be normed linear spaces, and let T : X Y
be a linear map. We say that T is a bounded operator if there exists a constant
k 0 so that |Tx| k|x| for all x X. When T is bounded, we dene
|T| = infk 0 : |Tx| k|x| for all x X.
We shall refer to |T| as the operator norm of T.
It is, of course, understood that the norm of Tx is computed using the Y-norm,
while the norm of x is computed using the X-norm. As we shall see below, the
operator norm does dene a bona de norm on the vector space of bounded linear
maps from X to Y, thereby justifying our terminology.
Our interest in bounded operators stems from the fact that they are precisely
the continuous operators from X to Y.
2.3. Theorem. Let X and Y be normed linear spaces and T : X Y be a
linear map. The following are equivalent:
(a) T is continuous on X.
(b) T is continuous at 0.
(c) T is bounded.
(d)
1
:= sup|Tx| : x X, |x| 1 < .
(e)
2
:= sup|Tx| : x X, |x| = 1 < .
(f)
3
:= sup|Tx|/|x| : 0 ,= x X < .
Furthermore, if any of these holds, then
1
=
2
=
3
= |T|.
Proof.
(a) implies (b): This is trivial.
(b) implies (c): Suppose that T is continuous at 0. Let = 1 and choose > 0
so that |x 0| < implies that |Tx T0| = |Tx| < = 1. If |y| /2,
then |Ty| 1, and so |x| 1 implies |T
_
2
x
_
| 1, i.e. |Tx| 2/.
Thus |T| 2/ < , and T is bounded.
(c) implies (d): This is trivial.
(d) implies (e): This is trivial.
(e) implies (f): Again, this is trivial.
(f) implies (a): Observe that for any x X, |Tx|
2
|x|. (For x ,= 0, this
follows from the hypothesis, while the linearity of T implies that T0 = 0, so
the inequality also holds for x = 0.) Thus if > 0 and |xy| < /(
2
+1),
then
|Tx Ty| = |T(x y)|
2
|x y| < .
The proof of the nal statement is left as an exercise for the reader.
2
2.4. Computing the operator norm of a given operator T is not always a simple
task. For example, suppose that H = (C
2
, | |
2
) is a two-dimensional Hilbert
space with standard orthonormal basis e
1
= (1, 0), e
2
= (0, 1). Let T : H H
be the map whose matrix with respect to this basis is
_
1 2
3 4
_
, so that T(x, y) =
(x + 2y, 3x + 4y). By denition,
|T| = sup|Tz| : z C
2
, |z| 1
= sup
_
[x + 2y[
2
+[3x + 4y[
2
: x, y C,
_
[x[
2
+[y[
2
1,
which involves non-linear equations. For Hilbert spaces of low dimension say,
less than dimension 5 alternate methods exist (but wont be developed just yet).
Instead, we turn our attention to special classes of operator which are simple enough
to allow us to obtain interesting results.
2.5. Example. Multiplication operators.
(a) Let X =
_
(([0, 1], C), | |
_
, and suppose that f X. Dene
M
f
: X X
g fg
.
It is routine to check that M
f
is linear. If |g|
1, then
|M
f
g|
= |fg|
= sup[f(x)g(x)[ : x [0, 1] |f|
|g|
.
Thus |M
f
|
|f|
< , and M
f
is bounded.
Setting g(x) = 1, x [0, 1], we have g X, |g|
= 1 and |M
f
g|
=
|f|
, so that |M
f
| |f|
and therefore |M
f
| = |f|
.
For (hopefully) obvious reasons, M
f
is referred to as a multiplication
operator .
(b) We now consider a similar operator acting on a Hilbert space. Let H =
L
2
(X, d), where d is a positive, regular Borel measure. Suppose that
f L
(X, d) and let

M
f
: X X
g fg
.
Once again, it is easy to check that M
f
is linear, while for g H,
|M
f
g|
2
=
_
_
1
0
[f(x)g(x)[
2
d
_1
2
_
_
1
0
|f|
2
[g(x)[
2
d
_1
2
= |f|
|g|
2
,
so that |M
f
| |f|
, and hence M
f
is bounded. As for a lower bound on
the norm of M
f
, for each n 1, let F
n
= x X : [f(x)[ |f|
1/n.
Then F
n
is measurable and (F
n
) > 0 by denition of |f|
. Let E
n
F
n
be a measurable set for which 0 < (E
n
) < , n 1. The existence of
such sets E
n
, n 1 follows from the regularity of the measure . Let
g
n
=
En
, the characteristic function of E
n
. Then g
n
L
2
(X, ) for all
n 1 and
|M
f
g
n
|
2
=
_
_
1
0
[f(x)g
n
(x)[
2
d
_1
2
=
_
_
En
[f(x)[
2
d
_1
2
_
_
En
(|f|
1/n)
2
d
_1
2
=
_
|f|
1/n
_ _
_
1
0
[g
n
(x)[d
_1
2
=
_
|f|
1/n
_
|g
n
|
2
.
From this we see that |M
f
| |f|
1/n. Since n 1 was arbitrary,

|M
f
| |f|
, and so |M
f
| = |f|
.
Observe that the computation of the norm of the operator depended
very much upon the underlying norms of the spaces involved.
(c) As a special case of this phenomenon, let X = N and suppose that d is
counting measure. Then H =
2
(N) and f
(N). As we are wont to

do when dealing with sequences, we denote by d
n
the value f(n) of f at
n N, so that f (d
n
)
n=1

(N). It follows that M

f
(x
n
)
n
= (d
n
x
n
)
n
for all (x
n
)
n

2
(N). By considering the matrix [M
f
] of M
f
with respect
to the standard orthonormal basis (e
n
)
n
for H, we see that
[M
f
] =
_
_
d
1
d
2
d
3
.
.
.
_
_
.
Thus, M
f
, often denoted in this circumstance as D = diagd
n
n
, is referred
to as a diagonal operator. The above calculation shows that
|D| = |M
f
| = |f|
= sup[f(n)[ : n 1 = sup[d
n
[ : n 1.
2.6. Example. Weighted shifts. With H =
2
(N) and (w
n
)
n

(N),
consider the map W : H H dened by
W(x
n
)
n
= (0, w
1
x
1
, w
2
x
2
, w
3
x
3
, ...) for all (x
n
)
n

2
(N).
We leave it as an exercise for the reader to show that W is a bounded linear operator
on H, and that
|W| = sup[w
n
[ : n 1.
Such an operator is referred to as a unilateral forward weighted shift.
If (v
n
)
n

(N) and we dene the linear map V : H H via

V (x
n
)
n
= (v
1
x
2
, v
2
x
3
, v
3
x
4
, ...) for all (x
n
)
n
H,
then once again V is bounded, |V | = sup[v
n
[ : n 1, and V is referred to as a
unilateral backward weighted shift.
Finally, consider H =
2
(Z), and with (u
n
)
n

(Z), dene the linear map

U : H H via
U(x
n
)
n
= (u
n1
x
n1
)
n
.
Again, U is bounded with |U| = sup[u
n
[ : n Z, and U is referred to as a
bilateral weighted shift. The reader should ask himself/herself why we do not
refer to forward and backward bilateral shift operators.
2.7. Example. Dierentiation operators. Consider the linear manifold
T(D) = p : p a polynomial (((D), | |
). Dene the map

D : T(D) T(D)
p p
,
the derivative of p. Then if p
n
(z) = z
n
, |p
n
|
= 1 for each n 1 and Dp

n
=
np
n1
, whence |D| n for all n 1. In particular, D is not bounded. That is,
dierentiation is not continuous on the linear space of polynomials.
2.8. Notation. The set of bounded linear operators from the normed linear
space X to the normed linear space Yis denoted by B(X, Y). If X = Y, we abbreviate
this to B(X). We now full an earlier promise by proving that the map T |T|
does indeed dene a norm on B(X, Y).
2.9. Proposition. Let X and Y be normed linear spaces. Then B(X, Y) is a
vector space and the operator norm is a norm on B(X, Y).
Proof. Since linear combinations of continuous functions between topological spaces
are continuous, B(X, Y) is a vector space.
As to the second assertion: for R, T B(X, Y) and k K,
(i) |T| = sup|Tx| : |x| 1 0;
(ii) |T| = 0 if and only if |Tx|/|x| = 0 for all x ,= 0, which in turn happens
if and only if Tx = 0 for all x X; i.e. if and only if T = 0.
(iii)
|kT| = sup|kTx| : |x| 1
= sup[k[ |Tx| : |x| 1
= [k[ sup|Tx| : |x| 1
= [k[ |T|.
(iv)
|R +T| = sup|(R +T)x| : |x| 1
sup|Rx| +|Tx| : |x| 1
sup|Rx| +|Ty| : |x|, |y| 1
= |R| +|T|.
This completes the proof.
2
2.10. Theorem. Let X and Y be normed linear spaces and suppose that Y is
complete. Then B(X, Y) is complete, and as such it is a Banach space.
Proof. Suppose that

n=1
T
n
is an absolutely summable series in B(X, Y). Given
x X,
n=1
|T
n
x|
n=1
|T
n
| |x| = |x|
_

n=1
|T
n
|
_
< ,
and thus, since Y is compete,

n=1
T
n
x Y exists. Moreover, the linearity of each
T
n
implies that the map T : X Y dened via Tx =

n=1
T
n
x is linear, while
|x| 1 implies from above that |Tx|
n=1
|T
n
|. Hence
|T|
n=1
|T
n
| < ,
implying that T is bounded.
Finally,
|Tx
N
n=1
T
n
x| = |
n=N+1
T
n
x|
|x|
n=N+1
|T
n
|,
from which it easily follows that T = lim
N
N
n=1
T
n
. That is, the series

n=1
T
n
is summable. By Proposition 1.20, B(X, Y) is complete.
2
As a particular case of Theorem 2.10, consider the case where Y = K, the base
eld.
2.11. Denition. Let X be a normed linear space. The dual of X is B(X, K),
and it is denoted by X
. The elements of X
are referred to as continuous linear

functionals or when no confusion is possible as functionals on X.
Since K is complete, Theorem 2.10 implies that X
is again a Banach space.

As such, we may consider the dual space of X
, namely X
(2)
= X
:= (X
,
known as the double dual of X, and more generally, the n
th
-iterated dual spaces
X
(n)
= (X
(n1)
)
, n 3. All of these are Banach spaces.

Before proceeding to some examples, let us rst introduce some notation and
terminology which will prove useful.
2.12. Denition. A collection e
n
n=1
in a Banach space X is said to be
a Schauder basis if every x X can be written in a unique way as a norm
convergent series
x =
n=1
x
n
e
n
for some choice x
n
K, n 1.
2.13. Example.
(a) For each n 1, let e
n
denote the sequence e
n
= (0, 0, ..., 0, 1, 0, 0, ...) K
N
,
where the unique 1 occurs in the n
th
positiion. Then e
n
n
is a Schauder
basis for c
0
and for
p
, 1 p < . Observe that it is not a Schauder basis
for
.
We shall refer to e
n
n
as the standard Schauder basis for c
0
and
for
p
.
(b) It is not as obvious what one should choose as the Schauder basis for
(([0, 1], R). It was Schauder [Sch27] who rst discovered a basis for this
space. The description of such a basis is non-trivial.
2.14. Example. Consider the Banach space
c
0
= (x
n
)
n=1
K
N
: lim
n
x
n
= 0,
equipped with the supremum norm |(x
n
)
n
|
= sup[x
n
[ : n 1. We claim that
c
0
is isometrically isomorphic to
1
=
1
(N). To see this, consider the map
:
1
c
0
z := (z
n
)
n

z
,
where
z
((x
n
)
n
) =

n=1
x
n
z
n
for all (x
n
)
n
c
0
. That is linear is readily seen.
That the sum converges absolutely is also easy to verify.
If |(x
n
)
n
|
1, then [x
n
[ 1 for all n 1, so that
[
z
((x
n
)
n
)[ = [
n=1
x
n
z
n
[
n=1
[x
n
z
n
[
[z
n
[ = |z|
1
.
Hence |
z
| |z|
1
. On the other hand, if we set v[n] = (w
1
, w
2
, w
3
, ..., w
n
, 0, 0, 0, ...)
for each n 1 (where w
j
= z
j
/[z
j
[ if z
j
,= 0, while w
j
= 1 if z
j
= 0), then v[n] c
0
,
|v[n]|
= 1 for all n 1, and
z
(v[n]) =
n
j=1
[z
j
[.
From this it follows that |
z
| |z|
1
. Combining these two estimates yields |
z
| =
|z|
1
.
Thus is an isometric injection of
1
into c
0
. There remains to prove that is
surjective.
To that end, suppose that c
0
. Let e
n
n
denote the standard Schauder
basis for c
0
, and for each n 1, let w
n
= (e
n
). Observe that if
n
:= w
n
/[w
n
[ for
w
n
,= 0, and
n
:= 0 if w
n
= 0, then
v[n] :=
n
k=1
n
e
n
c
0
and |v[n]|
= 1. Since, for all n 1,

n
k=1
[w
n
[ =
k=1
n
w
n
k=1
(
n
e
n
)
= [(v[n])[
|| |v[n]|
= ||,
we see that w := (w
1
, w
2
, w
3
, ...)
1
. A routine computation shows that
w
[
c
00
=
c
00
. Since ,
w
are both continuous and since c
00
is dense in c
0
, =
w
= (w).
Thus is onto.
2.15. Example. Let 1 p < . Recall from your Real Analysis courses that
there exists an isometric linear bijection :
q
(
p
)
, where as always q is the

Lebesgue conjugate of p satisfying 1/p + 1/q = 1.
The map is dened via:
:
q
(
p
)
z
z
,
where for z = (z
n
)
n

q
, we have
z
((x
n
)
n
) =

n
x
n
z
n
.
We normally abbreviate this result by saying that the dual of
p
is
q
, when
1 p < . We refer the reader to the Appendix to Section 2 for a proof of this
result.
2.16. Example. The above example can be extended to more general measure
spaces. Let 1 p < , and suppose that is a -nite, positive, regular Borel
measure on L
p
(X, ). Again, the map
: L
q
(X, ) L
p
(X, )
g
g
,
where
g
(f) =
_
X
fgd denes a linear, isometric bijection between L
q
(X, ) and
L
p
(X, )
.
If we drop the hypothesis that is -nite, the result still holds for 1 < p < .
For reasons we shall discuss in the next Section, when p = 2, we often consider
the related map
: L
2
(X, ) L
2
(X, )
g
g
,
where
g
(f) =
_
X
fgd denes a linear, isometric bijection between L
2
(X, ) and
L
2
(X, )
.
2.17. Example. A function f : [0, 1] K is said to be of bounded variation
if there exists > 0 such that for every partition 0 = t
0
< t
1
< t
2
< < t
n
= 1
of [0, 1],
n
i=1
[f(t
i
) f(t
i1
)[ .
The inmum of all such s for which the above inequality holds is denoted by
|f|
v
, and is called the variation of f.
Recall from your earlier courses in Analysis that if f is a function of bounded
variation, then for all x (0, 1], f(x
) := lim
tx
f(t) exists, and for all x [0, 1),
f(x
+
) := lim
tx
+ f(t) exists (though they might not be equal, of course). We set
f(0
) = f(0) and f(1

+
) = f(1). It is known that a function of bounded variation
admits at most a countable number of discontinuities in the interval [0, 1], and for
g (([0, 1], K), the Riemann-Stieltjes integral
_
1
0
g df
exists.
Let
BV [0, 1] = f : [0, 1] K : |f|
v
< , f is left-continuous on (0, 1) and f(0) = 0.
Then (BV [0, 1], | |
v
) is a Banach space with norm given by the variation.
Indeed, the dual of
_
(([0, 1]), | |
_
is isometrically isomorphic to BV [0, 1]. For
f BV [0.1] and g (([0, 1]), we dene a functional
f
((([0, 1], K) by
f
(g) :=
_
1
0
g df.
2.18. Proposition. Let X be a normed linear space. Then there exists a con-
tractive linear map : X X
.
Proof. Let z X and dene a map z : X
K via z(x
) = x
(z). It is routine to
check that z is linear, and if |x
| 1, then [ z(x
)[ = [x
(z)[ |x
| |z|, so that
| z| |z|; in particular, z X
.
It is also easy to verify that the map
: X X
z z
is linear, and the rst paragraph shows that is contractive.
2
2.19. The map is referred to as the canonical embedding of X into X
. It
is not necessarily the only embedding of interest, however. Once we have proven the
Hahn-Banach Theorem, we shall be in a position to show that is in fact isometric.
We point out that if is an isometric bijection from X onto X
, then X is said
to be reexive. These are in some sense amongst the best behaved Banach spaces.
We shall return to the notion of reexivity of Banach spaces in a later section.
2.20. Although norms of operators can be dicult to compute, there are cases
where useful estimates can be obtained.
Consider the Volterra operator
V : (([0, 1], K) (([0, 1], K)
f V f
,
where V f(x) =
_
x
0
f(t)dt for all x [0, 1]. (Since all functions are continuous, it
suces to consider Riemann integration.)
Then
|V f| = sup[V f(x)[ : x [0, 1]
= sup[
_
x
0
f(t)dt[ : x [0, 1]
sup
_
x
0
|f|
dt : x [0, 1]
= sup(x 0) |f|
: x [0, 1]
= |f|
.
Thus |V | 1. If 1(x) := 1 for all x [0, 1], then 1 (([0, 1], K), |1|
= 1,
and V 1 = j, where j(x) = x, x [0, 1]. But then |V 1|
= |j|
= 1, showing
that |V | 1, and hence |V | = 1.
Far more interesting (and useful) is the computation of |V
n
| for n 2.
Let us rst general the construction of the operator V . We may consider the
function k : [0, 1] [0, 1] C dened by
k(x, y) =
_
0 if x < y,
1 if x y.
Then
(V f)(x) =
_
x
0
f(y)dy =
_
1
0
k(x, y)f(y)dy.
The function k = k(x, y) is referred to as the kernel of the integral operator V .
This should not be confused with the notion of a null space of a linear map, also
referred to as its kernel.
Now
(V
2
f)(x) = (V (V f))(x)
=
_
1
0
k(x, t) (V f)(t)dt
=
_
1
0
k(x, t)
_
1
0
k(t, y) f(y)dy dt
=
_
1
0
f(y)
_
1
0
k(x, t) k(t, y)dt dy
=
_
1
0
f(y) k
2
(x, y)dy,
where k
2
(x, y) =
_
1
0
k(x, t) k(t, y)dt is a new kernel for the integral operator V
2
.
Note that
[k
2
(x, y)[ = [
_
1
0
k(x, t) k(t, y)dt[
= [
_
x
y
k(x, t) k(t, y)dt[
= (x y) for x > y,
while for x < y, k
2
(x, y) = 0.
In general, since x y < 1 0 = 1, we get
(V
n
f)(x) =
_
1
0
f(y) k
n
(x, y)dy, where
k
n
(x, y) =
_
1
0
k(x, t) k
n1
(t, y)dt, and where
[k
n
(x, y)[
1
(n 1)!
(x y)
n1
1
(n 1)!
.
It follows that
|V
n
| = sup
f=1
|V
n
f|
= sup
f=1
|
_
1
0
f(y) k
n
(x, y)dy|
sup
f=1
|f|
|k
n
(x, y)|
1/(n 1)!.
A simple consequence of these computations is that
lim
n
|V
n
|
1
n
= lim
n
1/(n 1)! = 0.
We shall have more to say about this in the Appendix to Section 3.
2.21. Example. Let e
1
, e
2
, ..., e
n
denote the standard basis for K
n
, so that
e
j
= (0, 0, ..., 0, 1, 0, ..., 0), where the 1 occurs in the j
th
position, 1 j n. Suppose
that 1 p , and that K
n
carries the p-norm from Example 1.6.
Let [t
ij
] M
n
(K), and dene the map
T : K
n
K
n
x [t
ij
]x.
It is instructive, while not dicult, to prove that if every row and every column
of [t
ij
] has at most one non-zero entry, then
|T| = max
1i,jn
[t
ij
[.
We leave this as an exercise for the reader.
2.22. Example. Let n 1 be an integer and consider M
n
= M
n
(C). For
T X, T
T is a hermitian matrix, and as such, has positive eigenvalues. Denote by

s
1
, s
2
, ..., s
n
the square roots of these eigenvalues (counted according to multiplicity)
and for 1 p < , set
|T|
p
=
_
n
k=1
s
p
k
_1
p
.
For p = , set
|T|
= maxs
1
, s
2
, ..., s
n
.
It can be shown that | |
p
is indeed a norm on M
n
for all 1 p . Let us
denote the space M
n
equipped with the norm | |
p
by (
n
p
. We shall refer to it as
the n-dimensional Schatten p-class of operators on C
n
. Then we shall leave it as an
exercise for the reader to prove that (
n
p
(
n
q
, where q is the Lebesgue conjugate
of p, i.e.,
1
p
+
1
q
= 1.
The above identication can be realized via the map:
: (
n
q
(
n
p
R
R
where
R
: (
n
p
C is the map
R
(T) = tr(RT), and where tr[x
ij
] =

n
k=1
x
kk
denotes the standard trace functional on M
n
.
The above result has a generalisation to innite-dimensional Hilbert spaces. We
refer the reader to [Dav88] for a more detailed treatment of this topic.
2.23. Example. We now return to the proof of the fact that the dual of
p
is
q
when 1 < p < , as stated in Example 2.15.
Given z = (z
n
)
n

q
, we dene
z
:
p
K
(x
n
)
n

n
x
n
z
n
.
Note that by Holders Inequality, Theorem 1.29, for x = (x
n
)
n

p
, we have
[
z
(x)[ =
n
x
n
z
n
n
[x
n
z
n
[
= |xz|
1
|x|
p
|z|
q
,
so that indeed
z
(x) K. Clearly
z
is linear, and so the above argument also
shows that |
z
| |z|
q
.
Furthermore, if we set x
n
=
n
z
q1
n
=
n
z
q/p
n
, where
n
K is chosen so that
[
n
[ = 1 and x
n
z
n
0 for all n 1, then
_
n
[x
n
[
p
_1
p
=
_
n
([z
n
[
q
p
)
p
_1
p
=
_
n
[z
n
[
q
_1
p
= |z|
q/p
q
,
so that x
p
, and
[
z
(x)[ =
n
x
n
z
n
=
n
[z
n
[
q
= |z|
q/p
q
|z|
q
= |x|
p
|z|
q
,
so that |
z
| |z|
q
, and therefore |
z
| = |z|
q
.
Consider the map dened via:
:
q
(
p
)
z
z
,
with
z
dened as above. Then is easily seen to be linear, and from above, it is
isometric (hence injective). There remains only to show that is surjective.
To that end, let (
p
)
. For each n 1, let z

n
:= (e
n
), where e
n
n
is the
standard Schauder basis for
p
. Set z[n] :=

n
k=1
z
k
e
k
, and x[n] :=

n
k=1
k
z
q1
k
,
where as before
k
is chosen so that [
k
[ = 1 and x
k
z
k
0 for all k 1. Then
z[n]
q
and x[n]
p
for all n 1.
Observe that if y = (y
n
)
n

p
, then by the continuity of ,
(y) =
_
n
y
n
e
n
_
=
n
y
n
(e
n
) =
n
y
n
z
n
.
Now
[(x[n])[ =
k=1
k
z
q1
k
z
k
=
n
k=1
[z
k
[
q
= |z[n]|
q1
q
|z[n]|
q
,
where
|z[n]|
q1
q
=
_
n
k=1
[z
k
[
q
_a1
q
=
_
n
k=1
([z
k
[
q/p
)
p
_
1
1
q
=
_
n
k=1
[x
k
[
p
_
1/p
= |x[n]|
p
.
Thus
[(x[n])[ = |x[n]|
p
|z[n]|
q
|x[n]|
p
|| for all n 1.
It follows that |z[n]|
q
|| for all n 1, so that if z := (z
n
)
n
, then z
q
with
|z|
q
||.
Finally, (y) =
z
(y) for all y
p
, so that =
z
= (z), proving that is
surjective, as required.
I handed in a script last year and the studio didnt change one word.
The word they didnt change was on page 87.
Steve Martin
3. HILBERT SPACE 31
3. Hilbert space
You should always go to other peoples funerals; otherwise, they wont
come to yours.
Yogi Berra
3.1. In this brief Chapter, we shall examine a class of very well-behaved Banach
spaces, namely the class of Hilbert spaces. Hilbert spaces are the generalizations
of our familiar (two- and) three-dimensional Euclidean space. There are two basic
approaches to studying Hilbert spaces. If one is interested in Banach space geometry
and many people are then one often tries to compare other Banach spaces to
Hilbert spaces. As an example of such a phenomenon, we mention the calculation
of the Banach-Mazur distance between Banach spaces, which we dene in the
Appendix to this Section.
In the second approach, one decides that because Hilbert spaces are so well-
behaved, they are in some sense understood, and for this reason they are less
interesting to study than the set of bounded linear operators acting upon them.
One can then study the operators individually or in sets which have no algebraic
structure this kind of analysis belongs to Single Operator Theory. Alternatively,
one can various Operator Algebras, of which there are myriads of examples. The
literature dealing with operators and operator algebras is vast.
3.2. Recall that a Hilbert space H is a vector space equipped with an inner
product , ) so that that the induced norm |x| := x, x)
1/2
gives rise to a complete
normed linear space, i.e. a Banach space. [When the corresponding normed linear
space is not complete, we refer only to inner product spaces.]
In any inner product space we have the Cauchy-Schwarz Inequality:
[x, y)[
2
|x|
2
|y|
2
,
for all x, y H. We say that x and y are orthogonal if x, y) = 0, and we write
x y.
3.3. Example.
(a) If (X, ) is a measure space, then L
2
(X, ) is a Hilbert space, with the
inner product given by
f, g) =
_
X
f(x)g(x)d(x).
(b)
2
= (x
n
)
n
: x
n
K, n 1 and

n=1
[x
n
[
2
< is a Hilbert space, with
the inner product given by
(x
n
)
n
, (y
n
)
n
) =
n
x
n
y
n
.
The reader with a background in measure theory will recognize that the second
example is merely a particular case of the rst. While these are the canonical inner
products on these spaces, they are not the only ones.
For example, if (r
n
)
n
is any sequence of strictly positive integers, one can dene
a weighted
2
space relative to this sequence by setting
2
(rn)n
:= (x
n
)
n
K
N
:
n
r
n
[x
n
[
2
<
with inner product
(x
n
)
n
, (y
n
)
n
) =
n
r
n
x
n
y
n
.
3.4. Theorem. Let H be a Hilbert space and suppose that x
1
, x
2
, ..., x
n
H.
(a) [The Pythagorean Theorem] If the vectors are pairwise orthogonal, then
|
n
j=1
x
j
|
2
=
n
j=1
|x
j
|
2
.
(b) [The Parallelogram Law]
|x
1
+x
2
|
2
+|x
1
x
2
|
2
= 2
_
|x
1
|
2
+|x
2
|
2
_
.
Proof. Both of these results follow immediately from the denition of the norm in
terms of the inner product.
2
The Parallelogram Law is a useful tool to show that many norms are not Hilbert
space norms.
3.5. Theorem. Let H be a Hilbert space, and K H be a closed, non-empty
convex subset of H. Given x H, there exists a unique point y K which is closest
to x; that is,
|x y| = dist(x, K) = min|x z| : z K.
Proof. By translating K by x, it suces to consider the case where x = 0.
Let d := dist(0, K), and choose k
n
K so that |0 k
n
| < d +
1
n
, n 1. By the
Parallelogram Law,
|
k
n
k
m
2
|
2
=
1
2
|k
n
|
2
+
1
2
|k
m
|
2
|
k
n
+k
m
2
|
2
1
2
(d +
1
n
)
2
+
1
2
(d +
1
m
)
2
d
2
,
as
k
n
+k
m
2
K because K is assumed to be convex.
We deduce from this that the sequence k
n
n=1
is Cauchy, and hence converges
to some k K, since K is closed and H is complete. Clearly lim
n
k
n
= k implies
that d = lim
n
|k
n
| = |k|.
3. HILBERT SPACE 33
As for uniqueness, suppose that z K and that |z| = d. Then
0 |
k z
2
|
2
=
1
2
|k|
2
+
1
2
|z|
2
|
k +z
2
|
2
1
2
d
2
+
1
2
d
2
d
2
= 0,
and so k = z.
2
3.6. Theorem. Let H be a Hilbert space, and let / H be a closed subspace.
Let x H, and m /. The following are equivalent:
(a) |x m| = dist(x, /);
(b) The vector x m is orthogonal to /, i.e., x m, y) = 0 for all y /.
Proof.
(a) implies (b): Suppose that |x m| = dist (x, /), and suppose to the
contrary that there exists y / so that k := x m, y) ,= 0. There is no
loss of generality in assuming that |y| = 1. Consider z = m + ky /.
Then
|x z|
2
= |x mky|
2
= x mky, x mky)
= |x m|
2
ky, x m) kx m, y) +[k[
2
|y|
2
= |x m|
2
[k[
2
< dist(x, /),
a contradiction. Hence x m /
.
(b) implies (a): Suppose that x m /
. If z / is arbitrary, then
y := z m /, so by the Pythagorean Theorem,
|x z|
2
= |(x m) y|
2
= |x m|
2
+|y|
2
|x m|
2
,
and thus dist (x, /) |x m|. Since the other inequality is obvious, (a)
holds.
2
3.7. Remarks.
(a) Given any non-empty subset o H, let
o
:= y H : x, y) = 0 for all x o.
It is routine to show that o
is a norm-closed subspace of H. In particular,

_
o
spano,
the norm closure of the linear span of o.
(b) Recall from Linear Algebra that if 1 is a vector space and J is a (vector)
subspace of 1, then there exists a (vector) subspace A 1 such that
(i) J A = 0, and
(ii) 1 = J +A := w +x : w J, y A.
We say that J is algebraically complemented by A. The existence of
such a A for each J says that every vector subspace of a vector space is
algebraically complemented. We shall write 1 = JA to denote the fact
that A is an algebraic complement for J in 1.
If X is a Banach space and Y is a closed subspace of X, we say that Y
is topologically complemented if there exists a closed subspace Z of X
such that Z is an algebraic complement to Y. The issue here is that both Y
and Z must be closed subspaces. It can be shown that the closed subspace
c
0
of
is not topologically complemented in
. This result is known as

Phillips Theorem (see the paper of R. Whitley [Whi66] for a short but
elegant proof). We shall write X = YZ if Z is a topological complement
to Y in X.
Now let H be a Hilbert space and let / H be a closed subspace
of H. We claim that H = / /
. Indeed, if z / /
, then
|z|
2
= z, z) = 0, so z = 0. Also, if x H, then we may let m
1
/ be
the element satisfying
|x m
1
| = dist(x, /).
The existence of m
1
is guaranteed by Theorem 3.5. By Theorem 3.6,
m
2
:= x m
1
lies in /
, and so x / + /
. Since / and /
are closed subspaces of a Banach space and they are algebraically comple-
ments, we are done.
In this case, the situation is even stronger. The space / may admit
more than one topological complement in H- however, the space /
above
is unique in that it is an orthogonal complement. That is, as well as
being a topological complement to /, every vector in /
is orthogonal
to every vector in /.
(c) With / as in (b), we have H = //
, so that if x H, then we may

write x = m
1
+ m
2
with m
1
/, m
2
/
in a unique way. Consider

the map:
P : H //
x m
1
,
relative to the above decomposition of x. It is elementary to verify that P
is linear, and that P is idempotent, i.e., P = P
2
. We remark in passing
that m
2
= (I P)x, and that (I P)
2
= (I P) as well.
In fact, for x H, |x|
2
= |m
1
|
2
+|m
2
|
2
by the Pythagorean Theorem,
and so |Px| = |m
1
| |x|, from which it follows that |P| 1. If
/ ,= 0, then choose m / with |m| ,= 0. Then Pm = m, and so
|P| 1. Combining these estimates, /, = 0 implies |P| = 1.
We refer to the map P as the orthogonal projection of H onto /.
The map Q := (I P) is the orthogonal projection onto /
, and we leave
it to the reader to verify that |Q| = 1.
3. HILBERT SPACE 35
(d) Let ,= o H. We saw in (a) that o
spano. In fact, if we let

/= spano, then / is a closed subspace of H, and so by (b),
H = //
.
It is routine to check that o
= /
. Suppose that there exists 0 ,= x

o
, x , /. Then x H, and so we can write x = m

1
+m
2
with m
1
/,
and m
2
/
= o
(m
2
,= 0, otherwise x /). But then 0 ,= m
2
o
and so
m
2
, x) = m
2
, m
1
) +m
2
, m
2
)
= 0 +|m
2
|
2
,= 0.
This contradicts the fact that x o
. It follows that o
= spano.
(e) Suppose that / admits an orthonormal basis e
k
n
k=1
. Let x H, and let
P denote the orthogonal projection onto /. By (b), Px is the unique
element of / so that x Px lies in /
. Consider the vector w =
n
k=1
x, e
k
)e
k
. Then
x w, e
j
) = x, e
j
)
n
k=1
x, e
k
)e
k
, e
j
)
= x, e
j
)
n
k=1
x, e
k
)e
k
, e
j
)
= x, e
j
) x, e
j
) |e
j
|
2
= 0.
It follows that xw /
, and thus w = Px. That is, Px =

n
k=1
x, e
k
)e
k
.
3.8. Theorem. The Riesz Representation Theorem Let 0 ,= H be a
Hilbert space over K, and let H
. Then there exists a unique vector y H so

that
(x) = x, y) for all x H.
Moreover, || = |y|.
Proof. Given a xed y H, let us denote by
y
the map
y
(x) = x, y). Our goal
is to show that H
=
y
: y H. First note that if y H, then
y
(kx
1
+ x
2
) =
kx
1
+x
2
, y) = kx
1
, y)+x
2
, y) = k
y
(x
1
)+
y
(x
2
), and so
y
is linear. Furthermore,
for each x H, [
y
(x)[ = [x, y)[ |x||y| by the Cauchy-Schwarz Inequality. Thus
|
y
| |y|, and hence
y
is continuous - i.e.
y
H
.
It is not hard to verify that the map
: H H
y
y
is conjugate-linear (if K = C), otherwise it is linear (if K = R). From the rst
paragraph, it is also contractive. But [(y)](y) =
y
(y) = y, y) = |y|
2
, so that
|(y)| |y| for all y H, and is isometric as well. It immediately follows that
is injective, and there remains only to prove that is surjective.
Let H
. If = 0, then = (0). Otherwise, let / = ker , so that

codim/= 1 = dim /
, since H// K /
. Choose e /
with |e| = 1.
Let P denote the orthogonal projection of H onto /, constructed as in Re-
mark 3.7. Then, as I P is the orthogonal projection onto /
, and as e is an
orthonormal basis for /
, by Remark 3.7 (d), for all x H, we have

x = Px + (I P)x = Px +x, e)e.
Thus for all x H,
(x) = (Px) +x, e)(e) = x, (e)e) =
y
(x),
where y := (e)e. Hence =
y
, and is onto.
2
3.9. Remark. The fact that the map dened in the proof the Riesz Rep-
resentation Theorem above induces an isometric, conjugate-linear automorphism of
H is worth remembering.
3.10. Denition. A subset e
of a Hilbert space H is said to be or-

thonormal if |e
| = 1 for all , and ,= implies that e
, e
) = 0.
An orthonormal set in a Hilbert space is called an orthonormal basis for H if
it is maximal in the collection of all orthonormal sets of H, partially ordered with
respect to inclusion.
If E = e
is any orthonormal set in H, then an easy application of Zorns

Lemma implies the existence of a orthonormal basis in H which contains E. The
reader is warned that if H is innite-dimensional, then an orthonormal basis for H
is never a vector space (i.e. a Hamel) basis for H.
3.11. Example.
(a) If H =
2
then the standard Schauder basis e
n
n=1
for
2
as dened in
Example 2.13 is an orthonormal basis for H.
(b) If H = L
2
(R, dm), where T = z C : [z[ = 1 and dm represents
normalised Lebesgue measure, then f
n
nZ
is an orthonormal basis for
L
2
(T, dm), where f
n
(z) = z
n
for all z T and for all n Z.
We recall from Linear Algebra:
3. HILBERT SPACE 37
3.12. Theorem. The Gram-Schmidt Orthogonalisation Process
If H is a Hilbert space over K and x
n
n=1
is a linearly independent set in H,
then we can nd an orthonormal set y
n
n=1
in H so that spanx
1
, x
2
, ..., x
k
=
spany
1
, y
2
, ..., y
k
for all k 1.
Proof. We leave it to the reader to verify that setting y
1
= x
1
/|x
1
|, and recursively
dening
y
k
:=
x
k

k1
j=1
x
k
, y
j
)y
j
|x
k

k1
j=1
x
k
, y
j
)y
j
|
, k 1
will do.
2
3.13. Theorem. Bessels Inequality
If e
n
n=1
is an orthonormal set in a Hilbert space H, then for each x H,
n=1
[x, e
n
)[
2
|x|
2
.
Proof. For each k 1, let P
k
denote the orthogonal projection of H onto
spane
n
k
n=1
. Given x H, we have seen that |P
k
| 1, and that P
k
x =
k
n=1
x, e
n
)e
n
. Hence |x|
2
|P
k
x|
2
=

k
n=1
[x, e
n
)[
2
for all k 1, from which
the result follows.
2
3.14. Before considering a non-separable version of the above result, we pause
to dene what we mean by a sum over an uncountable index set.
Given a Banach space X and a set x
of vectors in X, let T denote the

collection of all nite subsets of , partially ordered by inclusion. For each F T,
dene y
F
=

F
x
, so that (y
F
)
FF
is a net in X. If y = lim
FF
y
F
exists, then
we say that y =

, and we say that x
is unconditionally summable
to y.
In other words,

= y if for all > 0 there exists F

0
T so that F F
0
implies that |
F
x
y| < .
3.15. Corollary. Let H be a Hilbert space and c H be an orthonormal set.
(a) Given x H, the set e c : x, e) , = 0 is countable.
(b) For all x H,

eE
[x, e)[
2
|x|
2
.
Proof.
(a) Fix x H. For each k 1, dene T
k
= e c : [x, e)[
1
k
. Suppose
that there exists k
0
1 so that T
k
0
is innite. Choose a countably innite
subset e
n
n=1
of T
k
0
. By Bessels Inequality,
m
k
2
0
=
m
n=1
1
k
2
0
n=1
[x, e
n
)[
2
|x|
2
for all m 1. This is absurd. Thus [T
k
[ < for all k 1. But then
e c : x, e) , = 0 =
k1
T
k
is countable.
(b) This is left as a (routine) exercise for the reader.
2
3.16. Lemma. Let H be a Hilbert space, c H be an orthonormal set, and
x H. Then
eE
x, e)e
converges in H.
Proof. Since H is complete, it suces to show that if T as in Paragraph 3.14
denotes the collection of nite subsets of c, partially ordered by inclusion, and if for
each F T, y
F
=

eF
x, e)e, then y
F
FF
is a Cauchy net.
Let > 0. By Corollary 3.15, we can nd a countable subcollection e
n
n=1
c
so that e ce
n
n=1
implies that x, e) = 0. Moreover, by Bessels Inequality, we
can nd N > 0 so that

k=N+1
[x, e
k
)[
2
<
2
. Let F
0
= e
1
, e
2
, ..., e
N
.
If F, G T and F
0
F, F
0
G, then
|y
F
y
G
|
2
= |
eF\G
x, e)e
eG\F
x, e)e|
2
=
e(F\G)(G\F)
[x, e)[
2
k=N+1
[x, e
k
)[
2
<
2
.
Thus (y
F
)
FF
is Cauchy, and therefore convergent, as required.
2
3.17. Theorem. Let c be an orthonormal set in a Hilbert space H. The fol-
lowing are equivalent:
(a) The set c is an orthonormal basis for H. (That is, c is a maximal or-
thonormal set in H.)
(b) The set spanc is norm-dense in H.
(c) For all x H, x =

eE
x, e)e.
(d) For all x H, |x|
2
=

eE
[x, e)[
2
. [Parcevals Identity]
Proof. Let c = e
be an orthonormal set in H.
(a) implies (b): Let / = spanc. If / ,= H, then /
,= 0, so we can nd
z /
, |z| = 1. But then c z is an orthonormal set, contradicting

the maximality of c.
3. HILBERT SPACE 39
(b) implies (c): Let y =

x, e
)e
, which exists by Lemma 3.16. Then

y x, e
) = 0 for all , so y x is orthogonal to / = spanc = H.

But then y x y x, so that y x = 0, i.e. y = x.
(c) implies (d): |x|
2
=
eE
x, e)e,
fE
x, f)f) =

eE
[x, e)[
2
. [Check!]
(d) implies (a): If e e
for all , then

|e|
2
=
[e, e
)[
2
= 0,
so that c is maximal.
2
3.18. Proposition. If H is a Hilbert space, then any two orthonormal bases
for H have the same cardinality.
Proof. We shall only deal with the innite-dimensional situation, since the nite-
dimensional case was dealt with in linear algebra.
Let c and T be two orthonormal bases for H. Given e c, let T
e
= f T :
e, f) , = 0. Then [T
e
[
0
. Moreover, given f T, there exists e c so that
e, f) , = 0, otherwise f is orthogonal to spanc = H, a contradiction.
Thus T =
e
T
e
, and so [T[ (sup
eE
[T
e
[) [c[
0
[c[ = [c[. By symmetry,
[c[ [T[, and so [c[ = [T[.
2
The above result justies the following denition:
3.19. Denition. The dimension of a Hilbert space H is the cardinality of
any orthonormal basis for H, and it is denoted by dim H.
The appropriate notion of isomorphism in the category of Hilbert spaces involve
linear maps that preserve the inner product.
3.20. Denition. Two Hilbert spaces H
1
and H
2
are said to be isomorphic
if there exists a linear bijection U : H
1
H
2
so that
Ux, Uy) = x, y)
for all x, y H
1
. We write H
1
H
2
to denote this isomorphism.
We also refer to the linear maps implementing the above isomorphims as unitary
operators. Note that
|Ux|
2
= Ux, Ux) = x, x) = |x|
2
for all x H
1
, so that unitary operators are isometries. Moreover, the inverse map
U
1
: H
2
H
1
dened by U
1
(Ux) := x is also linear, and
U
1
(Ux)U
1
(Uy)) = x, y) = Ux, Uy),
so that U
1
is also a unitary operator. Furthermore, if L H
1
is a closed subspace,
then L is complete, whence UL is also complete and hence closed in H
2
.
Unlike the situation in Banach spaces, where two non-isomorphic Banach spaces
can have Schauder bases of the same cardinality, the case of Hilbert spaces is as nice
as one can imagine.
3.21. Theorem. Two Hilbert spaces H
1
and H
2
over K are isomorphic if and
only if they have the same dimension.
Proof. Suppose rst that H
1
and H
2
are isomorphic, and let U : H
1
H
2
be
a unitary operator. Let e
be an orthonormal basis for H

1
. We claim that
Ue
is an orthonormal basis for H

2
.
Indeed, Ue
, Ue
) = e
, e
) =
,
(the Kronecker delta function)), |Ue
| =
|e
| = 1 for all , and

H
2
= UH
1
= U(spane
) = spanUe
,
by the continuity of U.
Hence dim H
2
= [Ue
[ = [[ = dim H
1
.
Conversely, suppose that dim H
2
= dim H
1
. Then we can nd a set and
orthonormal bases e
: for H
1
, and f
: for H
2
. Consider the map
U : H
1

2
()
x (x, e
))
,
where
2
() :=
_
f : K :

[f()[
2
<
_
. This is an inner product space
using the inner product f, g) =

f()g(). The proof that

2
() is complete
is essentially the same as in the case of
2
(N).
It is routine to check that U is linear. Moreover,
Ux, Uy) =
x, e
) y, e
)
=
x, e
)e
, y, e
)e
)
=
x, e
)e
y, e
)e
)
= x, y)
for all x, y H, and so |Ux|
2
= Ux, Ux) = x, x) = |x|
2
for all x H
1
. It follows
that U is isometric and therefore injective.
If (r

2
() has nite support, then x :=

H
1
and Ux = (r
.
Thus ranU is dense. But from the comment above, UH
1
is closed, and therefore
UH
1
=
2
().
We have shown that U is a unitary operator implementing the isomorphism of
H
1
and
2
(). By symmetry once again, there exists a unitary V : H
2

2
().
But then V
1
U : H
1
H
2
is unitary, showing that H
1
H
2
.
2
3. HILBERT SPACE 41
3.22. Corollary. The spaces
2
(N),
2
(Q),
2
(Z) and L
2
([0, 1], dx) (where dx
represents Lebesgue measure) are all isomorphic, as they are all innite dimensional,
separable Hilbert spaces.
3.23. When dealing with Hilbert space operators and operator algebras, one
tends to focus upon complex Hilbert spaces. One reason for this is that the spectrum
provides a terribly useful tool for analyzing operators. For X a normed linear space,
let I (or I
X
if we wish to emphasize the underlying space) denote the identity map
Ix = x for all x X.
3.24. Denition. Let X and Y be Banach spaces, and let T B(X, Y). We
say that T is invertible if there exists a (continuous) linear map R B(Y, X) such
that RT = I
X
and TR = I
Y
.
If T B(X), we dene the spectrum of T to be:
(T) = K : (T I) is not invertible.
3.25. When X is a nite-dimensional space, the spectrum of T coincides with
the eigenvalues of T. The reader will recall from their Linear Algebra courses that
eigenvalues of linear maps need not exist when the underlying eld is not alge-
braically closed. When K = C, it can be shown that the spectrum of an operator
T B(X) is a non-empty, compact subset of C, and a so-called functional calculus
which allows one to naturally dene f(T) for any complex-valued function which is
analytic in an open neighbourhood of (T). This, however, is beyond the scope of
the present notes.
If X is a Banach space, an operator Q B(X) is said to be quasinilpotent if
(Q) = 0. The argument of Paragraph 2.20 says that the Volterra operator is
quasinilpotent.
A wonderful theorem of A. Beurling, known as Beurlings spectral radius
formula relates the spectrum of an operator T to a limit of the kind obtained in
Paragraph 2.20.
Theorem. Beurlings Spectral Radius Formula.
Let X be a complex Banach space and T B(X). Then
spr(T) := max[k[ : k (T) = lim
n
|T
n
|
1
n
.
The quantity spr(T) is known as the spectral radius of T. It is worth pointing
out that an implication of Beurlings Spectral Radius Formula is that the limit on
the right-hand side of the equation exists! A priori, this is not obvious.
3. HILBERT SPACE 43
3.26. As mentioned in paragraph 3.1, Hilbert spaces arise naturally in the study
of Banach space geometry. In this context, much of the literature concerns real
Hilbert spaces.
For example, for each n 1, let Q
n
denote the set of n-dimensional (real)
Banach spaces. Given Banach spaces X and Y in Q
n
, we denote by GL(X, Y) the
set of all invertible operators from X to Y. We can dene a metric on Q
n
via:
(X, Y) := log
_
inf|T| |T
1
| : T GL(X, Y)
_
.
It can be shown that (Q
n
, ) is a compact metric space, known as the Banach-
Mazur compactum. One also refers to the quantity
d(X, Y) = inf|T| |T
1
| : T GL(X, Y)
as the Banach-Mazur distance between X and Y, and it is an important problem
in Banach space geometry to calculate Banach-Mazur distances between the n-
dimensional subspaces of two innite-dimensional Banach spaces, say V and W.
Typically, one is interested in knowing something about the asymptotic behaviour
of these distances as n tends to innity.
We mention without proof two interesting facts concerning the Banach-Mazur
distance:
(a) If X, Y and Z are n-dimensional Banach spaces, then
d(X, Z) d(X, Y) d(Y, Z).
(b) A Theorem of Fritz John shows that (with
2
n
:= (R
n
, | |
2
)),
d(X,
2
n
)
n for all n 1.
It clearly follows from these two properties that d(X, Y) n for all X, Y Q
n
.
3.27. We mentioned earlier in this section that the Parallelogram Law is useful
in determining that a given norm is not induced by an inner product. In fact, in
can be shown that a norm on a Banach space X is the norm induced by some inner
product if and only if the norm satises the Parallelogram Law.
My friends tell me I have an intimacy problem. But they dont really
know me.
Garry Shandling
4. Topological Vector Spaces
Someday I want to be rich. Some people get so rich they lose all
respect for humanity. Thats how rich I want to be.
Rita Rudner
4.1. Let H be an innite-dimensional Hilbert space. The norm topology on
B(H) is but one example of an interesting topology we can place on this set. We are
also interested in studying certain weak topologies on B(H) generated by a family of
functions. The topologies that we shall obtain are not induced by a metric obtained
from a norm. In order to gain a better understand of the nature of the topologies we
shall obtain, we now turn our attention to the notion of a topological vector space.
4.2. Denition. Let J be a vector space over the eld K, and let T be a
topology on J. We say that the topology T is compatible with the vector space
structure on J if the maps
: J J J
(x, y) x +y
and
: KJ J
(k, x) kx
are continuous, where KJ and J J carry their respective product topologies.
A topological vector space (abbreviated TVS) is a pair (J, T ) where J is
a vector space with a compatible Hausdor topology. Informally, we refer to J as
the topological vector space.
4.3. Remark. Not all authors require T to be Hausdor in the above denition.
However, for all spaces of interest to us, the topology will indeed be Hausdor.
Furthermore, one can always pass from a non-Hausdor topology to a Hausdor
topology via a natural quotient map. (See Appendix T.)
4.4. Example. Let (X, | |) be any normed linear space, and let T denote the
norm topology. Suppose (x, y) XX and > 0. Choose a net (x
, y
XX
so that lim
(x
, y
) = (x, y). Then there exists

0
so that
0
implies
|x
x| < /2, |y
y| < /2. But then

0
implies |x
+ y
(x + y)|
|x
x| + |y
y| < . In other words, (x
, y
) tends to (x, y) and so is

continuous.
Similarly, if (k, x) K X, then we can choose a net (k
, x
so that
lim
(k
, x
) = (k, x). But then we can nd

0
so that
0
implies [k
k[ <
1, and so [k
[ < [k[+1. Next we can nd

1
so that
1
implies [k
k[ < /2|x|,
4. TOPOLOGICAL VECTOR SPACES 45
and
2
so that
2
implies |x
x| < /2([k[ + 1). Choosing

0
,
1
and
2
we get
|(k
, x
) (k, x)| = |k
kx|
|k
x| +|k
x kx|
[k
[|x
x| +[k
k[|x|
/2 +/2 = .
This proves that is continuous. Hence X is a TVS with the norm topology.
4.5. Example. As a concrete example of the situation in Example 4.4, let
n 1 be an integer and for (x
1
, x
2
, ..., x
n
) C
n
, set |(x
1
, x
2
, ..., x
n
)|
= max[x
k
[ :
1 k n. Then C
n
is a TVS with the induced norm topology. Of course, in this
example, the norm topology coincides with the usual topology on C
n
coming from
the Euclidean norm |(x
1
, x
2
, ..., x
n
)|
2
=
_
n
k=1
[x
k
[
2
_1
2
, since | |
and | |
2
are
equivalent norms on C
n
.
4.6. Remark. In the assignments we shall see how to construct a TVS which is
not a normed linear space. See also the discussion of Frechet spaces in the Appendix
to Section 5.
4.7. Remark. Let (J, T ) be a TVS, and let U |
0
be a nbhd of 0 in J. The
continuity of : J J J implies that
1
(U) is open in J J. As such,
1
(U) contains a basic nbhd N
1
N
2
of (0, 0), where N
1
, N
2
|
0
are open (see
Appendix T). But if N = N
1
N
2
, then N |
0
and N N N
1
N
2

1
(U).
Thus for all U |
0
there exists N |
0
so that (N N) = N + N := m + n :
m, n N U.
Similarly, we can nd a nbhd V
(0) of 0 in K and N |
W
0
open so that V
(0)
N
1
(U), or equivalently,
kn : n N, 0 [k[ < U.
4.8. Denition. A nbhd N of 0 in a TVS J is called balanced if kN N
for all k K satisfying [k[ 1.
4.9. Example. Let (X, | |) be a normed linear space. For all > 0, V
(0) =
x X : |x| < is a balanced nbhd of 0.
4.10. Proposition. Let (J, T ) be a TVS. Every nbhd of 0 contains a balanced,
open nbhd of 0.
Proof. By Remark 4.7, given U |
0
, we can nd > 0 and N |
0
open such
that k K, 0 < [k[ < implies kN U. Since multiplication by a non-zero scalar
is a homeomorphism, each kN is open.
Let M =
0<|k|<
kN. Then M U and M

2
N, so M |
0
. A routine
calculation shows that M is balanced. It is also open, being the union of open sets.
2
4.11. Suppose (J, T ) is a TVS, w
0
J and k
0
K. Dene
w
0
: J J via
w
0
(x) = w
0
+x.
By continuity of addition, we get that
w
0
is continuous, and clearly
w
0
is a bijection.
But
1
w
0
=
w
0
is also a translation, and therefore it is continuous by the above
argument. That is,
w
0
is a homeomorphism.
This simple observation underlies a particularly useful fact about TVSs, namely:
N |
W
0
if and only if w
0
+N |
W
w
0
.
That is,
the nbhd system at any point in J is determined
by the nbhd system at 0.
If k
0
,= K, then
k
0
: J J dened by
k
0
(x) = k
0
x is also a continuous bijection
(by continuity of scalar multiplication) and has continuous inverse
k
1
0
. Thus
N |
W
0
if and only if k
0
N |
W
0
for all 0 ,= k
0
K.
The following result shows that the assumption that a TVS topology be Haus-
dor may be replaced with a weaker assumption - namely that points be closed (i.e.
that T be T
1
) - from which the Hausdor condition follows.
4.12. Proposition. Let 1 be a vector space with a topology T for which
(i) addition is continuous;
(ii) scalar multiplication is continuous; and
(iii) points in 1 are closed in the T -topology.
Then T is a Hausdor topology and (1, T ) is a TVS.
Proof. Let x ,= y 1. Then y is closed (i.e. 1y is open) and so we can
nd an open nbhd U |
x
of x so that y , U. As above, by continuity of addition,
translation is a homeomorphism of 1 and so U = x+U
0
for some open nbhd U
0
of 0.
Also by continuity of addition, there exists an open nbhd V of 0 so that V +V U
0
.
By continuity of scalar multiplication, V is again an open nbhd of 0, and hence
W = V (V ) is an open nbhd of 0 as well, with W +W V +V U
0
.
Suppose that (x + W) (y + W) ,= . Then there exist w
1
, w
2
W so that
x + w
1
= y + w
2
, i.e. x + w
1
w
2
= y. But w
1
w
2
W + W, so that y
x + (W + W) x + U
0
= U, a contradiction. Hence (x + W) |
x
, (y + W) |
y
are disjoint open nbhds of x and y respectively, and (1, T ) is Hausdor.
2
4.13. Proposition. Let (J, T ) be a TVS and be a linear manifold in J.
Then
(a) is a TVS with the relative topology induced by T ; and
(b) is a subspace of J.
Proof.
(a) This is clear, since the continuity of [
Y
and [
Y
is inherited from the
continuity of and .
(b) Suppose y, z and k K. Choose a net (y
, z
) so that
lim
(y
, z
) = (y, z). By continuity of on J J, y
+ z
y + z.
But y
+ z
for all , and so y + z . Similarly, if we choose a net

(k
, y
) in K which converges to (k, y), then the continuity of implies

that k
ky. Since k
for all , ky .
2
4.14. Exercise. Let (1, T ) be a TVS. Prove the following.
(a) If C 1 is convex, then so is C.
(b) If E 1 is balanced, then so is E.
4.15. Denition. Let (1, T ) be a TVS, and let (x
be a net in 1. We say
that (x
is a Cauchy net if for all U |

0
there exists
0
so that
1
,
2

0
implies that x
1
x
2
U.
We say that a subset K 1 is Cauchy complete if every Cauchy net in K
converges to some element of K.
We pause to verify that if (x
is a net in 1 which converges to some x 1,

then (x
is a Cauchy net. Indeed, let U |

0
and choose a balanced, open nbhd
N |
0
so that N + N = N N U. Also, choose
0
so that
0
implies
that x
x+N. If
1
,
2

0
, then x
1
x
2
= (x
1
x)(x
2
x) NN U.
Thus (x
is a Cauchy net.
4.16. Example. If (X, ||) is a normed linear space, then X is Cauchy complete
if and only if X is complete. Indeed, the topology here being a metric topology, we
need only consider sequences instead of general nets.
4.17. Lemma. Let 1 be a TVS and / 1 be complete. Then / is closed in
1.
Proof. Suppose that z /. For each U |
z
, there exists y
U
/ so that y
U
U.
The family U : U |
z
forms a directed set under reverse-inclusion, namely:
U
1
U
2
if U
2
U
1
. Thus (y
U
)
UUz
is a net in /. By denition, this net converges
to the point z, i.e. lim
U
y
U
= z. (Since 1 is Hausdor, this is the unique limit
point of the net (y
U
)
U
.) From the comments following Denition 4.15, (y
U
)
UUz
is
a Cauchy net. Since / is complete, it follows that z /, and hence that / is closed.
2
4.18. Let (1, T ) be a TVS and J be a closed subspace of 1. Then 1/J exists
as a quotient space of vector spaces. Let q : 1 1/J denote the canonical quotient
map.
We can establish a topology on 1/J using the T topology on 1 by dening a
subset G 1/J to be open if q
1
(G) is open in 1.
If G
1/J and q
1
(G
) is open in 1 for all , then q

1
(
) =
q
1
(G
) is open, and thus
is open in 1/J.
If G
1
, G
2
1/J and q
1
(G
i
) is open in 1, i = 1, 2, then q
1
(G
1
G
2
) =
q
1
(G
1
) q
1
(G
2
) is open in 1, when G
1
G
2
is open in 1/J.
Clearly = q
1
() and 1 = q
1
(1/J) are open in 1, so that , 1/J
are open in 1/J and the latter is a topological space.
We refer to this topology on 1/J as the quotient topology. The quotient
map is continuous with respect to the quotient topology, by design. In fact, the
quotient topology is the largest topology on 1/J which makes q continuous.
We begin by proving that q is an open map. That is, if G 1 is open, then
q(G) is open in 1/J.
Indeed, for each w J, the set G + w is open in 1, being a translate of the
open set G. Hence G+J =
wW
G+w is open in 1, being the union of open sets.
But
G+J = q
1
(q(G)),
so that q(G) is open in 1/J by denition.
To see that addition is continuous in 1/J, let x+J, y+J and let E be a nbhd
of (x + y) +J in 1/J. Then q
1
(E) is open in 1 and (x + y) q
1
(E). Choose
nbhds U
x
of x and U
y
of y in 1 so that r U
x
, s U
y
implies that r +s q
1
(E).
Note that x + J q(U
x
), y + J q(U
y
) and that q(U
x
), q(U
y
) are open in 1/J
by the argument above. If a +J q(U
x
), b +J q(U
y
), then a +J = g +J and
b +J = h +J for some g U
x
, h U
y
. Thus
(a +b) +J = (g +h) +J q(q
1
(E)) E.
Hence addition is continuous.
That scalar multiplication is continuous follows from a similar argument which
is left to the reader.
Finally, to see that the resulting quotient topology is Hausdor, it suces (by
Proposition 4.12) to show that points in 1/J are closed. Let x+J 1/J. Then
q
1
(x + J) = x + w : w J is closed in 1, being a translation of the closed
subspace J. Hence the complement C = 1x + w : w J of x +J is open in
1. But then q(C) is open, since q is an open map, and q(C) is the complement of
x +J.
Finite-dimensional topological vector spaces. Our present goal is to prove
that there is only one topology that one can impose upon a nite-dimensional vector
space 1 to make it into a TVS. We begin with the one dimensional case.
4.19. Lemma. Let (1, T ) be a one dimensional TVS over K. Let e be a
basis for 1. Then 1 is homeomorphic to K via the map
: K 1
k ke
.
Proof. The map is clearly a bijection, and the continuity of scalar multiplication
in a TVS makes it continuous as well. We shall demonstrate that the inverse map
1
(ke) = k is also continuous. To do this, it suces to show that if lim
e = 0,
then lim
= 0. (Why?)
Let > 0. Then e ,= 0, and as 1 is Hausdor, we can nd a nbhd U of 0 so
that e , U. By Proposition 4.10, U contains a balanced nbhd V of 0. Obviously,
e , V . Since lim
e = 0, there exists
0
so that
0
implies that k
e V .
Suppose that there exists
0
with [k
[ . Then
e =
_

k
_
k
e V,
as V is balanced. This contradiction shows that
0
implies that [k
[ < . Since
> 0 was arbitrary, we have shown that lim
= 0.
2
4.20. Proposition. Let n 1 be an integer, and let (1, T ) be an n-dimensional
TVS over K with basis e
1
, e
2
, ..., e
n
. The map
: K
n
1
(k
1
, k
2
, ..., k
n
)

n
j=1
k
j
e
j
is a homeomorphism.
Proof. Lemma 4.19 shows that the result is true for n = 1. We shall argue by
induction on the dimension of 1.
It is clear that is a linear bijection, and furthermore, since 1 is a TVS, the
continuity of addition and scalar multiplication in 1 implies that is continuous.
There remains to show that
1
is continuous as well.
Suppose that the result is true for 1 n < m. We shall prove that it holds for
n = m as well. To that end, let F = e
t
1
, e
t
2
, ..., e
tr
for some 1 r < m, and let
E = e
1
, e
2
, ..., e
m
F = e
p
1
, e
p
2
, ..., e
ps
.
Now = spane
j
: j E is an s-dimensional space with s < m. By our
induction hypothesis, the map (k
1
, k
2
, ..., k
s
)
s
j=1
k
j
e
p
j
is a homeomorphism. It
follows that is complete (check! ) and therefore closed, by Lemma 4.17. By the
arguments of paragraph 4.18, 1/ is a TVS and the canonical map q
Y
: 1 1/
is continuous. Moreover, q(e
t
1
), q(e
t
2
), ..., q(e
tr
) is a basis for 1/. Since r < m,
our induction hypothesis once again shows that the map
Y
: 1/ K
r
r
j=1
k
j
q(e
t
j
) (k
1
, k
2
, ..., k
r
)
is continuous. Thus
:=
Y
q : 1 K
r
n
i=1
k
i
e
i

Y
(
n
i=1
k
i
q(e
i
)) = (k
t
1
, k
t
2
, ..., k
tr
)
is also continuous, being the composition of continuous functions.
To complete the proof, we rst apply the above argument with F = e
m
to get
that
1
: 1 K
n
i=1
k
i
e
i
k
m
is continuous, and then to F = e
1
, e
2
, ..., e
m1
to get that
2
: 1 K
m1
n
i=1
k
i
e
i
(k
1
, k
2
, ..., k
m1
)
is continuous. Since
1
= (
1
,
2
), it too is continuous, and we are done.
2
The previous result has a number of important corollaries:
4.21. Corollary. Let 1 be a TVS and J be a nite-dimensional subspace of
1. Then J is closed in 1.
Proof. This argument is embedded in the proof of the previous result; J is com-
plete because of the nature of the homeomorphism between J and K
n
, where n is
the dimension of J. Then Lemma 4.17 implies that J is closed.
2
4.22. Corollary. Let n 1 be an integer and 1 be an n-dimensional vector
space. Then there is a unique topology T which makes 1 a TVS. In particular,
therefore, all norms on a nite dimensional vector space are equivalent.
Proof. Since any topology on 1 which makes it a TVS is determined completely
by the product topology on K
n
, it is unique. If | |
1
and | |
2
are two norms on 1,
then they induce metric topologies which make 1 into a TVS. But these topologies
coincide, from the above argument. By Proposition 1.17, the norms are equivalent.
2
4.23. Denition. Let (1, T
V
) and (J, T
W
) be topological vector spaces and
suppose that f : 1 J is a (not necessarily linear) map. We say that f is
uniformly continuous if, given U |
W
0
there exists N |
V
0
so that x y N
implies that f(x) f(y) U.
4.24. The denition of uniform continuity given here derives from the fact that
the collection B = B(U) = (x, y) : x y U : U |
W
0
denes what is
known as a uniformity on the TVS (J, T
W
) whose corresponding uniform topology
coincides with the initial topology T . The interested reader is referred to the book
by Willard [Wil70] and to the books of Kadison and Ringrose [KR83] for a more
complete development along these lines. We shall focus only upon that part of the
theory which we require in this text.
4.25. Let us verify that in the case where (X, | |
X
) and (Y, | |
Y
) are normed
linear spaces, our new notion of uniform continuity coincides with our metric space
notion.
Observe rst that if (Z, | |) is a general normed linear space and > 0, then
x y V
Z
(0) if and only if |x y| < .

Suppose f : X Y is uniformly continuous in the sense of Denition 4.23.
Given > 0, V
Y
(0) = y Y : |y|
Y
< |
Y
0
and so there exists N |
X
0
such
that x y N implies f(x) f(y) V
Y
(0). But N |
X
0
implies that there exists
> 0 so that V
X
(0) N. Thus |x y|
X
< implies that x y N, and thus
f(x) f(y) V
Y
(0), i.e. |f(x) f(y)|

Y
< . Thus is the standard (metric) notion
of uniform continuity in a normed space.
Conversely, suppose that f : X Y is uniformly continuous in the standard
metric sense. Let U |
Y
0
. Then there exists > 0 so that V
Y
(0) U. By
hypothesis, there exists > 0 so that |xy|
X
< implies |f(x) f(y)|
Y
< , and
hence xy V
X
(0) implies that f(x)f(y) V

Y
(0) U. That is, f is continuous

in the sense of Denition 4.23.
It is useful to extend our notion of uniformly continuous functions between
topological vectors spaces to functions dened only upon a subset (not necessarily
a subspace) of the domain space 1.
4.26. Denition. If (1, T
V
) and (J, T
W
) are topological vector spaces and
C 1, then f : C J is uniformly continuous if for all U |
W
0
there exists
N |
V
0
such that x, y C and x y N implies f(x) f(y) U.
4.27. Example. Now (R, [ [) is a normed linear space, and so the comments of
paragraph 4.25 apply. The function f : (0, 1) R dened by f(x) = x
2
is therefore
uniformly continuous, whereas g : R R dened by g(x) = x
2
is not.
Another way in which uniform continuity in the TVS setting extends the notion
of uniform continuity in the metric setting is evinced by the following:
4.28. Proposition. Let (1, T
V
) and (J, T
W
) be topological vector spaces and
f : 1 J be uniformly continuous. Then f is continuous on 1.
Proof. Let x
0
1. Let U |
W
f(x
0
)
. Then by paragraph 4.11, U = f(x
0
) + U
0
where U
0
|
W
0
. By hypothesis, there exists N
0
U
V
0
so that x x
0
N
0
implies
that f(x) f(x
0
) U
0
. That is, x x
0
+N
0
implies f(x) f(x
0
) +N
0
= U. Since
N := x
0
+ N
0
|
V
x
0
, we see that f is continuous at x
0
. But x
0
1 was arbitrary,
and so f is continuous on 1.
2
4.29. Theorem. Let (1, T
V
) and (J, T
W
) be topological vector spaces over K.
Suppose that T : 1 J is linear. The following are then equivalent:
(a) there exists x
0
1 so that T is continuous at x
0
; and
(b) T is uniformly continuous on 1.
Proof. By Proposition 4.28, it suces to prove that (a) implies (b). To that end,
suppose that T is continuous at x
0
and let U
0
|
W
0
. Then U := Tx
0
+U
0
|
W
Tx
0
.
By continuity of T at x
0
, there exists N |
V
x
0
so that T(N) U. But N = x
0
+N
0
for some N
0
|
V
0
. Now if z N
0
, then x
0
+z N, and so T(x
0
+z) = Tx
0
+Tz
Tx
0
+U
0
. That is, Tz U
0
.
In particular, if xy N
0
, then T(xy) = TxTy U
0
, and so T is uniformly
continuous on 1.
2
4.30. Given vector spaces 1 and J over C, denote by 1
R
and J
R
the same
spaces of vectors, viewed as vector spaces over R. Observe that if (1, T
V
) and
(J, T
W
) are topological vector spaces over C and T : 1 J is conjugate-linear
(i.e. T(kx) = k x for all x 1 and k C), then T : 1
R
J
R
is linear (over R).
That is, T(kx+y) = kTx+Ty for all x, y 1
R
and k R. Moreover, the topologies
T
V
and T
W
are real-vector space topologies for 1
R
and J
R
respectively (as well as
C-vector space topologies for 1 and J).
From this it follows that T : 1 J is continuous if and only if T : (1
R
, T
V
)
(J
R
, T
W
) is continuous, and by the above Theorem, this happens precisely when T
is uniformly conitnuous on 1 (or equivalently on 1
R
, since all scalars appearing in
the denition of uniform continuity are real).
In other words, Theorem 4.29 holds for as conjugate-linear maps too.
4.31. Corollary. Let (1, T
V
) and (J, T
W
) be topological vector spaces over K
and suppose that dim 1 = n < . If T : 1 J is linear, then T is continuous.
Proof. By Theorem 4.29 and Proposition 4.28, it suces to prove that T is con-
tinuous at 0. Let e
1
, e
2
, ..., e
n
be a basis for 1, and suppose that (x
is a net
in 1 which converges to 0. For each we may express x
as a unique linear
combination of the e
j
s, say
x
= k
,1
e
1
+k
,2
e
2
+ +k
,n
e
n
.
By Proposition 4.20, lim
= 0 implies that for 1 j n, lim
k
,j
= 0.
Now Tx
=

n
j=1
k
,j
Te
j
, . But scalar multiplication in (J, T
W
) is
continuous, and lim
k
,j
= 0, so
lim
Tx
=
n
j=1
lim
k
,j
Te
j
=
n
j=1
0 Te
j
= 0 = T(lim
).
It follows that T is continuous at 0, as was required.
2
If we restrict our attention to subsets of 1 we get:
4.32. Proposition. Let 1, J be topological vector spaces and T : 1 J be
linear. Suppose that 0 C 1 is balanced and convex. If T[
C
is continuous at 0,
then T[
C
is uniformly continuous.
Proof. Our assumption is that T[
C
is continuous at 0, and thus for all U |
W
0
,
there exists N |
V
0
so that x C N implies Tx
1
2
U.
By Proposition 4.10, every nbhd N |
V
0
contains a balanced nbhd N
0
, and so
by replacing N by N
0
if necessary, we may assume that N is balanced. Note that
1
2
N |
V
0
.
Now suppose that x, y C and that x y N. Then C balanced implies
that y C, and C convex implies that
1
2
x +
1
2
(y) =
1
2
(x y) C. Since N is
balanced,
1
2
(x y) N and so T(
1
2
(x y)) =
1
2
(Tx Ty)
1
2
U. Hence x, y C,
x y N implies Tx Ty U. That is, T is uniformly continuous on C.
2
If A, B, and C are sets with A B, and if f : A C is a map, the we say that
the map g : B C extends f (or that g is an extension of f) if g[
A
= f.
4.33. Proposition. Suppose that 1 and J are topological vector spaces and
that J is Cauchy complete. If A 1 is a linear manifold and T
0
: A J is
continuous and linear, then T
0
extends to a continuous linear map T : A J.
Proof. Let x A and choose (x
1
in A so that lim
= x. Clearly, if T is to
be continuous, we shall need Tx = lim
Tx
. The issue is whether or not this limit

exists and is independent of the choice of (x
.
Now (x
is a Cauchy net. Take U U

W
0
. Since T
0
is continuous, there exists
N |
X
0
such that w N implies that T
0
w U. Hence, there exists
0
such
that
1
,
2

0
implies that x
1
x
2
N and therefore that T
0
x
1
T
0
x
2
=
T
0
(x
1
x
2
) U. Thus (T
0
x
is also a Cauchy net. Our assumption that J is

Cauchy complete implies that there exists z (depending a priori upon (x
) such
that z = lim
T
0
x
.
Suppose that (y
2
X and that lim
= x. Arguing as above, there exists

z
2
J so that z
2
= lim
T
0
y
. If we set
y
(,)
:= y
,
1
x
(,)
:= x
,
2
,
and =
1
2
, then
lim
(,)
x
(,)
= lim
(,)
y
(,)
= x.
Also, lim
(,)
T
0
x
(,)
= z
1
, lim
(,)
T
0
y
(,)
= z
2
. Thus lim
(,)
x
(,)
y
(,)
=
0 A and so by the continuity of T
0
on A,
0 = T
0
0 = T
0
( lim
(,)
x
(,)
y
(,)
)
= lim
(,)
T
0
(x
(,)
y
(,)
)
= z
1
z
2
.
That is, we can set Tx = lim
T
0
x
and this is well-dened.

That T is linear on A is left as an exercise.
Finally, to see that T is continuous on A, let U |
W
0
and choose U
1
|
W
0
so
that U
1
+ U
1
U. Choose N |
X
0
so that x N implies Tx = T
0
x U
1
. Then
N = G A for some G |
V
0
.
Let M = G A so that M |
X
0
. If z M, then z = lim
for some x
N,
. Now Tz = lim
T
0
x
, so that there exists

0
so that
0
implies that
Tz T
0
x
U
1
.
But T
0
x
U
1
for all and so Tz = (Tz T
0
x
) +T
0
x
U
1
+U
1
U. That
is, z M implies that Tz U. Hence T is continuous at 0, and consequently T is
uniformly continuous on A.
2
4.34. Corollary. Suppose that X and Y are Banach spaces and that M X is
a linear manifold. If T
0
: M Y is bounded, then T
0
extends to a bounded linear
map T : MY, and |T| = |T
0
|.
Proof. Invoking Proposition 4.33, there remains only to show that |T| = |T
0
|.
That |T| |T
0
| is clear.
Conversely, given x M with |x| = 1 and > 0, there exists y M so that
|y| = 1 and |x y| < . Then
|T
0
y Tx| = |Ty Tx| |T| |y x| < |T|.
(Recall that T is bounded since T is continuous!) Since > 0 was arbitrary,
sup|T
0
y| : y M, |y| = 1 sup|Tx| : |x| = 1,
so |T
0
| |T|, completing the proof.
2
Do you know what it means to come home at night to a woman wholl
give you a little love, a little aection, a little tenderness? It means
youre in the wrong house, thats what it means.
Henny Youngman
5. SEMINORMS AND LOCALLY CONVEX SPACES 55
5. Seminorms and locally convex spaces
The secret of life is honesty and fair dealing. If you can fake that,
youve got it made.
Groucho Marx
5.1. Our main interest in topological vector spaces is to develop the theory
of locally convex topological vector spaces, which appear naturally in dening cer-
tain weak topologies naturally associated with Banach spaces, including the Banach
space of all bounded operators on a Hilbert space. Locally convex spaces are also
the most general spaces for which (in our opinion) interesting versions of the Hahn-
Banach Theorem will be shown to apply. As we shall see in this section, there is an
intimate relation between locally convex topological vector spaces and separating
families of seminorms on the underlying vector spaces, a phenomenon to which we
now turn our attention.
5.2. Denition. Let 1 be a vector space over K. A seminorm on 1 is a map
p : 1 R satisfying
(i) p(x) 0 for all x 1;
(ii) p(x) = [[p(x) for all x 1, K;
(iii) p(x +y) p(x) +p(y) for all x, y 1.
It follows from this denition that a norm on 1 is simply a seminorm which
satises the additional property that p(x) = 0 if and only if x = 0.
5.3. Remark. A few remarks are in order. If p is a seminorm on a vector space
1, then for all x, y 1,
p(x +y) p(x) +p(y)
implies that
p(x +y) p(y) p(x).
Equivalently, with z = x+y, p(z) p(x) p(z x). Thus p(x) p(z) p(xz) =
p(z x). Hence
[p(x) p(z)[ p(z x).
5.4. Example. Let 1 = (([0, 1], C). For each x [0, 1], the map
p
x
: 1 R
dened by setting p
x
(f) = [f(x)[ is a seminorm on 1 which is not a norm.
5.5. Example. Let n 1 and consider 1 = M
n
= M
n
(C). Fix 1 k, l n.
The map
kl
: 1 R dened by
kl
([x
ij
]) = [x
kl
[ denes a seminorm on 1 which,
once again, is not a norm.
5.6. Convexity. Recall that a subset E of a vector space 1 is said to be convex
if x, y E and 0 t 1 imply tx+(1 t)y E. Geometrically, we are asking that
the line segment between any two points in E must lie in E.
It is a simple but useful fact that any linear manifold of 1 is necessarily convex.
Note also that if p
1
, p
2
, ...., p
m
is a family of seminorms on a 1, x
0
1, > 0
and E = x 1 : p
j
(x x
0
) < , 1 j m, then E is convex. Indeed, if x, y E
and 0 t 1, then for all 1 j m,
p
j
(tx + (1 t)y x
0
) = tp
j
(x x
0
) + (1 t)p
j
(y x
0
)
< t + (1 t) = .
Thus tx + (1 t)y E, and E is convex.
We leave it as an exercise for the reader to show that if 1 is a TVS and E 1
is convex, then so is E.
Another elementary but useful observation is that if C 1 is convex and
T : 1 J is a linear map (where J is a second vector space), then T(C) is
convex as well. Finally, if E 1 is convex, then for all r, s > 0, rE+sE = (r +s)E.
Indeed,
r
r+s
e
1
+
s
r+s
e
2
E for all e
1
, e
2
E, from which the desired result easily
follows.
5.7. The Minkowski functional. Let 1 be a TVS and suppose that E |
V
0
is convex. Now for any x 1, lim
r0
rx = 0 (by continuity of multiplication), and
thus there exists r
0
> 0 so that r
0
x E. This allows us to dene the map
p
E
: 1 R
x infr (0, ) : x rE,
which we call the gauge functional or the Minkowski functional for E.
Note: the name is misleading, since the map is clearly not linear - its range is
contained in [0, ). By convexity of E, if x rE and 0 < r < s, then x = re for
some e E, so x = (1
r
s
)0 +
r
s
(se) co(sE) = sE. In particular, x sE for all
s > p
E
(x).
5.8. Denition. Let 1 be a vector space over K. A function p : 1 R is
called a sublinear functional if it satises:
(i) p(x +y) p(x) +p(y) for all x, y 1, and
(ii) p(rx) = rp(x) for all 0 < r R.
5.9. It is clear from the denition that every seminorm (and hence every norm)
on a vector space is a sublinear functional on that space. The converse is false in
general.
For example, the identity map : R R is a sublinear functional on R. It is
not a seminorm since it is not even a non-negative valued function.
5.10. Proposition. Let J be a TVS and E |
0
be convex. Then
(a) The Minkowski functional p
E
is a sublinear functional on J for E.
(b) If E is open, then
E = x J : p
E
(x) < 1.
(c) If E is balanced, then p
E
is a seminorm.
Proof.
(a) Suppose that x, y E and that r, s (0, ) with r > p(x), s > p(y). Then
x rE, y sE and so x +y (r +s)E. That is,
p(x +y) r +s for all r > p(x), s > p(y).
Thus p(x +y) p(x) +p(y).
Also, if k > 0, then x rE if and only if kx krE, so that
p(kx) = infs : kx sE
= infkr : kx krE
= k infr : x rE
= kp(x).
Thus p is a sublinear functional, as claimed.
(b) Suppose that x E and that E is open. Since the map f : R J
given by f(t) = tx is continuous, 1 f
1
(E) is open in R and therefore
(1 , 1 + ) f
1
(E) for some > 0. But then (1 +

2
)x E, or
equivalently, x
2
2+
E, implying that p(x) <
2
2+
< 1.
Conversely, suppose that p(x) < 1. Then x = re for some p(x) < r < 1
and e E. But then x = (1 r)0 +re coE = E.
(c) Suppose now that E is balanced. First observe that if k ,= 0, then
k
|k|
E = E.
Note that p is subadditive since it is a sublinear functional by (a). Also,
p(x) 0 for all x J by denition of p.
Finally, if k = 0, then p(kx) = p(0). But 0 rE for all r > 0, and so
p(0x) = p(0) = infr > 0 : x rE = 0 = 0p(x).
If k ,= 0, then
p(kx) = infr > 0 : kx rE
= infs[k[ > 0 : kx s([k[E)
= infs[k[ > 0 : kx s(kE)
= [k[ infs > 0 : x sE
= [k[p(x).
Thus p is a seminorm.
2
5.11. Proposition. Let J be a TVS and p be a seminorm on J. The following
are equivalent:
(a) p is continuous on J;
(b) there exists a set U |
W
0
such that p is bounded above on U.
Proof.
(a) implies (b): It follows from our observations in paragraph 5.6 that the set
0 E := x J : p(x) < 1
is convex. Since p is assumed to be continuous and E = p
1
(, 1), E is
also open. Thus p is bounded above (by 1) on the open set E |
W
0
.
(b) implies (a): Suppose that p is bounded above, say by M > 0 on an open
set U |
W
0
. Let > 0. If x, y J and x y (

M
)U, say x y =

M
u
for some u U, then
[p(x) p(y)[ p(x y) = p(

M
u) =

M
p(u) < .
Thus p is (uniformly) continuous on J.
2
5.12. Example. Recall from Example 5.4 that for each x [0, 1],
p
x
: (([0, 1], C) R
f [f(x)[
is a seminorm. Now B
1
(0) := f (([0, 1], C) : |f|
< 1 is open and f B

1
(0)
implies that p
x
(f) = [f(x)[ |f|
< 1.
Thus each such p
x
is continuous on (([0, 1], C).
5.13. Denition. A topology T on a topological vector space J is said to be
locally convex if it admits a base consisting of convex sets. We shall write LCS
for locally convex topological vector spaces, and for the sake of brevity, we shall refer
to them as locally convex spaces.
Since the topology on J is determined by the nbhds at a single point, it suces
to require that J admit a nbhd base at 0 consisting of convex sets; that is, given
any nbhd U |
0
, there exists a convex nbhd N |
0
so that N U. In verifying
that a space is a LCS, we shall often only verify this condition.
5.14. Proposition. Let J be a TVS, and suppose that U |
0
is convex. Then
U contains a balanced, open, convex nbhd of 0.
Proof. By Proposition 4.10, U contains a balanced, open nbhd H of 0. Set N =
co(H). Then U convex and H U implies that N U. Since H is balanced, a
routine calculation shows that N is also balanced. For any choice of t
1
, t
2
, ..., t
m

[0, 1] with

m
k=1
t
k
= 1, and for any h
1
, h
2
, ..., h
m
H we have
_
m1
k=1
t
k
h
k
_
+t
m
H N.
Since H is open, so is
_
m1
k=1
t
k
h
k
_
+t
m
H. Since N = co(H), it follows that N is
a union of open sets of this form, and hence N is also open.
Thus N is an open, balanced, convex nbhd of 0 contained in U, the existence of
which proves our claim.
2
As an immediate consequence we obtain:
5.15. Corollary. Let (1, T ) be a LCS. Then 1 admits a nbhd base at 0 con-
sisting of balanced, open, convex sets.
5.16. Example. Let (X, | |) be a normed linear space. For each > 0, the
argument of paragraph 5.6 shows that B
(0) = x X : |x| < is convex. Since

B
(0) : > 0 is a nbhd base at 0 for the norm topology, (X, | |) is a LCS.
More concretely, (K
n
, | |
2
) is a LCS, as is any Hilbert space H. So is B(H).
We have already seen that the quotient of a TVS by one of its closed subspaces
is a TVS. Let us rst obtain the same result for locally convex spaces.
5.17. Proposition. Let (1, T ) be a LCS and J 1 be a closed subspace.
Then 1/J is a LCS in the quotient topology.
Proof. As mentioned above, that 1/J is a TVS follows from paragraph 4.18.
There remains only to show that 1/J admits a nbhd base at 0 consisting of convex
sets (see the remarks following Denition 5.13).
Let q : 1 1/J denote the canonical quotient map, and let U |
V/W
0
. Then
q
1
(U) |
V
0
, as q is continuous. Since 1 is a LCS, we can nd a convex nbhd
N |
V
0
so that 0 N q
1
(U). Let M = q(N). Since q is an open map, we have
M |
V/W
0
. Since q is linear, M is convex.
Finally, since N q
1
(U), M = q(N) U, and we are done.
2
5.18. Denition. A family of seminorms on a vector space J is said to be
separating if for all 0 ,= x J there exists p so that p(x) ,= 0.
5.19. Example. Let J = (([0, 1], C) and consider = p
x
: x Q [0, 1],
where - as before - p
x
(f) = [f(x)[ for all f J.
If 0 ,= f J, then there exists y [0, 1] so that f(y) ,= 0. By continuity of f,
there exists a nbhd N of y such that f(y) ,= 0 for all y N. Thus there exists a
rational number q N so that 0 ,= f(q) and hence p
q
(f) ,= 0. Thus is a separating
family of seminorms.
5.20. Let be a family of seminorms on a vector space J. For F nite,
x J and > 0, set
N(x, F, ) = y J : p(x y) < , p F.
Permitting ourselves a slight abuse of notation, we shall write N(x, p, ) in the case
where F = p.
5.21. Theorem. If is a separating family of seminorms on a vector space
J, then
B = N(x, F, ) : x J, > 0, F nite
is a base for a locally convex topology T on J. Moreover, each p is T -
continuous.
Proof.
Step One: We begin by showing that B is a base for a Hausdor topology T on
J.
Let x J and choose 0 ,= p . (Such a p exists since is assumed to
be separating.) Then x N(x, p, 1). Thus
B : B B N(x, p, 1) : x J = J.
Next suppose that B
1
= N(x, F
1
,
1
) and B
2
= N(y, F
2
,
2
) lie in B and
that z B
1
B
2
. We must nd B
3
B so that z B
3
B
1
B
2
.
To that end, let = min
1
p(x z),
2
q(y z) : p F
1
, q F
2
,
so that > 0. If w N(z, F
1
F
2
, ), then
p(w x) p(w z) +p(z x) < +p(z x)
1
for all p F
1
, and so w B
1
. An analogous argument proves that w B
2
.
That is, B
3
:= N(z, F
1
F
2
, ) satises the required condition.
It now follows from our work in the homework assignments that B is a
base for a topology on J.
If x, y J and x ,= y, then our assumption that is separating implies
the existence of an element p so that := p(x y) > 0. But then
N(x, p,

2
) and N(y, p,

2
) are disjoint nbhds of x and y respectively in the
T -topology, proving that T is Hausdor.
Step Two: That T is locally convex follows readily from the fact that B is a base
for T and each N(x, F, ) is itself convex, as is easily veried.
Step Three: Next we verify that (J, T ) is a TVS; namely, that the topology T
is compatible with the vector space operations.
Suppose that x
0
, y
0
J and let U be a nbhd of x
0
+y
0
in the T -topology.
Then there exists a basic nbhd B = N(x
0
+y
0
, F, ) of x
0
+y
0
with B U.
Let B
1
= N(x
0
, F,

2
) and B
2
= N(y
0
, F,

2
). If (x, y) B
1
B
2
, then
p
_
(x +y) (x
0
+y
0
)
_
p(x x
0
) +p(y y
0
) <

2
+

2
=
for all p F, and thus (B
1
B
2
) B U. This shows that addition is
continuous relative to T .
As for scalar multiplication, let
0
K, x
0
J and U be a nbhd of
0
x
0
in the T -topology. As before, choose a basic nbhd B = N(
0
x
0
, F, ) U.
Let > 0. If K := C : [
0
[ < and B = N(x
0
, F, ), then
(, x) K B implies that
p(x
0
x
0
) p(x x
0
) +p(x
0
0
x
0
)
[[p(x x
0
) +[
0
[p(x
0
)
< ([
0
[ +) +p(x
0
)
for all p F. Since F is nite, it is clear that can be chosen such that
p(x
0
x
0
) < , p F, which proves that scalar multiplication is also
continuous relative to T .
Together, these two observations prove that (J, T ) is a TVS.
Step Four: Finally, let us show that each p is continuous relative to T .
If p = 0, then clearly p is continuous relative to T .
Otherwise, let B = N(0, p, 1). Then B T , and for x B,
p(x) = p(x 0) < 1,
so that p is bounded on some open set in J. It now follows from Proposition 5.11
that p is (uniformly) continuous on J.
2
5.22. The above result says that a separating family of seminorms on a vector
space J gives rise to a locally convex topology on J. Our next goal is to show that
all locally convex spaces arise in this manner.
5.23. Theorem. Suppose that (1, T
V
) is a LCS. Then there exists a separating
family of seminorms on 1 which generate the topology T
V
.
Proof. By Corollary 5.15, (1, T
V
) admits a nbhd base (
0
at 0 consisting of balanced,
open, convex sets. By Proposition 5.10, for each E (
0
, the Minkowski functional
p
E
is a seminorm and E = x 1 : p
E
(x) < 1.
Let = p
E
: E (
0
. We rst show that is separating. Indeed, suppose
that 0 ,= x 1. Since T
V
is Hausdor by hypothesis, there exists G (
0
so that
x , G (this is actually a bit weaker than the statement that T
V
is Hausdor, but
certainly implied by it). Since G (
0
, p
G
. But x , G implies that p
G
(x) 1,
and hence p
G
(x) ,= 0. Thus is separating. This is required before passing to the
next step.
By Theorem 5.21,
B = N(x, F, ) : x 1, > 0, F nite
is a base for a locally convex topology T
on 1. Our goal, of course, is to prove that

T
V
= T
.
Let E (
0
be a T
V
-open, balanced, convex nbhd of 0. Since E = N(0, p
E
, 1) B,
it follows that T
contains a nbhd base at 0 for the topology T

V
. Since both topologies
are TVS-topologies, they are determined by their nbhd bases at any point (for eg.,
at 0), and from this it follows that T
T
V
.
On the other hand, each p
E
is bounded above by 1 on E, and E is a T
V
-
open nbhd of 0. By Proposition 5.11, p
E
is continuous on (1, T
V
). It follows that
N(0, p
E
, ) = p
1
E
(, ) T
V
for all > 0. Thus T
V
contains a nbhd subbase for
T
at 0, and arguing as before, we get that T
T
V
.
Hence T
V
= T
, and the topology T

V
is determined by the family of seminorms.
2
5.24. Example. Let (X, | |) be a normed linear space. The norm topology on
X is the metric topology induced by the metric d(x, y) = |x y|. That is, a nbhd
base at x
0
X for the norm topology is
B
x
0
= V
(x
0
) : > 0
= y X : |y x
0
| < : > 0
= N(x
0
, | |, ) : > 0.
Thus we see that the norm topology on X is exactly the locally convex topology
generated by = | |. Observe that since | | is a norm, 0 ,= x X implies that
|x| , = 0, and thus is indeed separating, as required.
5.25. In Corollary 5.15, we saw that any LCS (1, T ) admits a nbhd base at 0
consisting of open, balanced, convex sets.
In fact, each N(0, p
1
, p
2
, ..., p
m
, ) is balanced, open and convex for all choices
of m 1, p
1
, p
2
, ..., p
m
and > 0, where is a separating family of seminorms
which generate T . It is clear from Theorems 5.21 and 5.23 that the collection of
such sets is a nbhd base at 0 for T .
Having generated a topology on a vector space using a separating family of
seminorms, let us now examine what it means for a net to converge in this topology.
5.26. Proposition. Let 1 be a vector space and be a separating family of
seminorms on 1. Let T denote the locally convex topology on 1 generated by .
A net (x
in 1 converges to a point x 1 if and only if

lim
p(x x
) = 0 for all p .
Proof.
Suppose rst that (x
converges to x in the T -topology. Given p

and > 0, the set N(x, p, ) T and so there exists
0
so that
0
implies that x
N(x, p, ). That is,

0
implies that p(x x
) < .
Thus lim
p(x x
) = 0.
Alternatively, one may argue as follows: suppose that (x
converges
to x in the T -topology. Given p , we know that p is continuous in the
T -topology by Theorem 5.21. Since lim
x x
= 0,
lim
p(x x
) = p(lim
(x x
)) = p(0) = 0.
Conversely, suppose that lim
p(x x
) = 0 for all p . Let U |

x
is the T -topology. Then there exist p
1
, p
2
, ..., p
m
and > 0 so that
N(x, p
1
, p
2
, ..., p
m
, ) U. For each 1 j m, choose
j
so that
j
implies that p
j
(x
x) < . Choose
0

1
,
2
, ...,
m
. If
0
, then
p
j
(x
x) < for all 1 j m so that x
N(x, p
1
, p
2
, ..., p
m
, ) U.
Hence lim
= x in (1, T ).
2
5.27. Remarks. Let 1 be a vector space as above and let be a separating
family of seminorms on 1. Recall that if T
w
is the weak topology on 1 induced by ,
then T
w
is the weakest topology for which each of the functions p is continuous.
By Theorem 5.21, the LCS topology T generated by
B = N(x, F, ) : x 1, > 0, F nite
has the property that each p is continuous on (1, T ). It follows, therefore, that
T
w
T . In other words, if (x
is a net in (1, T ) which converges to x 1, then

(x
converges to x in (1, T
w
); i.e. lim
p(x
) = p(x) for all p .

That these two topologies do not, in general, coincide can be seen by examining
a simple example.
Let 1 = K and let = p, where p(x) = [x[ for each x K. The LCS topology
on K generated by is a TVS topology, and thus must agree with the usual topology
on K, since the latter admits a unique TVS topology, by Lemma 4.19. The weak
topology T
w
on K generated by is the weakest topology for which p is continuous.
In particular, a net (x
converges to x K if and only if lim
[x
[ = [x[. For
example, the sequence (x
n
)
n
, where x
n
= (1)
n
, n 1 converges to x = 1 in
(K, T
w
). Since it clearly doesnt converge in (K, T ), the two topologies are necessarily
dierent, and again by Theorem 5.21 it follows that (K, T
w
) is not a TVS.
There is, however, a situation where we can say a bit more than this. Let 1
be a vector space and let (X
, | |
)
A
be a collection of Banach spaces (in fact,
normed linear spaces will do). For each such , suppose that T
: 1 X
is a
linear map. Suppose furthermore that the family T
is separating in the sense

that if 0 ,= x 1, then there exists A so that 0 ,= T
x X
. Then each of the

functions
p
: 1 [0, )
x |T
x|
is easily seen to be a seminorm. It is routine to verify that the fact that T
is separating implies that = p
is a separating family of seminorms. Let T

denote the LCS topology on 1 generated by . By Proposition 5.26, a net (x
converges to x (1, T ) if and only if

lim
(x x
) = lim
|T
(x x
)|
= 0 for all A.
That is, lim
= x if and only if
lim
= T
x for all A.
Since this is nothing more than the statement that each T
is continuous, we nd
that in this case, the T topology on 1 coincides with the weak topology generated
by the family T
A
. This is still not the same as the weak topology generated
by the family , however.
5.28. Example. Let H =
2
(N) and recall that H is a Hilbert space when
equipped with the inner product (x
n
), (y
n
)) =

n=1
x
n
y
n
.
Recall also that B(H) is a normed linear space with the operator norm |T| :=
sup|Tx| : x H, |x| 1.
From above, we see that the norm topology on B(H) admits as a nbhd base at
T B(H) the collection
N(T, | |, ) : > 0 = V
(T) : > 0,
and that this is the locally convex topology generated by the separating family
= | | of (semi)norms.
Convergence of a net of operators (T
in the norm topology is uniform con-

vergence - i.e. lim
|T
T| = 0.
This is certainly not the only interesting topology one can impose upon B(H).
Let us rst consider the topology of pointwise convergence.
The strong operator topology (SOT) For each x H, consider
p
x
: B(H) R
T |Tx|.
Then
(i) p
x
(T) 0 for all T B(H);
(ii) p
x
(T) = |Tx| = [[ |Tx| = [[ p
x
(T) for all K;
(iii) p
x
(T
1
+T
2
) = |T
1
x +T
2
x| |T
1
x| +|T
2
x| = p
x
(T
1
) +p
x
(T
2
),
so that p
x
is a seminorm on B(H) for each x H.
In general, p
x
is not a norm because we can always nd T B(H) so that 0 ,= T
but p
x
(T) = 0. Indeed, let y H with 0 ,= y and y x. Dene T
y
: H H via
T
y
(z) = z, y)y. Then |T
y
(z)| |z| |y|
2
by the Cauchy-Schwarz Inequality and
in particular T
y
(y) = |y|
2
y ,= 0, but T
y
(x) = x, y)y = 0y = 0. Thus 0 ,= T
y
but
p
x
(Ty) = 0.
On the other hand, if 0 ,= T B(H), then there exists x H so that Tx ,= 0.
Thus p
x
(T) = |Tx| , = 0, proving that
SOT
:= p
x
: x H separates the points
of B(H).
The locally convex topology on B(H) generated by
SOT
is called the strong
operator topology and is denoted by SOT.
By Proposition 5.26 above, we see that a net (T
B(H) converges to T
B(H) in the SOT if and only if
lim
p
x
(T
T) = lim
|T
x Tx| = 0 for all x H.

Thus the SOT is the topology of pointwise convergence. That is, it is the weakest
topology that makes all of the evaluation maps T Tx, x H continuous.
A nbhd base for the SOT at the point T B(H) is given by the collection
N(T, x
1
, x
2
, ..., x
m
, ) : m 1, x
j
H, 1 j m, > 0 =
R B(H) : |Rx
j
Tx
j
| < , 1 j m.
The weak operator topology (WOT) Next, for each pair (x, y) H H,
consider the map
q
x,y
: B(H) R
T [Tx, y)[.
Again, it is routine to verify that each q
x,y
is a seminorm but not a norm on B(H).
The locally convex topology on B(H) generated by
WOT
:= q
x,y
: (x, y)
HH is called the weak operator topology on B(H) and is denoted by WOT.
A net (T
B(H) converges to T B(H) in the WOT if and only if

lim
[(T
T)x, y)[ = [T
x, y) Tx, y)[ = 0
for all x, y H. In other words, the WOT is the weakest topology that makes all
of the functions T Tx, y), x, y H continuous.
A nbhd base for the WOT at the point T B(H) is given by the collection
N(T, x
1
, x
2
, ..., x
m
, y
1
, y
2
, ..., y
m
, ) : m 1, x
j
, y
j
H, 1 j m, > 0 =
R B(H) : [Rx
j
Tx
j
, y
j
)[ < , 1 j m.
5.29. Proposition. Let (1, T ) be a LCS, and let be a separating family
of seminorms on 1 which generate the locally convex topology on 1. Let p be a
seminorm on 1. The following are equivalent:
(a) p is continuous on 1;
(b) there exists a constant > 0 and p
1
, p
2
, ..., p
m
so that
p(x) max(p
1
(x), p
2
(x), ..., p
m
(x)) for all x 1.
Proof.
(a) implies [(b)] Suppose that p is continuous on 1. Then M := p
1
((1, 1)) =
p
1
([0, 1)) is a T -open nbhd of 0, and as such, it must contain a basic nbhd
N := N(0, p
1
, p
2
, ..., p
m
, ) for some p
1
, p
2
, ..., p
m
and > 0. It
follows that if p
j
(x) < for 1 j m, then x N M, and hence
p(x) < 1.
More generally, let y 1 and let r = max(p
1
(y), p
2
(y), ..., p
m
(y)).
If r = 0, then for all k > 0, p
j
(ky) = 0 < , 1 j m, so that from
above, p(ky) = kp(y) < 1. But then
p(y) = 0 1 max(p
1
(y), p
2
(y), ..., p
m
(y)).
If r > 0, then x =

2r
y satises p
j
(x) < , 1 j m, and so
2r
p(y) = p(x) < 1. That is,
p(y) <
2r
=
2
max(p
1
(y), p
2
(y), ..., p
m
(y)).
We conclude that with = max(1,
2
),
p(y) max(p
1
(y), p
2
(y), ..., p
m
(y))
for all y 1.
(b) implies [(a)] Suppose that (b) holds. Now N := N(0, p
1
, p
2
, ..., p
m
, 1) is
an open nbhd of 0 in the T -topology. If x N, then p
j
(x) < 1 for all
1 j m, and so p(x) . But then p is bounded above on the T -open
nbhd N of 0, and hence is continuous by Proposition 5.11.
2
V
) and (J, T
W
) be locally convex spaces. Let
V
and
W
denote separating families of seminorms which generate the corresponding
locally convex topologies on 1 and J respectively. Finally, let T : 1 J be a
linear map.
The following are equivalent:
(a) T is continuous.
(b) For all q
W
there exists > 0 and p
1
, p
2
, ..., p
m

V
so that
q(Tx) max(p
1
(x), p
2
(x), ..., p
m
(x)) for all x 1.
Proof.
(a) implies (b): Suppose that T is continuous and that q
W
. Clearly q is
continuous as well. It is routine to verify that qT is a seminorm on 1. Since
the composition of continuous functions is continuous, q T is a continuous
seminorm on 1, and the result now follows from Proposition 5.29.
(b) implies (a): Conversely, suppose that for all q
W
there exists > 0 and
p
1
, p
2
, ..., p
m

V
so that
q(Tx) max(p
1
(x), p
2
(x), ..., p
m
(x)) for all x 1.
As before, we observe that qT is a seminorm on 1 for all q J. Moreover,
by Proposition 5.29, each such q T is continuous.
Let U |
W
0
and choose q
1
, q
2
, ..., q
n

W
so that
N(0, q
1
, q
2
, ..., q
n
, ) U. Since each q
j
T is continuous on 1, we have
that N(0, q
1
T, q
2
T, ..., q
n
T, ) is a nbhd of 0 in 1. Moreover,
x N(0, q
1
T, q
2
T, ..., q
n
T, )
implies that
Tx N(0, q
1
, q
2
, ..., q
n
, ) U.
It follows that T is continuous at 0.
By Theorem 4.29 and paragraph 4.30, T is continuous on 1.
2
We shall require the following special case of the above result.
5.31. Corollary. Let (1, T ) be a LCS. A linear functional f on 1 is continuous
if and only if there exists a continuous seminorm q on 1 such that
[f(x)[ q(x) for all x 1.
Proof. Observe that if f is a continuous linear functional on 1, then q(x) := [f(x)[,
x 1 denes a continuous seminorm on 1; indeed, that q is continuous follows from
the fact that f is continuous on 1 and [ [ is continuous on K respectively. Obviously
[f(x)[ q(x) for all x 1.
Conversely and more interestingly suppose that there exists a continuous
seminorm q on 1 such that [f(x)[ q(x) for all x 1. As before, we may choose
a separating family of seminorms on 1, and without loss of generality, we may
assume that q . (Otherwise we replace by q.). The result now follows
immediately from Proposition 5.30.
2
5.32. In the assignment questions we exhibited an example of a TVS which is
not normable, i.e. it is not a normed linear space with respect to any norm. The
technique for constructing that example can be extended to produce a large variety
of such examples. The spaces we have in mind are called Frechet spaces, and we
dene them now.
5.33. Denition. A metric d on a vector space 1 is said to be translation
invariant if
d(x, y) = d(x +z, y +z)
for all x, y, z 1.
We shall also say that a metric d on 1 is complete if (1, d) is a complete metric
space.
Finally, let us say that a countable family
n
n
of pseudo-metrics on 1 is
complete if, whenever (x
k
)
k
is a sequence in 1 which is Cauchy relative to each
n
(i.e. for all n 1 and > 0 there exists N = N(, n) > 0 so that j, k N implies
n
(x
j
, x
k
) < ), there exists x 1 so that lim
k
n
(x
k
, x) = 0 for each n 1.
5.34. Example. Most, but certainly not all metrics we deal with are translation
invariant. For example, if d(x, y) = [xy[ for x, y R, then d is obviously translation
invariant.
On the other hand, the metric d on R dened via:
d(x, y) = [x
3
y
3
[
for all x, y R is not translation invariant, since d(0, 1) = 1 ,= 7 = d(1, 2).
5.35. Denition. Let (1, T ) be a LCS. If the topology T on 1 is induced by
a translation invariant, complete metric d, then we say that (1, T ) is a Frechet
space.
5.36. Constructing Frechet spaces. We know from Theorem 5.21 that if
is a separating family of seminorms on a vector space 1, then generates a
LCS topology T on 1. Suppose now that the family possesses the following two
additional properties, namely:
the set = p
n
n
is countable, and
the family
n
n
of pseudo-metrics dened on 1 via
n
(x, y) = p
n
(x y)
is a complete family.
Then the metric
d(x, y) =
n=1
1
2
n
p
n
(x y)
1 +p
n
(x y)
=
n=1
1
2
n
n
(x, y)
1 +
n
(x, y)
is easily seen to be translation-invariant. It is not too dicult to verify that a
sequence (x
k
)
k
in 1 converges to x 1 relative to the metric topology induced
by d if and only if lim
k
p
n
(x
k
x) = 0 for all n 1. That is, the d-metric
topology coincides with the LCS topology induced by . Furthermore, observe that
(x
k
)
k
is Cauchy in the d-metric topology if and only if (x
k
)
k
is Cauchy relative to
each pseudo-metric
n
, n 1. By the second item above, it follows that (1, d) is
complete, and hence that (1, T ) is a Frechet space.
5.37. Example.
(a) Let 1 = (
(R) denote the vector space of all functions f : R R which

are innitely dierentiable at each point x R. Let = p
n,k
n,k0
, where
for f 1,
p
n,k
(f) := sup[f
(n)
(x)[ : x [k, k].
Let T denote the LCS topology on 1 generated by the separating family
of seminorms. Then (1, T ) is a Frechet space.
A sequence (f
k
)
k
in 1 converges to f 1 if and only if
lim
j
sup[f
(n)
j
(x) f
(n)
(x)[ : x [k, k] = 0
for all n 0, k 0.
(b) If (X, | |) is a normed linear space, then with = | |, X becomes a
Frechet space.
5.38. Many authors dene a Frechet space as a LCS with a translation-invariant
metric which is complete as a uniform topological space. The denition of a
uniform space is rather long, and instead we refer the interested reader to the book
of Willard [Wil70] for a development of this concept.
I once spent a year in Philadelphia. I think it was on a Sunday.

W.C. Fields
6. The Hahn-Banach theorem
When I wake up in the morning, I just cant get started until Ive had
that rst, piping hot pot of coee. Oh, Ive tried other enemas...
Emo Philips
6.1. It is somewhat of a misnomer to refer to the Hahn-Banach Theorem. In
fact, there is a large number of variations on this theme. These variations fall into
two groups: the separation theorems, and the extension theorems. The crucial rela-
tion between these two classes of theorems is that they all refer to linear functionals.
Having said this, when one wishes to apply a version of the Hahn-Banach Theorem,
one tends to say only: by the Hahn-Banach Theorem..., usually leaving it to the
reader to determine which version of the Theorem is being applied.
The importance of these theorems in Functional Analysis can not be overstated.
6.2. Denition. Let J be a vector space over K. A linear functional on
J is a linear map f : J K. The vector space of all linear functionals on J is
denoted by J
#
and is referred to as the algebraic dual of J.
If J is a TVS, the (vector) space of continuous linear functionals is denoted
by J
, and is referred to as the (topological) dual of J. Obviously J
J
#
.
6.3. Example. Let n 1 be an integer and consider J = K
n
equipped with
the norm |(x
1
, x
2
, ..., x
n
)|
= max [x
j
[. For any choice of k
1
, k
2
, ..., k
n
K, the
map
f : J K
(x
1
, x
2
, ..., x
n
)

n
i=1
k
i
x
i
is a continuous linear functional.
6.4. Remarks.
(a) Recall from basic linear algebra that every linear functional on K
n
is of this
form for some choice of k
1
, k
2
, ..., k
n
K. As such, every linear functional
on K
n
is continuous.
(b) Recall from Proposition 4.20 that if 1 is an n-dimensional TVS with basis
e
1
, e
2
, ..., e
n
, then 1 is homeomorphic to K
n
via the map
n
i=1
k
i
e
i
(k
1
, k
2
, ..., k
n
). Since the product topology on K
n
is in turn
equivalent to the norm topology induced by the innity norm, it follows
from (a) above that every linear functional on a nite-dimensional TVS is
continuous.
6. THE HAHN-BANACH THEOREM 71
6.5. Example. Let us next consider c
00
(K) = (x
n
)
n=1
: x
n
K for all n
1 and x
n
= 0 for all but nitely many ns. Recall that this forms a normed linear
space when equipped with the norm
|(x
n
)
n
|
= sup
n
[x
n
[.
Dene
f : c
00
(K) K
(x
n
)
n

n=1
x
n
.
Then f is a non-continuous linear functional on c
00
(K). Indeed, if
y
n
= (
1
n
,
1
n
, ...,
1
n
, 0, 0, 0, ...)
(where the
1
n
term is repeated n times), then |y
n
|
=
1
n
, and so lim
n
y
n
= 0,
while f(y
n
) = 1 for all n, and hence lim
n
f(y
n
) ,= 0 = f(0).
For a number of the results we shall obtain below, we shall assume that the
underlying eld is R. In order to translate the results to the case of complex vector
spaces, the following Lemma will be useful.
6.6. Lemma. Let 1 be a vector space over C.
(a) If f : 1 R is an R-linear functional, then the map
f
C
(x) := f(x) if(ix)
is a C-linear functional on 1, and f = Re f
C
.
(b) If g : 1 C is C-linear, f = Re g and f
C
is dened as in (a), then g = f
C
.
(c) If p is a C-seminorm on 1 and f, f
C
are as in (a) above, then [f(x)[ p(x)
for all x 1 if and only if [f
C
(x)[ p(x) for all x 1.
(d) If 1 is a NLS and f, f
C
are as in (a), then |f| = |f
C
|.
Proof.
(a) This is routine and is left to the reader.
(b) Let x 1 and write g(x) = a + ib, where a = Re g(x) = f(x) and b =
Img(x) are real. By C-linearity of g, g(ix) = ig(x) = b +ia, and so
Img(x) = b = Re g(ix) = f(ix).
That is, g(x) = f(x) +i(f(ix)) = f(x) if(ix) = f
C
(x).
(c) First suppose that [f
C
(x)[ p(x) for all x 1. Then [f(x)[ = [Re f
C
(x)[
[f
C
(x)[ p(x) for all x 1.
Next suppose that [f(x)[ p(x) for all x 1. Given x 1, choose
C, [[ = 1 so that [f
C
(x)[ = f
C
(x). Then
[f
C
(x)[ = f
C
(x)
= f
C
(x)
= Re f
C
(x) (as this quantity is non-negative)
= f(x)
p(x)
= [[p(x) = p(x).
(d) It is routine to verify that |f| |f
C
|, and this step is left to the reader.
Conversely, given x 1 with |x| = 1, we can nd
x
so that [f
C
(x)[ =
x
f
C
(x) = f
C
(
x
x) = Re f
C
(
x
x) = f(
x
x). Note that |
x
x| = 1 because
1 is a C-vector space and |
x
x| = [
x
[ |x| = |x| = 1. Thus
|f
C
| = sup[f
C
(z)[ : |z| = 1
= sup[f(
z
z)[ : |z| = 1
sup[f(y)[ : |y| = 1
= |f|.
2
6.7. Proposition. Let 1 be a vector space over K and let f 1
#
.
(a) If g 1
#
and g[
ker f
= 0, then g = kf for some k K.
(b) If g, f
1
, f
2
, ..., f
N
1
#
and g(x) = 0 for all x
N
j=1
ker f
j
, then
g spanf
1
, f
2
, ..., f
N
.
Proof.
(a) If g = 0, then set k = 0 and we are done.
Otherwise, choose z 1 so that g(z) ,= 0. By hypothesis, f(z) ,= 0.
Let k = g(z)/f(z). Now ker f has codimension 1 in 1, and so if x 1,
then x = z +y for some y ker f and K. Hence
g(x) = g(z) +g(y) = g(z) + 0
= kf(z)
= k(f(z) +f(y))
= kf(x).
Since x 1 was arbitrary, g = kf.
(b) We may assume that f
1
, f
2
, ..., f
N
are linearly independent. Let ^ =
N
j=1
ker f
j
. Then dim(1/^) N. For 1 j N, dene f
j
: 1/^ K
via f
j
(x + ^) = f
j
(x). Since ^ ker f
j
, each f
j
is well-dened, and
f
j
(1/^)
#
.
We claim that f
1
, f
2
, ..., f
N
is also linearly independent. Otherwise,
we can nd k
1
, k
2
, ..., k
N
K so that

N
j=1
[k
j
[ ,= 0, but

N
j=1
k
j
f
j
= 0.
But then

N
j=1
k
j
f
j
,= 0, so we can nd z 1 with 0 ,=

N
j=1
k
j
f
j
(z) =
N
j=1
k
j
f
j
(z +^), a contradiction.
Thus dim(1/^)
#
N. Combining this with the fact that dim(1/^)
N yields dim(1/^) = dim(1/^)
#
= N, and that f
1
, f
2
, ..., f
N
is a ba-
sis for (1/^)
#
.
Now dene g : 1/^ K via g(x+^) = g(x). Again, since ^ ker g,
g is well-dened. Since g (1/^)
#
, we can write
g =
N
j=1
k
j
f
j
for some k
1
, k
2
, ..., k
N
K.
For x 1,
0 = (g
N
j=1
k
j
f
j
)(x +^)
= g(x)
N
j=1
k
j
f
j
(x),
so that g =

N
j=1
k
j
f
j
.
2
The rst part of the above Proposition shows that if f and g are distinct linear
functionals on a vector space 1, then they have the same kernel if and only if one
functional is a non-zero multiple of the other.
6.8. Denition. Let 1 be a vector space over K. A hyperplane / in 1 is a
linear manifold for which dim(1//) = 1.
6.9. If 0 ,= 1
#
, then from elementary linear algebra theory we see that
/ := ker is a hyperplane in 1 and that induces an (algebraic) isomorphism
between 1// and K via
(x +/) := (x) for all x +/ 1//.
Conversely, if / 1 is a hyperplane, then 1//is (algebraically) isomorphic to
K. Let : 1//K denote such an isomorphism. If q : 1 1//is the canonical
quotient map, then q : 1 K is a linear functional with ker ( q) = /.
Thus we have established a correspondence between linear functionals and hy-
perplanes. Proposition 6.7 implies that, up to a factor of a non-zero scalar multiple,
this correspondence is bijective.
6.10. Proposition. If (1, T ) is a TVS and / 1 is a hyperplane, then either
/ is closed in 1, or / is dense.
Proof. Since / is a vector space satisfying / / 1, and since dim(1//) =
1, we either have /= /, or /= 1.
2
It is worth noting that both possibilities can occur.
6.11. Example.
(a) Let X = ((([0, 1], C), | |
), and let 1
2
: X C be the map 1
2
(f) := f(
1
2
),
f X. Then ker 1
2
= f : [0, 1] C[f is continuous and f(
1
2
) = 0. This
is clearly closed.
(b) Let 1 = c
0
(C), and let e
k
= (
1k
,
2k
,
3k
, ...), where
ij
=
_
0 if i ,= j
1 if i = j
.
Let z = (1, 1/2, 1/3, 1/4, ...). Then z, e
1
, e
2
, e
3
, ... is linearly independent
in 1, and as such it can be extended to a Hamel basis (i.e. a vector space
basis) for 1, say
B = z, e
1
, e
2
, e
3
, ... b
: .
Given v 1, say v = z +
k=1
k
e
k
+
for some ,
k
,

K with only nitely many coecients not equal to zero, dene g(v) = .
It is clear that this denes a linear functional on 1. Since ker g is a
subspace containing e
k
, k 1, and since spane
k
k=1
is dense in 1, ker g
is dense in 1. Since g ,= 0, and since ker g is dense in 1, we see that ker g
is not closed.
6.12. Proposition. Let 1 be a TVS and 1
#
. Suppose that there exists an
open nbhd U |
0
of 0 and a constant > 0 so that Re (x) < for all x U.
Then is uniformly continuous on 1.
Proof. By Proposition 4.10, we can nd a balanced, open nbhd N of 0 with N U.
Observe that for x N U, there exists
x
K, [
x
[ = 1 so that
[(x)[ = (
x
x) = Re (
x
x).
But
x
x N since N is balanced, and so [(x)[ < for x N. Consider the
function
p : 1 R
x [(x)[
which is easily seen to be a seminorm. Since p is bounded above by on the open
nbhd N of 0, we can invoke Proposition 5.11 to conclude that p is continuous on 1.
By linearity, it follows that is continuous at 0, and hence is uniformly continuous
on 1 by Theorem 4.29.
2
6.13. Corollary. Let 1 be a TVS and 1
#
. The following are equivalent:
(a) is continuous on 1 - i.e. 1
;
(b) ker is closed.
Proof.
(a) implies (b): This is clear. If is continuous, then ker =
1
(0) is closed
in 1 because 0 is closed in K.
(b) implies (a): Suppose next that ker is closed. If = 0, then is obviously
continuous. Suppose therefore that ,= 0. Then J := 1/ ker is a one-
dimensional TVS and
: J K
x + ker (x)
is a linear functional on J. By Remark 6.4 (b), is continuous. If
q : 1 J is the canonical quotient map, then by Paragraph 4.18, q
is also continuous, and thus = q is continuous as well.
2
Recall that X
= B(X, K) is a Banach space whenever X is a NLS. Let us recall

a couple of results from Measure Theory which provide us with interesting examples
of classes of linear functionals.
6.14. Theorem. Let (X, , ) be a measure space and 1 < p < . If
1
p
+
1
q
= 1,
and if g L
q
(X, , ), then
g
(f) :=
_
X
fgd
denes a continuous linear functional on L
p
(X, , ), and the map g
g
is an
isometric linear bijection of L
q
(X, , ) onto L
p
(X, , )
.
If (X, , ) is -nite, then the same conclusion holds in the case where p = 1
and q = .
Recall that if X is a locally compact space, then M
K
(X) denotes the space of
K-valued regular Borel measures on X with the total variation norm.
6.15. Theorem. If X is locally compact and M
K
(X), then
: (
0
(X, K) K
f
_
X
fd
denes an element of (
0
(X, K), and the map
is an isometric linear isomor-

phism of M
K
(X) onto (
0
(X, K)
.
The Extension Theorems
6.16. The Hahn-Banach Theorem is probably the most important result in
Functional Analysis. It has a great many applications, and its usefulness cannot
be overstated. There are two basic formulations of this result (each with a variety
of consequences); the rst in terms of extensions of linear functionals from linear
submanifolds of a LCS to the entire LCS, and the second in terms of so-called
separation theorems, which we shall examine later.
6.17. Proposition. Let 1 be a vector space over R and p : 1 R be a
sublinear functional. Suppose that / is a (proper) hyperplane and that f : /R
is a linear functional for which f(x) p(x) for all x /. Then there exists a
linear functional g : 1 R such that g[
M
= f, and g(x) p(x) for all x 1.
Proof. Let z 1//, so that 1 = spanz, /. Then v 1 implies that v = tz+m
for some t R, m /.
For each r R we may dene h
r
: 1 R by setting h
r
(z) = r, setting
h
r
(m) = f(m), m /, and then extending h
r
by linearity to all of 1. Clearly
h
r
1
#
and h
r
extends f. The problem is that we do not know that h
r
(x) p(x)
for all x 1 - in fact, this is generally not true. The question of nding a g as
in the statement of the Proposition amounts to showing that for some s R, we
will have h
s
(x) p(x) for all x 1. To nd such an s, we rst examine which
properties it must satisfy. We then demonstrate that these properties are also
sucient. Finally, the existence of s is a byproduct of reconciling these necessary
and sucient conditions.
If h
s
(x) p(x) for all x 1, then for all t R, m / we must have
h
s
(tz +m) = ts +f(m) p(tz +m).
If t > 0, then setting m
1
= t
1
m yields:
s t
1
f(m) +t
1
p(tz +m)
f(t
1
m) +p(z +t
1
m) for all m /
f(m
1
) +p(z +m
1
) for all m
1
/. (1)
If t < 0, then setting m
2
= t
1
m yields:
s t
1
f(m) +t
1
p(tz +m)
= f(t
1
m) p(z t
1
m) for all m /
= f(m
2
) p(z +m
2
) for all m
2
/. (2)
The key issue is that we can reverse engineer this process. Suppose that s R
satises both (1) and (2), namely
(3) f(m
2
) p(z +m
2
) s f(m
1
) +p(z +m
1
) for all m
1
, m
2
/.
If t > 0, then
h
s
(tz +m) = ts +f(m)
t(f(m/t) +p(z + (m/t))) +f(m)
= p(tz +m) for all m /,
while
if t < 0, then
h
s
(tz +m) = ts +f(m)
t(f(m/t) p(z (m/t))) +f(m)
= f(m) + (t)p(z (m/t)) +f(m)
= p(tz +m) for all m /.
If t = 0, then h
s
(tz + m) = h
s
(m) = f(m) p(m) = p(tz + m) for all
m /.
There remains to show, therefore, that we can nd s R which satises (3) (or
equivalently, which satises both (1) and (2)). Now this can be done if
f(m
2
) +f(m
1
) p(z +m
2
) +p(z +m
1
) for all m
1
, m
2
/.
But
f(m
1
) +f(m
2
) = f(m
1
+m
2
) p(m
1
+m
2
)
p(m
2
z) +p(z +m
1
)
for all m
1
, m
2
/, and so we can choose
s
0
:= supf(m
2
) p(z +m
2
) : m
2
/.
Letting g = h
s
0
completes the proof.
2
6.18. Theorem. The Hahn-Banach Theorem 01
Let 1 be a vector space over R and let p be a sublinear functional on 1. If /
is a linear manifold in 1 and f : / R is a linear functional with f(m) p(m)
for all m /, then there exists a linear functional g : 1 R with g[
M
= f, and
g(x) p(x) for all x 1.
Proof. Let = (^, h) : ^ a linear manifold in 1, / ^, h[
M
= f, and h(n)
p(n) for all n ^. For (^
1
, h
1
), (^
2
, h
2
) , dene (^
1
, h
1
) _ (^
2
, h
2
) if
^
1
^
2
and h
2
[
N
1
= h
1
. Then (, _) is a partially ordered set with respect
to _. Moreover, , = , since (/, f) .
Let ( = (^
, h
) : be a chain in , and let ^ :=
. Dene
h : ^ R by setting h(n) = h
(n) if n ^
. Then h is well-dened because (

is a chain (check!), h is linear and h(n) p(n) for all n ^. Thus (^, h)
and it is an upper bound for (. By Zorns Lemma, (, _) has a maximal element
(, g). Suppose that ,= 1. Choosing z 1 and letting
0
= spanz, ,
Proposition 6.17 implies the existence of a functional g
0
:
0
R which extends g
and satises g
0
(y) p(y) for all y
0
. This contradicts the maximality of (, g).
Hence = 1, and g has the required properties.
2
The complex version of this theorem can now be established.
Let 1 be a vector space over K. Let / 1 be a linear manifold and let p : 1 R
be a seminorm on 1. If f : / K is a linear functional and [f(m)[ p(m) for
all m /, then there exists a linear functional g : 1 K so that g[
M
= f and
[g(x)[ p(x) for all x 1.
Proof. Suppose that K = R. Then f(m) [f(m)[ p(m) for all m /, and
p is a sublinear functional (by virtue of the fact that it is a seminorm). By the
Hahn-Banach Theorem 01, there exists g : 1 R linear so that g[
M
= f and
g(x) p(x) for all x 1. Thus g(x) = g(x) p(x) = p(x) for all x 1, so
that [g(x)[ p(x) for all x 1.
Now suppose that K = C. Let f
1
= Ref. Then by Lemma 6.6, [f
1
(m)[ p(m)
for all m /. By the argument of the rst paragraph of this proof, there exists
an R-linear functional g
1
: 1 R so that g
1
[
M
= f
1
and [g
1
(m)[ p(m) for all
m /. Let g = (g
1
)
C
denote the complexication of g
1
, as obtained in Lemma 6.6.
Then g : 1 C is C-linear, g[
M
= f, and by part (c) of that Lemma,
[g(x)[ p(x) for all x 1.
2
6.20. Corollary. Let (1, T ) be a LCS and J 1 be a linear manifold. If
f J
, then there exists g 1
so that g[
W
= f.
Proof. Since (1, T ) is a LCS, so is J. Let be a separating family of seminorms
which generate the LCS topology on 1 (see Theorem 5.23). Then it is routine to
verify that
W
:= p[
W
: p is a separating family of seminorms on J which
generates the relative LCS topology on J.
Suppose that f J
. By Corollary 5.31, there exist > 0 and p

1
, p
2
, ..., p
m

so that
[f(w)[ max(p
1
(w), p
2
(w), ..., p
m
(w)) for all w J.
Let q(x) := max(p
1
(x), p
2
(x), ..., p
m
(x)) for all x 1. Then, as is easily veried,
q is a seminorm on 1, and q is continuous by Proposition 5.29. Moreover,
[f(w)[ q(w) for all w J.
By the Hahn-Banach Theorem 02, we can nd a linear functional g : 1 K so that
g[
W
= f and
[g(x)[ q(x) for all x 1.
Another application of Corollary 5.31 shows that g is continuous, as was required.
2
The following is simply an application of Corollary 6.20 to the context of normed
linear spaces. It is often the version that comes to mind when the Hahn-Banach
Theorem is quoted.
Let (X, | |) be a NLS, / X be a linear manifold, and f /
be a bounded
linear functional. Then there exists g X
such that g[
M
= f and |g| = |f|.
Proof. Consider the map
p : X R
x |f| |x|
.
It is easy to check that p is a seminorm on X. (In fact, it is a norm unless f = 0.)
Since [f(m)[ p(m) for all m /, it follows from the Hahn-Banach Theorem 02
that there exists g : X K so that g[
M
= f and [g(x)[ p(x) = |f| |x| for all
x X. This last inequality shows that |g| |f|. That |g| |f| is clear, and
hence |g| = |f|.
2
6.22. Corollary. Let (1, T ) be a LCS and x
j
m
j=1
be a linearly independent
set of vectors in 1. If k
j
m
j=1
K are arbitrary, then there exists g 1
so that
g(x
j
) = k
j
, 1 j m.
Proof. Let / = spanx
j
m
j=1
, so that / is a nite-dimensional subspace of 1.
Dene f : /K via
f(
m
j=1
a
j
x
j
) =
m
j=1
a
j
k
j
.
Then f is linear on /, and thus, by Corollary 4.31, it is continuous. By Corol-
lary 6.20, there exists g 1
so that g[
M
= f. Hence g(x
j
) = k
j
, 1 j m.
2
A special case of the above Corollary which is worth pointing out is the following.
6.23. Corollary. Let (1, T ) be a LCS and 0 ,= y 1. Then there exists g 1
so that g(y) ,= 0.
Proof. Simply let x
1
= y and k
1
= 1 in the previous Corollary.
2
As an application of these results, let us show that nite dimensional subspaces
of locally convex spaces are topologically complemented.
6.24. Denition. A closed subspace J of a LCS (1, T ) is said to be topolog-
ically complemented if there exists a closed subspace of 1 so that 1 = J.
That is, x 1 implies that x = y+w for some y and w J, while J = 0.
6.25. Remark. Every vector subspace W of a vector space V over K is alge-
braically complemented. If w
is a basis for W, then it can be extended to a

basis w
for V . Letting Y = spany
, we get V = Y W.
The key issue in the above denition is that if J is a closed subspace in the
LCS 1, then we are asking that the complement of J also be closed. This is not
always possible. For example, c
0
is a closed subspace of (
, | |
). Nevertheless,
it does not possess a topological complement. The proof is omitted.
When J is nite-dimensional, the situation is somewhat better.
6.26. Proposition. Let J be a nite-dimensional subspace of a LCS (1, T ).
Then J is topologically complemented in 1.
Proof. First observe that J is closed in 1 by Corollary 4.21. Let w
1
, w
2
, ..., w
n
be a basis for J.
By Corollary 6.22, we can nd continuous linear functionals
1
,
2
, ...,
n
1
so that
j
(w
i
) =
ij
, where
ij
is the Kronecker delta function. Let =
n
j=1
ker
j
.
Since each
j
is continuous, ker
j
is closed for all j, and hence is also closed.
Suppose v 1. Let k
j
=
j
(v), 1 j n. Then w =

n
i=1
k
i
w
i
J. If
y := v w, then
j
(y) =
j
(v)
j
(w) = k
j
k
j
= 0, 1 j n. Hence y .
Finally, if z J, then z =

n
i=1
r
i
w
i
for some r
i
K, 1 i n. But then
z , so for each 1 j n, 0 =
j
(z) = r
j
. Hence z = 0 and 1 = J.
2
6.27. Corollary. Let (1, T ) be a LCS and J 1 be a closed subspace of 1.
If x 1, x , J, then there exists g 1
so that g[
W
= 0 but g(x) ,= 0.
Proof. By Proposition 5.17, 1/J is a LCS. Also, x , J implies that q(x) ,= 0 in
1/J, where q : 1 1/J is the canonical quotient map. We can therefore apply
Corollary 6.23 to produce a functional f (1/J)
so that f(q(x)) ,= 0. Since q is

continuous, so is g := f q, and so g 1
satises g(w) = 0 for all w J, while

g(x) ,= 0.
2
6.28. Theorem. Let (1, T ) be a LCS and J 1 be a linear manifold. Then
J = ker f : f 1
and J ker f.
Proof. Clearly f 1
implies that ker f is closed, so if J ker f, then J ker f.

Thus
J ker f : f 1
and J ker f.
Conversely, suppose that x 1, x , J. By Corollary 6.27, there exists g 1
so that g[
W
= 0 but g(x) ,= 0. This proves the reverse inclusion, and combining the
two inclusions yields the desired result.
2
6.29. Corollary. Let (1, T ) be a LCS and J 1 be a linear manifold. The
following are equivalent:
(a) J is dense in 1.
(b) f 1
and f[
W
= 0 implies that f = 0.
Let us now describe some quantitative versions of the above results, in the setting
of normed linear spaces.
6.30. Corollary. Let (X, | |) be a NLS and x X. Then
|x| = max[x
(x)[ : x
, |x
| 1.
Proof. For the rest of the proof, the vector x X is xed.
Let := sup[x
(x)[ : x
, |x
| 1. Then for any x
with |x
| 1,
[x
(x)[ |x
| |x| |x|, and so |x|.

Dene Y = Kx, so that Y is a one-dimensional normed, linear subspace of X.
Dene f Y
#
via f(kx) = k|x|. Then [f(kx)[ = [k[ |x| = |kx|, and so |f| = 1.
By the Hahn-Banach Theorem 03 (Theorem 6.21), there exists y
so that
y
[
Y
= f, and |y
| = |f| = 1.
Thus
[y
(x)[ = y
(x) = f(x) = |x|,

which proves that |x|, and hence that = |x|. It also shows that the
supremum is attained at y
.
2
Recall from Proposition 2.18 that if X is a normed linear space, then the canon-
ical embedding : X X
which sends x X to x X
, where x(x
) = x
(x) for
all x
is a contractive linear mapping.

As a simple consequence of Corollary 6.30, we obtain:
6.31. Corollary. The canonical embedding : X X
is an isometry.
6.32. Corollary. Let (X, | |) be a NLS and Y X be a closed subspace, with
z X but z , Y. Let d := d(z, Y) = |z + Y|. Then there exists x
so that
|x
| = 1, x
[
Y
= 0, and x
(z) = d.
Proof. Let q : X X/Y denote the canonical quotient map. Since X/Y is a
NLS and |q(z)| = d, Corollary 6.30 guarantees the existence of a linear functional
(X/Y)
so that |
| = 1 and
(q(z)) = |q(z)| = d.
Let x
q. Obviously x
(z) = d. Since |q| 1, |x
| |
| |q| 1. Also,
for y Y, x
(y) =
(q(y)) =
(0) = 0.
To see that |x
| 1, note that |
| = 1 and so we can nd a sequence

(q(x
n
))
n=1
in X/Y with |q(x
n
)| < 1 for all n 1 and lim
n
[
(q(x
n
))[ = 1.
Choose y
n
Y so that |x
n
+y
n
| < 1 for all n. Then
lim
n
[x
(x
n
+y
n
)[ = lim
n
[
(q(x
n
+y
n
)[
= lim
n
[
(q(x
n
))[
= 1.
Hence |x
| 1, whence |x
| = 1.
2
THE SEPARATION THEOREMS
6.33. Proposition. Let (1, T ) be a LCS over the eld K and let ,= G 1
be an open, convex subset of 1 with 0 , (. Then there exists a closed hyperplane
/ in 1 such that G /= .
Proof. Let us rst consider the case where K = R.
Fix x
0
G and let H = x
0
G. Then H |
V
0
is open and convex. Let p
H
denote
the Minkowski functional on H. By Proposition 5.10, H = x 1 : p
H
(x) < 1.
Observe that 0 , G implies that x
0
, H. Thus p
H
(x
0
) 1. Let J = Rx
0
, and
dene f : J R via f(kx
0
) = kp
H
(x
0
). Clearly f J
#
. Moreover,
if k 0, then f(kx
0
) = kp
H
(x
0
) = p
H
(kx
0
), while
if k < 0, then f(kx
0
) = kp
H
(x
0
) < 0 p
H
(kx
0
).
It follows from the Hahn-Banach Theorem 01 (Theorem 6.18) that there exists a
linear functional g : 1 R with g[
W
= f and g(x) p
H
(x) for all x 1. Suppose
that y H. Then Re g(y) = g(y) p
H
(y) < 1.
By Proposition 6.12, g is continuous on 1. Thus / := ker g is a closed hyper-
plane in 1. [Note: obviously g ,= 0 since f ,= 0.]
Suppose z G. Then x
0
z H, so g(x
0
) g(z) = g(x
0
z) p
H
(x
0
z) < 1.
On the other hand, g(x
0
) = f(x
0
) = p
H
(x
0
) 1, and so
g(z) > g(x
0
) 1 0, and z , /.
Thus G /= .
Next, suppose that K = C.
Then 1 is also an R-linear space, and so as above we can nd a continuous R-
linear functional g
R
: 1 R so that Gker g
R
= . Let g
C
be the complexication
of g
R
, g
C
(x) = g
R
(x) ig
R
(ix), x 1. By Lemma 6.6, g
C
is a C-linear functional
and g
R
= Re g
C
.
Now g
C
(x) = 0 if and only if g
R
(x) = g
R
(ix) = 0. Let / = ker g
C
. Then
/= ker g
R
[i ker g
R
] is a closed C-hyperplane in 1 and /G ker g
R
G = .
2
6.34. Denition. An ane hyperplane / in a TVS (1, T ) is a translate
of a hyperplane; that is, / is an ane hyperplane if there exists x / so that
/x is a hyperplane.
More generally, L 1 is an ane manifold (resp. ane subspace) of 1 if
there exists m L so that L m is a manifold (resp. subspace) of 1.
We remark that if there exists m L so that Lm is a manifold in 1, then for
all m L we must have L m is a manifold. The verication of this is left to the
reader.
6.35. Corollary. Let (1, T ) be a LCS and ,= G 1 be open and convex. If
L 1 is an ane subspace of 1 and L G = , then there exists a closed, ane
hyperplane 1 so that L and G = .
Proof. Choose m L and let L
0
= L m, so that L
0
is a closed subspace of 1.
Let G
0
= Gm. Since L G = , it follows that L
0
G
0
= . Let q : 1 1/L
0
denote the canonical quotient map.
Since G is open, so is G
0
. Since q is an open map (see paragraph 4.18), q(G
0
)
is open. Furthermore, G is convex and hence so are G
0
and q(G
0
). Again, since
L
0
G
0
= , 0 , q(G
0
). By Proposition 6.33, there exists a closed hyperplane ^
0
in 1/L
0
so that ^
0
q(G
0
) = . Let
0
= q
1
(^
0
). It is routine to check that
0
is a linear manifold in 1, and
0
is closed since q is continuous. Moreover,
dim1/
0
= dim(1/L
0
)/(
0
/L
0
)
= dim(1/L
0
)/^
0
= 1,
and so
0
is a closed hyperplane in 1 with L
0

0
. Translating back, let =
0
+ m. Then is a closed ane hyperplane of 1, L and if z G, then
q(z m) q(
0
) q(G
0
) = ^
0
q(G
0
) = , a contradiction.
2
6.36. Denition. Let (1, T ) be a TVS over the eld R. By an open half-
space (resp. closed half-space) we shall mean a subset o 1 for which there
exist a non-zero continuous linear functional f : 1 R and k R so that
o = x 1 : f(x) > k
(resp. o = x 1 : f(x) k.)
We say that two subsets A and B of 1 are separated if we can nd closed half-
spaces o
A
and o
B
so that A o
A
, B o
B
and o
A
o
B
is a closed ane hyperplane
of 1. We say that A and B are strictly separated if we can nd disjoint open
half-spaces o
A
and o
B
with A o
A
and B o
B
.
Note that if f 1
and k R, then o = x 1 : f(x) < k is also an open

half-space, since g = f 1
and o = x 1 : g(x) > k. A similar statement

holds for closed half-spaces.
6.37. Example.
(a) Consider R
2
equipped with the Euclidean norm. Let A = (x, y) R
2
:
x < 0 and y 1/x
2
, B = (x, y) R
2
: x > 0 and y 1/x
2
. Then
f : R
2
R, f(x, y) = x denes a continuous linear functional on R
2
. Let
o
A
= (x, y) R
2
: f(x) < 0 and o
B
= (x, y) R
2
: f(x) > 0. Then
o
A
, o
B
are disjoint open half-spaces with A o
A
and B o
B
. Hence A
and B are strictly separated.
(b) With R
2
, f A, o
A
and o
B
as above, set C = (0, y) : y R. Then
A o
A
= (x, y) R
2
: f(x) 0, C o
B
= (x, y) R
2
: f(x) 0
and o
A
, o
B
are closed half-spaces for which o
A
o
B
= (x, y) R
2
: x = 0
is a closed (ane) hyperplane.
Thus A and C are separated. One can check that A and C are not
strictly separated.
6.38. Theorem. The Hahn-Banach Theorem 04 - R
Let (1, T ) be a LCS over R and suppose that A and B are non-empty, disjoint,
open, convex subsets of 1. Then A and B are strictly separated.
Proof. Let G := A B = a b : a A, b B. We claim that ,= G is
open and convex. That ,= G is obvious as both A and B are non-empty. Since
G =
bB
A b, G is the union of open sets (each A b is a translate of the open
set A), and thus G is open.
Suppose that g
1
= a
1
b
1
and g
2
= a
2
b
2
lie in G. Let t [0, 1]. Then
tg
1
+ (1 t)g
2
= [ta
1
+ (1 t)a
2
] [tb
1
+ (1 t)b
2
].
Since A and B are convex, it follows that so is G. Observe that A B = also
implies that 0 , G.
It now follows from Proposition 6.33 that there exists a closed hyperplane /
in 1 such that / G = . Let f denote a continuous linear functional on 1 such
that ker f = /. (That a linear functional with this kernel exists was demonstrated
in paragraph 6.9, and that it is continuous follows from Corollary 6.13.)
Now G is convex and f is linear, whence f(G) is again convex. But Gker f = ,
so 0 , f(G) and hence either f(x) > 0 for all x G, or f(x) < 0 for all x G.
By replacing f by f if necessary, we may assume that the rst condition holds. If
a A, b B, then c = a b G, so f(c) = f(a) f(b) > 0, i.e. f(b) < f(a). We
deduce that there exists k R so that
supf(b) : b B k inff(a) : a A.
Now A is open and hence f(A) is open (check!). Similarly, f(B) is open. Hence
f(b) < k < f(a) for all a A, b B. It follows that o
A
= x 1 : f(x) > k
and o
B
= x 1 : f(x) < k are disjoint open half-spaces with A o
A
, B o
B
.
Hence A and B are strictly separated.
2
6.39. Remark. If A were open but not B, then G =
bB
Ab would still be
a union of open sets and hence would still be open. The conclusion would be there
there exist a continuous linear functional g : 1 R and a constant k R such that
A x 1 : g(x) > k and B x 1 : g(x) k.
6.40. Theorem. The Hahn-Banach Theorem 04 - C
Let (1, T ) be a LCS over C and suppose that A, B 1 are non-empty, disjoint,
open, convex subsets of 1. Then there exist a continuous C-linear functional f on
1 and k R so that
Ref(a) > k > Ref(b) for all a A, b B.
Proof. Thinking of (1, T ) as a vector space over R, we may apply Theorem 6.38
(i.e. the HB04-R) above to obtain a continuous R-linear functional f
R
: 1 R and
a constant k R so that
f
R
(a) > k > f
R
(b) for all a A, b B.
Let f
C
(x) = f
R
(x) if
R
(ix) be the complexication of f
R
. By Lemma 6.6, f
C
is
continuous and
Ref(a) > k > Ref(b) for all a A, b B.
Thus f = f
C
is the desired C-linear functional.
2
Let (1, T ) be a LCS and suppose that A, B 1 are non-empty, disjoint, closed,
convex subsets of 1. Suppose furthermore that B is compact. Then there are real
numbers , and a continous linear functional f 1
so that
Ref(a) > Ref(b)
for all a A, b B. In particular, A and B are strictly separated.
Proof. Observe that 1A is open and that b B implies that b 1A. It follows
from Corollary 5.15 that we can nd a balanced, convex, open nbhd N
b
of 0 so
that b + N
b
1A. The collection b + N
b
: b B is an open cover of B, and
bB
b +N
b
1A. Since B is compact, we can nd a nite subcover b
j
+N
b
j
n
j=1
of B. Let N
0
=
n
j=1
N
b
j
. Then N
0
|
V
0
, and N :=
1
3
N
0
|
V
0
is balanced, convex
and open.
Let A
0
= A+N = a+n : a A, n N and B
0
= B+N. Clearly A
0
,= ,= B
0
.
Then A
0
=
aA
a +N is open and similarly B
0
is open. If a
1
+n
1
, a
2
+n
2
A and
t [0, 1], then
t(a
1
+n
1
) + (1 t)(a
2
+n
2
) = (ta
1
+ (1 t)a
2
) + (tn
1
+ (1 t)n
2
) A+N,
since each of A and N is convex. Thus A
0
, and similarly B
0
, is convex.
Suppose z A
0
B
0
. Then there exists a A, b B and n
1
, n
2
N so that
a+n
1
= b+n
2
. Thus a = b+(n
2
n
1
). Now n
1
, n
2
N implies that 2n
2
, 2n
1
N
0
since N =
1
3
N
0
and N
0
is balanced. Since N
0
is convex,
1
2
(2n
2
)+
1
2
(2n
1
) = n
2
n
1

N
0
. Thus a b+N
0
, contradicting the fact that b+N
0
1A. Hence A
0
B
0
= .
By Theorem 6.40 (HB04-C), there exists f 1
and R so that
Re f(a) > > Re f(b)
for all a A
0
, b B
0
.
But B is compact, and Ref is continuous on B, so that = supRef(b) : b B
is attained at some point b
0
B. Thus
Re f(a) > > = Re f(b
0
) Re f(b)
for all a A, b B.
2
6.42. Corollary. Let (1, T ) be a LCS over R and ,= A 1. Then the closed,
convex hull of A, co(A), is the intersection of the closed half spaces that contain A.
Proof. Let = o : A o, o 1 is a closed half-space. Since each o is
closed and convex, B =
S
o is again closed and convex. Clearly A B, and so
the closed convex hull of A is also a subset of B.
If z , co(A), then z and co(A) are disjoint, non-empty, closed and convex
subsets of 1. Since z is compact, we can apply Theorem 6.41 (i.e. HB05) to
obtain a linear functional f 1
and , R such that

f(z) > f(y)
for all y co(A). Thus if o
0
= x 1 : f(x) , then o
0
is a closed half-space of
1, A co(A) o
0
, and z , o
0
. Thus
S
o co(A), proving that
co(A) =
S
o.
2
6.43. Corollary 6.30 admits an interesting interpretation. Given (X, | |) a NLS
and x
, we do not in general expect x
to achieve its norm. For example, if

X = c
0
, equipped with the supremum norm, and if
x
((x
n
)
n
) :=
n
x
n
2
n
,
then x
has norm one, but there is no x = (x

n
)
n
c
0
with |x| = 1 and [x
(x)[ =
|x
| = 1.
Nevertheless, Corollary 6.30 allows us to conclude that we may nd x
with |x
| = 1 for which [x
(x
)[ = 1. If : X
is the canonical embedding

of X
into its second dual, then |
| = 1 by Corllary 6.31 and

[
(x
)[ = [x
(x
)[ = 1 = |
|.
One can think of this as saying that the domain of x
is not large enough to

allow x
to attain its norm, but hat X
extends the domain of x
enough to allow
the extension

x
of x
to attain its norm. Of course, given an arbitrary z
,
it need not attains its norm at an element of X
and so unless X
is reexive the
game is once again afoot.
6.44. The proof of Proposition 6.33 also gives us an indication of how one may
try to interpret the Hahn-Banach Theorem 01, namely Theorem 6.18, geometrically.
Let 1 be a vector space over R and let p be a sublinear functional on 1. Suppose
that / is a linear manifold in 1 and f : / R is a linear functional with
f(m) p(m) for all m /.
Let H = x 1 : p(x) < 1. It is routine to verify that H is convex. Since
0 ,= f, there exists m
0
/ such that f(m
0
) > 1. This forces p(m
0
) f(m
0
) > 1,
and so m
0
, H. Let K = m
0
H = m
0
h : h H. Clearly K is also convex.
Since m
0
, H, 0 , K. In fact, we claim that K ker f = .
Suppose otherwise: let k K ker f. Then k / and k = m
0
h for some
h H, which forces h H /, since k, m
0
/. But
0 = f(k) = f(m
0
h) = f(m
0
) f(h) > 1 p(h) > 0,
a contradiction. Thus K ker f = . Theorem 6.18 then says that we can extend
f to a linear functional g : 1 R with g[
M
= f, such that
g(x) p(x) for all x 1.
A similar analysis to that above shows that K ker g = . In other words,
one can translate the unit ball H of 1 (as measured by the sublinear functional
p note, p doesnt even have to be a non-negative-valued function, and hence the
interpretation of H as a unit ball here is very, very loose) so that it doesnt
intersect the linear manifold ker f in such a way that the manifold may be extended
to a hyperplane (ker g) which also doesnt intersect the translation.
This is only intended as a heuristic. Proposition 6.33 shows how to correctly
use HB01, i.e. Theorem 6.18, to obtain an interesting geometric result in locally
convex spaces.
There is less in this than meets the eye.

Tallulah Bankhead
7. WEAK TOPOLOGIES AND DUAL SPACES 89
7. Weak topologies and dual spaces
Last week I stated that this woman was the ugliest woman I had ever
seen. I have since been visited by her sister and now wish to withdraw
that statement.
Mark Twain
7.1. In Remarks 5.27, we observed that if 1 is a vector space over K, if
(X
, | |
)
A
is a family of normed linear spaces, and if for each A we have a
linear map T
: 1 X
, then each
p
: 1 R
x |T
x|
is a seminorm. Furthermore, if T
A
is separating i.e. for each 0 ,= x 1,
there exists
0
A so that T
0
x ,= 0 then the family = p
A
is a separating
family of seminorms. Finally, we saw there that the LCS topology on 1 generated
by was nothing more (nor was it anything less) than the weak topology generated
by T
A
.
Let us now consider the following special instance of this phenomenon. Again,
we begin with a vector space 1 over K, and we assume that we are given a separating
family 1
#
. Of course, for each , we have
: 1 K,
and (K, [ [) is a normed linear space. Since was assumed to be separating for 1,
the family =
: of functions dened by
(x) = [(x)[, x 1,
is a separating family of seminorms which generates a LCS topology on 1. From
above, this topology coincides with the weak topology generated by , and we shall
denote it by (1, ).
Thus a base for the (1, ) topology on 1 is given by
B = N(x, F, ) : x 1, > 0, F nite,
where for each x, F and > 0 as above,
N(x, F, ) = y 1 :
(x y) = [(x) (y)[ < , F.

In particular, a net (x
in (1, (1, )) converges to x 1 if and only if

lim
(x
x) = lim
[(x
) (x)[ = 0,
or equivalently,
lim
(x
) = (x)
for all .
7.2. Denition. Let 1 be a vector space over K, and suppose that L 1
#
is
both a linear manifold and a separating family of linear functionals. We say that
(1, L) is a dual pair.
7.3. Example. Suppose that (1, T ) is a LCS and that L = 1
. By Corol-
lary 6.23, L separates points of 1 and hence (1, 1
) is a dual pair. The (1, 1
)
topology is suciently important to merit its own name, and we refer to it as the
weak topology on 1. If (x
is a net in 1 which converges to some x in the weak

topology, we say that (x
converges weakly to x.
Suppose that (x
is a net in 1 which converges to x 1 in the initial topology

T . For any f 1
, the fact that f is continuous implies that

lim
f(x
) = f(x).
Thus (x
converges to x weakly. It follows that the weak topology on 1 induced

by 1
is weaker than the initial topology: in other words, (1, 1
) T .
Let (V, L) be a dual pair. By Paragraph 7.1, each L is continuous on V
relative to the (V, L) topology. Our present goal is to show that these are the only
(V, L)-continuous linear functionals on V .
7.4. Theorem. Let (V, L) be a dual pair. Then
L = (V, (V, L))
.
Proof. That L is contained in (V, (V, L))
was shown in Paragraph 7.1.

Suppose now that V
#
is (V, L)-continuous. Then the map p
: V R
satisfying p
(x) = [(x)[ denes a (V, L)-continuous seminorm on V . By Proposi-

tion 5.29, there exist
1
,
2
, ...,
n
L and 0 < R so that
p
(x) = [(x)[ max([

1
(x)[, [
2
(x)[, ..., [
n
(x)[) for all x V.
It follows that ker
n
j=1
ker
j
. By Proposition 6.7, span
j
n
j=1
L.
2
7.5. Remark.
We rst remark that if 1
#
is a separating family of linear functionals
but is not a linear manifold, then after setting L = span, one can verify
that the (1, L)-topology on 1 agrees with the (1, )-topology, and hence
that
L = (1, (V, ))
.
It follows from Theorem 7.4 that the only weakly continuous linear func-
tionals on a locally convex space (1, T ) are the elements of (1, T )
.
7.6. Denition. Suppose that (1, T ) is a LCS. Then 1
1
#
is a vector space
over K. For each x 1, dene
x : 1
K
x() := (x).
Then

1 := x : x 1 is a linear manifold in (1
)
#
. If 0 ,= 1
, then obviously
there exists x 1 such that (x) ,= 0. In other words,

1 is a separating family of
linear functionals on 1
. Hence (1
1) is a dual pair.
By convention, the weak topology on 1
induced by the family

1 is usually de-
noted by (1
, 1) (as opposed to (1
1)), and is referred to as the weak
-topology
on 1
.
7.7. Remark. It follows that a base for the weak
-topology on 1
is given by
B = N(, F, ) : 1
, > 0, F 1 nite,
where
N(, F, ) = 1
: [ x( )[ = [(x) (x)[ < , x F.

Moreover, a net (
in 1
converges in the weak
-topology to a functional 1
if and only if lim
(x) = (x) for all x 1. In other words, convergence in the

weak
-topology on 1
is convergence at every point of 1.

By Theorem 7.4, a functional is weak
-continuous on 1
if and only if = x
for some x 1.
V
) and (J, T
W
) be locally convex spaces, and sup-
pose that T : (1, T
V
) (J, T
W
) is a continuous linear operator. Then T is contin-
uous as a linear map between 1 and J when they are equipped with their respective
weak topologies.
Proof. Suppose that (x
is a net in 1 which converges weakly to x 1. We must

show that the net (Tx
converges weakly to Tx in J. Now, if J
, then T
is continuous with respect to the T
V
topology on 1, and hence T 1
. But the
weakly continuous linear maps on 1 coincide with 1
, and therefore T is weakly

continuous on 1, i.e. lim
T(x
) = T(x), as was to be shown.

2
When 1 is a LCS and C 1 is convex, we get a particularly nice result con-
cerning the weak topology.
7.9. Theorem. Let C be a convex set in a LCS (1, T ). Then the closure of C
in (1, T ) coincides with its weak closure in (1, (1, 1
)).
Proof. First observe that we can always view (1, T ) and (1, (1, 1
)) as locally
convex spaces over R. Since C is assumed to be convex already, Corollary 6.42
implies that the closure of C in (1, T ) (resp. in (1, (1, 1
))) is the intersection of

the T -closed (resp. (1, 1
)-closed) half spaces which contain C.

But a closed half space in a LCS corresponds to (a constant and) a continuous
linear functional on that space. Since (1, T ) and (1, (1, 1
)) share the same dual

space, namely 1
, it follows that they also share the same closed half-spaces, and
hence the closure of C in these two topologies must coincide.
2
7.10. Let (X, | |) be a Banach space. Recall from Proposition 2.18 that the
canonical embedding
: X X
x x,
where x(x
) := x
(x) for all x
, is a contractive map. In Corollary 6.31 we saw

that as a consequence of the Hahn-Banach Theorem is in fact an isometry.
By Theorem 7.4 and Remark 7.5, (X) corresponds exactly to the weak
-
continuous linear functionals on X
.
7.11. Proposition. Let X be a nite-dimensional Banach space. Then the
norm, weak and weak
-topologies on X all coincide.

Proof. First we must decide what we mean by the weak
-topology on X. Observe
that if dim X = n < , then dim X
= n as well, and thus dim X
= n = dim X.
Since : X X
is a linear isometry, it must be a bijection in this case and therefore

we can identify X with X
= (X
. In this sense X (X) comes equipped with a

weak
-topology induced by X
, namely the ((X), X
)-topology. But since we are

identifying X with (X) = X
, this is really just the (X, X
)-topology, namely the

weak topology on X.
Since the weak and the norm topologies on X are both TVS topologies, and
since nite-dimensional vector spaces admit a unique TVS topology, we see that all
three topologies cited above must coincide.
2
We now wish to examine some of the properties of the weak and weak
-topologies
in the context of normed linear spaces. We shall rst require a result from Real
Analysis, which we shall then adapt to the setting of normed linear spaces.
7.12. Theorem. The Uniform Boundedness Principle
Let (X, d) be a complete metric space and let H ((X, K) be a non-empty
family of continuous functions on X such that for each x X,
M
x
:= sup
hH
[h(x)[ < .
Then there exists an open set G X and a constant M > 0 so that
[h(x)[ M for all h H, x G.
Proof. For each m 1, let E
m,h
= x X : [h(x)[ m, and let E
m
=
hH
E
m,h
.
Since each E
m,h
is closed (as h H implies that h is continuous), so is E
m
. Also,
for any x X, there exists m > M
x
, and so x E
m
. Thus
X =
m=1
E
m
.
Since X is complete, the Baire Category Theorem implies the existence of k 1
so that the interior int (E
k
) of E
k
is non-empty. This clearly leads to the desired
conclusion.
2
7.13. Corollary. The Uniform Boundedness Principle - Banach space
version
Let (X, | |
X
) and (, | |
Y
) be Banach spaces and let A B(X, Y) denote a
family of continuous linear operators from X to Y. Suppose that for each x X, we
have
M
x
:= sup|Tx| : T A < .
Then
sup|T| : T A < .
Proof. For each T A, let p
T
: X R be the continuous seminorm given by
p
T
(x) = |Tx|. Since X is complete, the metric space version of the Uniform Bound-
edness Principle (Theorem 7.12) implies that there exists an open set ,= G X
and a constant M > 0 so that
|Tx| M for all T A, x G.
Now ,= G open implies that there exists z G and > 0 so that V
(z) = x
X : |x z| < G. Consider y V
(0) and T A.
Then
|Ty| |T(y +z)| +| Tz|
M +|Tz|
2M,
as z and y +z G. It follows that if x X, |x| 1, then
2
|Tx| = |T(
2
x)| M +|Tz| 2M,
and hence that
|Tx|
2
(2M), T A.
That is,
sup|T| : T A
4M
.
2
7.14. Corollary. Let X be a Banach space and o X. Then o is bounded if
and only if for all x
,
sup[x
(s)[ : s o < .
Proof. Suppose that o is bounded by M > 0. If x
, then [x
(s)[ |x
| |s|
M|x
| < .
Conversely, if sup[x
(s)[ : s o < , then sup[ s(x
)[ : s o < for all

x
. By the Uniform Boundedness Principle,

sup| s| : s o = sup|s| : s o < .
2
7.15. Corollary. Let X be a Banach space and S X
. Then S is bounded if
and only if for all x X,
sup[s
(x)[ : s
S < .
Proof. This is an immediate consequence of the Uniform Boundedness Principle,
Theorem 7.13.
2
7.16. Theorem. The Banach-Steinhaus Theorem
Let X and Y be Banach spaces and suppose that T
n
n=1
B(X, Y) is a sequence
which satises the property that for each x X, there exists y
x
Y so that
lim
n
T
n
x = y
x
.
Then the map T : X Y dened by Tx = y
x
is a bounded linear map, and
sup
n
|T
n
| < .
Proof. For each x X, we have that T
n
x
n=1
converges to some y
x
, and therefore
it is bounded. That is, sup
n1
|T
n
x| < for each x X. By the Uniform
Boundedness Principle, M := sup
n1
|T
n
| < .
Let Tx := lim
n
T
n
x, for each x X. Linearity of T is readily checked. Also,
|Tx| sup
n1
|T
n
x|
_
sup
n1
|T
n
|
_
|x| M|x|, x X.
Hence |T| M < , and T is bounded.
2
In general, we do not expect weak topologies to be determined by sequential
convergence. If we do have weak convergence of a sequence, therefore, something
strong is implied.
7.17. Proposition. Let X be a Banach space.
(a) If (x
n
)
n=1
is a sequence which converges weakly to x X, then
(i) sup|x
n
| < ; and
(ii) |x| liminf |x
n
|.
(b) If (y
n
)
n=1
is a sequence which converges in the weak
-topology to y
,
then
(iii) sup|y
n
| < ; and
(iv) |y
| liminf |y
n
|.
Proof.
(a) (i) For each x
, lim
n
x
(x
n
) = x
(x). It follows that the sequence

(x
(x
n
))
n=1
is bounded for each x
. By Corollary 7.14, x
n
n=1
is bounded.
(ii) If we choose x
with |x
| = 1 so that [x
(x)[ = |x|, then

|x| = [x
(x)[ = [ lim
n
x
(x
n
)[ liminf |x
| |x
n
| = liminf |x
n
|.
(b) (iii) For each x X, lim
n
y
n
(x) = lim
n
x(y
n
) = x(y
) = y
(x), so
that (y
n
(x))
n=1
is bounded. By the Uniform Boundedness Principle,
sup
n1
|y
n
| < .
(iv) Fix > 0. Choose y X such that |y| = 1, and [y
(y)[ |y
| .
Then
|y
| [y
(y)[
= [ y(y
)[
= [ lim
n
y(y
n
)[
liminf | y| |y
n
|
= liminf |y
n
|
for all > 0. Hence
|y
| liminf |y
n
|.
2
It is worth pointing out that by Theorem 7.9, if (x
n
)
n=1
converges weakly to x,
then x co
(x
n
n=1
), since the latter is a convex set in X, closed in the norm
(and hence in the weak) topology.
Before considering our next example, let us rst recall a result from Measure
Theory, alternately referred to as the Riesz Representation Theorem or the Riesz-
Markov Theorem.
7.18. Theorem. Let X be a locally compact, Hausdor topological space, and
denote by /(X) the space of K-valued, nite, regular, Borel measures on X, equipped
with the total variation norm: || = [[(X).
If /(X), then
: (
0
(X, K) K given by
(f) =
_
X
fd is an ele-
ment of (
0
(X, K)
, and the map : /(X) (

0
(X, K)
is an isometric linear
isomorphism.
For example, if X = N with counting measure, then (
0
(X, K) = c
0
(N, K) and
/(X) =
1
(N, K).
When X = [0, 1], we can in turn identify /([0, 1]) = (([0, 1], K)
with the space

BV [0, 1] of left-continuous functions of bounded variation on [0, 1].
7.19. Proposition. Let X be a compact, Hausdor space. Then a sequence
f
n
n=1
in ((X) converges weakly to f ((X) if and only if
(i) sup
n
|f
n
| < ; and
(ii) For each x X, (f
n
(x))
n=1
converges to f(x).
Proof. Suppose rst that f
n
n=1
converges weakly to f. By Proposition 7.17,
sup
n
|f
n
| < . Let
x
: ((X) K be the evaluation functional,
x
(f) = f(x) for
all x X. Then
x
is linear and for f ((X), [
x
(f)[ = [f(x)[ |f|, so that
|
x
| 1 and
x
((X)
. (In fact,
x
corresponds to the point mass measure at x.)
Thus lim
n
x
(f
n
) =
x
(f), i.e. lim
n
f
n
(x) = f(x).
Conversely, suppose that (i) and (ii) hold. If ((X)
, then by the Riesz

Representation Theorem above, there exists /(X) with || = || so that
(f) =
_
X
fd,
for all f ((X). By the Lebesgue Dominated Convergence Theorem,
(f) =
_
X
fd = lim
n
_
X
f
n
d = lim
n
(f
n
).
In other words, (f
n
)
n
converges weakly to f.
2
7.20. Theorem. Tychonos Theorem
Suppose that (X
, T
) is a non-empty collection of compact, topological spaces.

Then X =

is compact in the product topology.

Proof. Recall from Real Analysis that it suces to prove that if T is a collection
of closed subsets of X with the Finite Intersection Property (FIP), then F : F
T ,= . To that end, let T be a collection of closed subsets of X with the FIP.
Let J = T(X) : T and has the FIP, partially ordered by inclu-
sion, so that
1

2
if
1

2
. Since T J, J ,= . Suppose that ( =
is a chain if J. Clearly T / :=
, and if H
1
, H
2
, ..., H
m
/, then the fact
that ( is totally ordered implies that there exists
0
so that H
1
, H
2
, ..., H
m

0
.
Since
0
has the FIP,
m
i=1
H
i
,= . Thus / has the FIP, and so / J is an upper
bound for (. By Zorns Lemma, J admits a maximal element, say /.
We make two observations: rst, if we set /
0
=
r
k=1
M
k
: M
k
/, 1 k
r, r 1, then the elements of /
0
are nite intersections of elements of /. It
follows that /
0
has the FIP. Moreover, T /
0
. Since / /
0
, the maximality
of / implies that / = /
0
. In other words, nite intersections of elements of /
lie in /.
Second, if R X and R M ,= for all M /, then T / / R
and / R has the FIP. Again, the maximality of / implies that R /.
Our goal now is to prove that M : M / ,= . Since F : F T
M : M /, this will suce to prove the Theorem.
For each , let : X X
denote the canonical projection map. Then ,=

/
:=
(M) : M / is a family of subsets of X
with the FIP. Since X
is
compact,
(M)
X
: M / , = . Choose x
(M)
X
: M /, and
let x = (x
. We want to show that x M : M /.

To do this, we must show that G |
X
x
implies G M ,= for all M /.
Clearly it suces to do this when G is a basic nbhd of x, say
G =
n
j=1
j
(U
j
),
where U
j
X
j
is open, 1 j n. Now for any
0
and x
0
U
0
T
0
,
x
0
(M)
X
0
for all M / implies that U
0
(M) ,= for all M /.
But then
1
0
(U
0
) M ,= whenever x
0
U
0
X
0
is open. By maximality
of / and the second observation above,
1
0
(U
0
) /. Since / is closed under
nite intersections by the rst observation,
G =
n
j=1
j
(U
j
) / whenever G is a basic nbhd of x.
Thus x M : M /, and we are done.
2
7.21. Theorem. The Banach-Alaoglu Theorem
Let X be a Banach space. Then the closed unit ball X
1
:= x
: |x
| 1
of X
is weak
-compact.
Proof. For each x X, x
1
, we have
[ x(x
)[ = [x
(x)[ |x
| |x| |x|.
Thus x(X
1
) D
x
:= z K : [z[ |x|. Now each such D
x
is compact, and so by
Tychonos Theorem above,
D :=
xX
D
x
is also compact in the product topology. To complete the proof, we shall show that
X
1
is homeomorphic to a closed, and therefore compact, subset of D.
Dene
: X
1
D
f ( x(f))
xX
= (f(x))
xX
.
Clear is injective. Now a net (f
converges weak
to f if and only if
lim
(x) = lim
x(f
) = x(f) = f(x) for all x X,

that is, if and only if lim
(f
) = (f).
Thus X
1
is homeomorphic to (X
1
). There remains to show that (X
1
) is closed
in D.
Suppose that (f
is a net in X
1
, and that ((f
))
converges to d = (d
x
)
xX
D. Then
lim
(x) = d
x
for all x X.
Dene f(x) := d
x
, x X. Then f is linear since each f
is, and
[f(x)[ = [d
x
[ |x| for all x X,
so that f X
1
. Clearly (f) = lim
((f
)), so that ran is closed, and we are

done.
2
7.22. Corollary. Every Banach space X is isometrically isomorphic to a sub-
space of (((L, K), | |
) for some compact, Hausdor space L.

Proof. Let L := X
1
. Then L is weak
-compact, by the Banach-Alaoglu Theorem,

and is Hausdor since X separates the points of X
1
. Dene
: X ((L, K)
x x[
L
.
Then is easily seen to be linear, and | x[
L
| | x| |x|.
By the Hahn-Banach Theorem [Corollary 6.30], there exists x
1
such that
[x
(x)[ = |x|, and so | x[

L
| [ x(x
)[ = [x
(x)[ = |x|; that is, is an isometry.

2
7.23. Corollary. Let X be a Banach space and suppose that / X
is weak
-
closed and bounded. Then / is weak
-compact.
7.24. Theorem. Goldstines Theorem
Let X be a Banach space and : X X
denote the canonical embedding. Then

(X
1
) is weak
-dense in X
1
. Thus (X) is weak
-dense in X
.
Proof. Clearly (X
1
) =

X
1
is convex, since X
1
is. Observe that the closure of
(X
1
) in the weak
-topology, namely (X
1
)
w
, is weak
-closed and convex. Being

weak
-closed in the weak
-compact set X
1
, it is also weak
-compact. Suppose that

X
1
and , (X
1
)
w
. Then, by the Hahn-Banach Theorem 6.41 (HB05), we

can nd a weak
-continuous linear functional

x

X
(X
) X
so that
Re
() = b
> a := supRe
() : (X
1
)
w
= sup[
()[ : (X
1
)
w
.
(The last equality follows from the fact that X
1
and hence (X
1
) and (X
1
)
w
are
balanced.)
But
sup[
()[ : (X
1
)
w
= sup[(x
)[ : (X
1
)
w
sup[ x(x
)[ : x X
1
= sup[x
(x)[ : x X
1
= |x
|,
while
[Re
()[ [
()[ = [(x
)| || |x
| |x
|.
This contradicts our choice of

x
, and thus (X
1
)
w
= X
1
, as claimed.
Since X
=
n1
X
n
, and since each (X
n
) is weak
-dense in X
n
by a routine
modication of the above proof, (X) is weak
-dense in X
.
2
7.25. Example. Since c
0
(K)
=
1
(K) and
1
(K)
(K)
, the unit ball

(c
0
(K))
1
of c
0
(K) is weak
-dense in the closed unit ball of
(K), and thus c

0
(K) is
weak
-dense in
(K).
Of course, c
00
(K) is norm dense in c
0
(K), and so c
00
(K) is also weak
-dense in
(K).
Culture: Although we shall not have time to prove this, the non-commutative
analogue of the above statement is that the set of nite rank operators T(H) on an
innite-dimensional Hilbert space is weak
-dense in B(H).
Let us now establish a relation between compactness and reexivity of a Banach
space.
7.26. Proposition. Let X be a Banach space. The following are equivalent.
(a) X is reexive.
(b) X
1
is weakly compact.
Proof.
(a) implies (b): First suppose that X is reexive. Then

X = X
and

X
1
= X
1
is weak
-compact the Banach-Alaoglu Theorem 7.21. But then the weak
-
topology on

X
1
is just the weak topology on X
1
, so X
1
is weakly compact.
(b) implies (a): Next suppose that X
1
is weakly compact. Then

X
1
is weak
-
compact, and since the weak
-topology is Hausdor,

X
1
is weak
-closed.
But by Goldstines Theorem,

X
1
is weak
-dense in X
1
. Thus

X
1
= (X
1
) =
(X
1
)
w
= X
1
. This in turn implies that

X = X
, or in other words, that

X is reexive.
2
Although in general, weak topologies are not metrizable, sometimes their re-
strictions to bounded sets can be:
7.27. Theorem. Let X be a Banach space. Then X
1
is weak
-metrizable if and
only if X is separable.
Proof. First assume that X is separable, and let x
n
n=1
be a dense subset of X.
Dene a metric d on X
1
via
d(x
, y
) =
n=1
[x
(x
n
) y
(x
n
)[
2
n
|x
n
|
.
Then a net (x
in X
1
converges in the metric topology to x
1
if and only if
(x
(x
n
))
converges to x
(x
n
) for all n 1 (exercise). If x X and > 0, we can
choose n 1 so that |x
n
x| < /3. Choose
0
so that
0
implies that
[x
(x
n
) x
(x
n
)[ < /3. Then
0
implies that
[x
(x) x
(x)[ [x
(x) x
(x
n
)[ +[x
(x
n
) x
(x
n
)[ +[x
(x
n
) x
(x)[
|x
| |x x
n
| +/3 +|x
| |x
n
x|
< /3 +/3 +/3 = .
Thus x
(x
n
) converges to x
(x
n
) for all n 1 if and only if (x
converges in the
weak
-topology to x
. Hence the weak
-topology on X
1
is metrizable.
Next, assume that X
1
is weak
-metrizable. Then we can nd a countable se-

quence G
n=1
of weak
-open nbhds of 0 X
1
so that
n=1
G
n
= 0. There is no
harm in assuming that each G
n
is a basic weak
-open nbhd, so for each n 1 there

exists
n
> 0 and a nite set F
n
X so that
G
n
= x
1
: [x
(x)[ <
n
, x F
n
.
Let F =
n=1
F
n
. If x
1
, x
(F) = 0, then x
n
for all n 1, and therefore
x
= 0. That is, if Y = span
F, then Y is separable and x
1
, x
[
Y
= 0 implies
that x
= 0. By the Hahn-Banach Theorem [Corollary 6.29], Y = X.

2
In a similar vein, we have
7.28. Theorem. Let X be a Banach space. Then X
1
is weakly metrizable if
and only if X
is separable.
Proof. Assignment.
2
7.29. Denition. Let X be a Banach space and M X, N X
. Then the
annihilator of M is the set
M
= x
: x
(m) = 0 for all m M,

while the pre-annihilator of N is the set
N = x X : n
(x) = 0 for all n
N.
Observe that M
and

N are linear manifolds in their respective spaces. More-
over, both are norm-closed and hence Banach spaces in their own right.
7.30. Theorem. Let X be a Banach space, and let M X be a closed subspace.
Let q : X X/M denote the canonical quotient map. Then
: (X/M)
q
is an isometric isomorphism of Banach spaces.
Proof. Clearly is linear. Let us show that is injective.
If (
1
) =
1
q =
2
q = (
2
), then
1
(q(x)) =
2
(q(x)) for all x X,
and so
1
=
2
.
Next we show that is surjective.
Let z
and dene
z
: X/M K via
z
(q(x)) = z
(x). Since
M ker z
, the map is well-dened. Furthermore, if x X and |q(x)| < 1,

then there exists m M so that |x +m| < 1, and
[
z
(q(x))[ = [z
(x)[ = [z
(x +m)[ |z
|,
so that |
z
| |z
| < . Hence
z
(X/M)
. Clearly (
z
) = z
.
Thus is bijective, and |()| = | q| || |q| ||, so that || 1.
Conversely, let > 0 and choose q(x) X/M with |q(x)| < 1 so that [(q(x))[
|| . Choose m M so that |x +m| < 1. Then
| q| [ q(x +m)[ = [(q(x))[ || ,
so that |()| = | q| ||, implying that is in fact isometric.
2
7.31. Theorem. Let X be a normed linear space and M X be a closed linear
subspace. Then the map
: X
/M
+M
[
M
is an isometric isomorphism.
Proof. Note that M
closed implies that X
/M
is a Banach space. We check

that is well-dened.
If x
+M
= y
+M
, then x
, so that (x
)[
M
= 0. That is,
(x
+ M
) = (y
+ M
). Working our way backwards through this argument

proves that is injective. That is linear is easily veried.
Next suppose that m
. By the Hahn-Banach Theorem, we can nd

x
, |x
| = |m
| so that x
[
M
= m
. Then (x
+M
) = x
[
M
= m
, so that
is onto. Thus is bijective.
Suppose that |x
+M
| < 1. Then there exists n
so that |x
+n
| < 1.
Thus
|(x
+M
)| = |x
[
M
| = |(x
+n
)[
M
| |x
+n
| < 1.
It follows that || 1.
From above, given m
, there exists x
with |x
| = |m
| so that
1
(m
) = x
+M
. Now
|
1
(m
)| = |x
+M
| |x
| = |m
|,
so that
1
is also contractive. But then is isometric, and we are done.
2
7.32. It is a worthwhile exercise to think about the relationship between the an-
nihilator M
of a subspace Mof a Hilbert space H and the orthogonal complement

of M in H, for which we used the same notation.
In particular, one should interpret what Theorem 7.31 says in the Hilbert space
setting, where M
refers to the orthogonal complement of M.
If you had a face like mine, youd punch me right on the nose, and
Im just the fella to do it.
Stan Laurel
The following result characterizes weak convergence of sequences in
p
. We
leave its proof as an exercise for the reader.
7.33. Proposition. Suppose that 1 < p < . A sequence (x
n
)
n=1
in
p
(N)
(i.e. each x
n
= (x
n1
, x
n2
, x
n3
, ...)
p
(N) converges weakly to z = (z
1
, z
2
, z
3
, ...)
p
(N)) if and only if
(i) sup
n1
|x
n
| < , and
(ii) lim
n
x
nk
= z
k
for all k N.
8. Local compactness and extremal points
Somewhere on this globe, every ten seconds, there is a woman giving
birth to a child. She must be found and stopped.
Sam Levenson
8.1. The main result of this section is the Krein-Milman Theorem, which asserts
that a non-empty, compact, convex subset of a LCS has extreme points; so many,
in fact, that we can generate the compact, convex set as the closed, convex hull of
these extreme points.
Extreme points of convex sets appear in many dierent contexts in Functional
Analysis. For example, it is an interesting exercise (so interesting that it may appear
as an Assignment question) to calculate the extreme points of the closed unit ball
B(C
n
)
1
of the locally convex space B(C
n
), where C
n
is endowed with the Euclidean
norm | |
2
and B(C
n
) is given the operator norm.
Recall that a linear map T B(C
n
) is said to be positive and we write T 0
if Tx, x) 0 for all x C
n
. An equivalent formulation of this property says that
T is positive if there exists an orthonormal basis for C
n
with respect to which the
matrix [T] of T is diagonal, and all eigenvalues of T are non-negative real numbers.
A linear functional B(C
n
)
is said to be positive if (T) 0 whenever T 0.

For example, if e
1
, e
2
, ..., e
n
is an orthonormal basis for C
n
, then the so-called
normalized trace functional
(T) :=
1
n
n
k=1
Te
k
, e
k
)
for T B(C
n
) can be shown to be a positive linear functional of norm one.
The state space o(B(C
n
)) of B(C
n
), consisting of all positive, norm-one linear
functionals on B(C
n
) called states forms a non-empty, compact, convex subset
of B(C
n
)
. The extreme points of the state space are called pure states. For
example, if x C
n
, then the map
x
: B(C
n
) C
T Tx, x)
denes a pure state. States on B(C
n
) (and more generally states on so-called
C
-algebras) are of extreme importance in determining the representation theory

of these algebras. This, however, is beyond the scope of the present manuscript.
8.2. Denition. A topological space (X, T ) is said to be locally compact if
each point in X has a nbhd base consisting of compact sets.
Suppose that X is locally compact and Hausdor, and that x
0
X. Then for
all U |
x
0
, there exists K |
x
0
so that K is compact and K U. Choose G T
so that x
0
G K. Then G K = K U, and so G is compact. That is, if X is
8. LOCAL COMPACTNESS AND EXTREMAL POINTS 105
Hausdor and locally compact, then for any U |
x
0
, there exists G T so that G
is compact and x
0
G G U.
8.3. Example. Let n 1 be an integer and consider (K
n
, | |
2
). Then for each
x K
n
, the collection
B
(x) := y K
n
: |y x|
2
: > 0
is a nbhd base at x consisting of compact sets, so (K
n
, | |
2
) is locally compact.
Our next result says that this is essentially the only example amongst locally
convex spaces.
8.4. Theorem. A LCS (1, T ) is locally compact if and only if 1 is nite-
dimensional.
Proof. If dim1 < , then 1 is homeomorphic to (K
n
, | |
2
) by Proposition 4.20.
Since (K
n
, | |
2
) is locally compact from above, so is 1.
Conversely, suppose that 1 is locally compact. Choose N T |
0
so that N
is compact. Now
1
2
N T |
0
and N
xN
x +
1
2
N. Since the latter is an open
cover of N, we can nd x
1
, x
2
, ..., x
r
N so that
N
r
i=1
x
i
+
1
2
N = x
1
, x
2
, ..., x
r
+
1
2
N.
Let /= spanx
1
, x
2
, ..., x
r
. Then N /+
1
2
N. Moreover,
1
2
N
1
2
N
1
2
/+
1
4
N,
so that N /+ (
1
2
/+
1
4
N) = /+
1
4
N.
Repeating this argument ad nauseum shows that
N /+
1
2
k
N for all k 1.
We claim that N /.
If we can prove this, then the fact that / is nite-dimensional implies that /
is closed, and so N / = /. But then 1 =
r=1
rN /, proving that 1 is
also nite-dimensional, as required.
Let w
1
, w
2
, ..., w
s
be a basis for / (here s r). Since dim / < , /
is topologically complemented in 1, so we can nd a closed subspace of 1 with
/ = 1. It is routine to verify that w
1
+, w
2
+, ..., w
s
+ is a basis for
1/ /.
As such, we can choose
1
,
2
, ...,
s
(1/)
so that
j
(w
k
+) =
j k
, 1 j, k s.
Dene
j
:=
j
q, where q : 1 1/ is the canonical quotient map. Then
j
1
,
1 j s,
j
(w
k
) =
j k
for all 1 j, k s, and
s
j=1
ker
j
= .
Since N is compact, given any 1
, (N) is again compact. Thus there exists
> 0 so that
[(n)[
for all n N.
Since N /+
1
2
k
N for all k 1, given z N, we can nd m
k
/ and u
k

1
2
k
N
so that z = m
k
+ u
k
. Also, since 1 = / , we can nd a unique m / and
u so that z = m+y.
For each 1 j s and k 1,
j
(z) =
j
(m
k
) +
j
(u
k
).
But [
j
(u
k
)[
1
2
k
j
for all k 1 by linearity of
j
, and so
j
(z) = lim
k

j
(m
k
). Of
course, y =
s
j=1
ker
j
implies that
j
(z) =
j
(m), 1 j s and thus
j
(m) =
j
(z) = lim
k
j
(m
k
), 1 j s.
Now m =

s
j=1
j
(m)y
j
and m
k
=

s
j=1
j
(m
k
)y
j
for each k 1. Hence
m = lim
k
m
k
.
Also, u
k

1
2
k
N for all k 1 implies that
lim
k
u
k
= 0.
Since z = m
k
+u
k
for all k 1,
z = lim
k
m
k
= m /.
This completes the proof.
2
An interesting and useful consequence of this result is the following.
8.5. Corollary. Let (X, | |) be a NLS. Then the closed unit ball X
1
of X is
compact if and only if X is nite-dimensional.
Proof. First suppose that X
1
is compact. If U |
X
0
is any nbhd of 0, then there
exists > 0 so that |x| < 2 implies that x U. But then X
U, and X
= X
1
is compact, being a homeomorphic image of X
1
. By denition, X is locally compact,
hence nite-dimensional, by Theorem 8.4.
Conversely, suppose that X is nite-dimensional. By Theorem 8.4, X is locally
compact. By hypothesis, there exists a compact nbhd K of 0. As above, there exists
> 0 so that X
K. Since X
is a closed subset of a compact set, it is compact.

Since X
1
= (
1
)X
is a homeomorphic image of a compact set, it too is compact.

2
8.6. Denition. Let V be a vector space and C V be a convex set. A point
e C is called an extreme point of C if whenever there exist x, y C and t (0, 1)
for which
e = tx + (1 t)y,
it follows that x = y = e. We denote by Ext(C) the (possibly empty) set of all
extreme points of C.
8.7. Example.
(a) Let V = C. Let D = w C : [w[ < 1 denote the open disk. It is easy
to see that D is convex. However D has no extreme points. If w D, then
[w[ < 1, so there exists > 0 so that (1 + )[w[ < 1. Let x = (1 + )w,
y = (1 )w. Then x, y D and w =
1
2
x +
1
2
y.
(b) With V = C again, let D = w C : [w[ 1. Then every z T :=
z C : [z[ = 1 is an extreme point of D. The proof of this is left as an
exercise.
(c) Let V = R
2
, and let p
1
, p
2
, p
3
be three non-collinear points in V . The
triangle T whose vertices are p
1
, p
2
, p
3
has exactly p
1
, p
2
, p
3
as its set of
extreme points.
The following generalizes the concept of an extreme point.
8.8. Denition. Let V be a vector space and let ,= C V be convex. A
non-empty convex set F C is called a face of C if whenever x, y C and t (0, 1)
satisfy tx + (1 t)y F, then x, y F.
We emphasize the fact that F is convex is part of the denition of a face.
8.9. Remarks. Let V be a vector space and C V be convex.
(a) If e is an extreme point of C, then F = e is a face of C. Conversely, if
F = z is a face of C, then z Ext(C).
(b) Let F be a face of C, and let D be a face of F. Then D is a face of C.
Indeed, let x, y C and t (0, 1), and suppose that tx +(1 t)y D.
Then D F implies that tx + (1 t)y F. Since F is a face of C, we
must have x, y F. But then D is a face of F, and so it follows from
tx + (1 t)y D that x, y D.
(c) From (b), it follows that if e is an extreme point of a face F of C, then e
is an extreme point of C.
8.10. Example.
(a) Let V = R
2
and let p
1
, p
2
, p
3
be three non-collinear points in V . Denote
by T the triangle whose vertices are p
1
, p
2
, p
3
. Then T is a face of itself.
Also, each line segment p
i
p
j
is a face of T. Finally, each extreme point p
j
is a face of T.
(b) Let V = R
3
and C be a cube in V , for e.g.,
C = (x, y, z) R
3
: 0 x, y, z 1.
Then C has itself as a face. Also, the 6 (square) sides of the cube are faces.
The 12 edges of the cube are also faces, as are the 8 corners. The corners
are extreme points of the cube.
The denition of a face currently requires us to consider convex combinations of
two elements of C. In fact, we may consider arbitrary nite convex combinations of
elements of C.
8.11. Lemma. Let (1, T ) be a LCS, ,= C 1 be convex and ,= F C
be a face of C. Suppose that x
j
n
j=1
C and that x =

n
j=1
t
j
x
j
is a convex
combination of the x
j
s. If x F and t
j
(0, 1) for all 1 j n, then x
j
F for
all 1 j n.
Proof. We argue by induction on n. If n = 1, there is nothing to prove, and the
case n = 2 is nothing more than the denition of a face. Let k 3, and suppose
that the result is true for n < k.
Suppose that x =

k
j=1
t
j
x
j
F, where t
j
(0, 1) for all 1 j k and
k
j=1
t
j
= 1. Then
x = (1 t
k
)
_
_
k1
j=1
t
j
1 t
k
x
j
_
_
+t
k
x
k
.
Since C is convex, y :=

k1
j=1
t
j
1t
k
x
j
C. But then x = (1 t
k
)y +t
k
x
k
F, and
F is a face, so that y and x
k
must lie in F. Since y F, our induction hypothesis
next implies that x
j
F for all 1 j k 1, which completes the proof.
2
8.12. Lemma. Let (1, T ) be a LCS and ,= K 1 be a compact, convex set.
Let 1
, and set
r = supRe (w) : w K.
Then F = x K : Re (x) = r is a non-empty, compact face of K.
Proof. Since Re : K R is continuous and K is compact, r = maxRe (w) :
w K, and so F is non-empty. Moreover, F = (Re )
1
(r) and r R is
closed, so F is closed in K, and hence F is compact.
Next, observe that if x, y F K and t (0, 1), then tx + (1 t)y K as K
is convex. But Re (tx + (1 t)y) = tRe (x) + (1 t)Re (y) = tr + (1 t)r = r,
so that tx + (1 t)y F and F is convex.
Suppose that x, y K, t (0, 1), and tx + (1 t)y F. As before,
r = Re (tx + (1 t)y)
= t Re (x) + (1 t) Re (y).
But Re (x) r, Re (y) r, so the only way that equality can hold is if x, y F.
Hence F is a face of K.
2
The following result is a crucial step in the proof of the Krein-Milman Theorem.
8.13. Lemma. Let (1, T ) be a LCS and ,= K 1 be a compact, convex set.
Then Ext(K) ,= .
Proof. Let = F K : ,= F is a closed face of K, and partially order by
reverse inclusion: i.e. F
1
F
2
if F
2
F
1
. Observe that K and so ,= .
Suppose that ( = F
is a chain in . We claim that F =
is
an upper bound for (. Since F
has the Finite Intersection Property and K

is compact, F ,= . Moreover, each F
is assumed to be closed and convex, and

thus so is F. Suppose that x, y K, t (0, 1), and tx + (1 t)y F. Then
tx + (1 t)y F
for each . But F
is a face of K, so x, y F
for all , whence

x, y F and F is face of K. Clearly it is an upper bound for (.
By Zorns Lemma, contains a maximal element, say E. Since E , it is
non-empty, convex, and closed in K, hence compact. We claim that E is a singleton
set, and therefore corresponds to an extreme point of K.
Suppose to the contrary that there exist x, y E with x ,= y. By the Hahn-
Banach Theorem 05 (Theorem 6.41), there exists a continuous linear functional
1
so that
Re (x) > Re (y).
Since E is non-empty, convex and compact, we can apply Lemma 8.12. Let r =
supRe (w) : w E, and set H = x E : Re (x) = r. Then H is a non-empty,
compact face of E, and hence of K.
But at least one of x and y does not belong to H, and so E < H, contradicting
the maximality of E. Thus E = e is a singleton set, and e Ext (E) Ext(K),
proving that the latter is non-empty.
2
8.14. Theorem. The Krein-Milman Theorem
Let (1, T ) be a LCS and ,= K 1 be a compact, convex set. Then
K = co(Ext(K)),
the closed, convex hull of the extreme points of K.
Proof. By Lemma 8.13, Ext(K) ,= . Thus ,= co(Ext(K)) K, as K is closed
and convex.
Suppose that m K(co(Ext(K))). By the Hahn-Banach Theorem (Theo-
rem 6.41), there exists 1
and real numbers > so that

Re (m) > Re (b) for all b co(Ext(K)).
Let s := supRe (w) : w K. Then s Re (m) , and L := z K :
Re (z) = s is a non-empty, compact face of K, by Lemma 8.12. But then ,= L
is a compact, convex set in 1, and so by Lemma 8.13, Ext(L) ,= . Furthermore,
Ext(L) Ext(K), by virtue of the fact that L is a face of K (see Remark 8.9 (c)).
Hence there exists e Ext(L) co(Ext(K)) so that
Re (e) = s > Re (b) for all b co(Ext(K)),
an obvious contradiction.
It follows that Kco(Ext(K)) = , and thus K = co(Ext(K)).
2
8.15. Corollary. Let (1, T ) be a LCS and ,= K 1 be a compact, convex
set. If 1
, then there exists e Ext(K) so that

Re (w) Re (e) for all w K.
Proof. Let r := supRe (w) : w K. By Lemma 8.11, F = x K : Re (x) =
r is a non-empty, compact face of K. By the Krein-Milman Theorem, Theo-
rem 8.14, Ext(F) ,= . Let e Ext(F). Then e Ext(K), and
Re (w) r = Re (e) for all w K.
2
Equipped with the Krein-Milman Theorem 8.14 above, we are able to extend
Corollary 7.23.
8.16. Corollary. Let X be a Banach space and suppose that / X
is weak
-
closed and bounded. Then / is weak
-compact. If / is also convex, then / =

co
w
(Ext/).
I want to go back to Brazil, get married, have lots of kids, and just
be a couch tomato.
Ana Beatriz Barros
9. THE CHAPTER OF NAMED THEOREMS 111
9. The chapter of named theorems
I believe that sex is one of the most beautiful, natural, wholesome
things that money can buy.
Steve Martin
9.1. In general, if f : X Y is a continuous map between topological spaces
X and Y , one does not expect f to take open sets to open sets. Despite this, we
have seen that if 1 is TVS and J is a closed subspace of 1, then the quotient map
does just this.
The Open Mapping Theorem extends this result to surjections of Banach spaces.
Many of the theorems in this Chapter are a consequence - either direct or indirect
- of the Open Mapping Theorem. We begin with a Lemma which will prove crucial
in the proof of the Open Mapping Theorem.
9.2. Lemma. Let X and Y be Banach spaces and suppose that T B(X, Y).
If Y
1
TX
m
for some m 1, then Y
1
TX
2m
.
Proof. First observe that Y
1
TX
m
implies that Y
r
TX
rm
for all r > 0.
Choose y Y
1
. Then there exists x
1
X
m
so that |y Tx
1
| < 1/2. Since
y Tx
1
Y
1/2
TX
m/2
, there exists x
2
X
m/2
so that |(y Tx
1
) Tx
2
| < 1/4.
More generally, for each n 1, we can nd x
n
X
m/2
n1 so that
|y
n
j=1
Tx
j
| <
1
2
n
.
Since X is complete and
n=1
|x
n
|
n=1
m
2
n1
= 2m, we have x =

n=1
x
n

X
2m
. By the continuity of T,
Tx = T
_

n=1
x
n
_
= lim
N
T
_
N
n=1
x
n
_
= y.
2
9.3. Theorem. The Open Mapping Theorem
Let X and Y be Banach spaces and suppose that T B(X, Y) is a surjection.
Then T is an open map - i.e. if G X is open, then TG Y is open.
Proof. Since T is surjective, Y = TX =
n=1
TX
n

n=1
TX
n
. Now Y is a
complete metric space, and so by the Baire Category Theorem , there exists m 1
so that the interior int(TX
m
) ,= . As TX
m
is dense in TX
m
, we can choose
y int(TX
m
) TX
m
.
Let > 0 be such that y + V
Y
(0) int(TX
m
) (TX
m
). Then V
Y
(0)
y +TX
m
TX
m
+TX
m
TX
2m
. (This last step uses the linearity of T.)
Thus Y
/2
V
Y
(0) TX
2m
. By Lemma 9.2 above,
Y
/2
TX
4m
,
or equivalently,
TX
r
Y
r/8m
for all r > 0.
Suppose that G X is open and that y TG, say y = Tx for some x G.
Since G is open, we can nd > 0 so that x +V
X
(0) G. Thus
TG Tx +T(V
X
(0))
y +TX
/2
y +Y
/16m
y +V
Y
/16m
(0).
Thus y TG implies that y int TG, and so TG is open.
2
9.4. Corollary. The Inverse Mapping Theorem
Let X and Y be Banach spaces and suppose that T B(X, Y) is a bijection.
Then T
1
is continuous, and so T is a homeomorphism.
Proof. If G X is open, then (T
1
)
1
(G) = TG is open in Yby the Open Mapping
Theorem above. Hence T
1
is continuous.
2
9.5. Corollary. The Closed Graph Theorem
Let X and Y be Banach spaces and suppose that T : X Y is linear. If the
graph
((T) := (x, Tx) : x X
is closed in X
1
Y, then T is continuous.
Proof. The
1
norm on XY was chosen only so as to induce the product topology
on XY. We could have used any equivalent norm (for example, the
2
or
norms).
Let
1
: X
1
Y X be the canonical projection
1
(x, y) = x, (x, y) XY. Then
1
is clearly linear, and
|x|
X
= |
1
(x, y)|
X
|x|
X
+|y|
Y
= |(x, y)|,
so that |
1
| 1. Moreover, ((T) is easily seen to be a linear manifold in X
1
Y,
and by hypothesis, it is closed and hence a Banach space.
The map
G
: ((T) X
(x, Tx) x
is a linear bijection with |
G
| = |
1
[
G(T)
| |
1
| 1.
By the Inverse Mapping Theorem 9.4 above,
1
G
is also continuous, hence
bounded.
Thus
|Tx|
Y
|x|
X
+|Tx|
Y
= |(x, Tx)| = |
1
G
(x)| |
1
G
| |x|
for all x X, and therefore |T| |
1
G
| < . That is, T is continuous.
2
Let X and Y be Banach spaces. A linear map T : X Y is continuous if and
only if for all sequences (x
n
)
n
in X converging to x X, we have lim
n
Tx
n
= Tx. Of
course, given a linear map T : X Y, and given a sequence (x
n
)
n
converging to x,
there is no reason a priori to assume that (Tx
n
)
n
converges to anything at all in Y.
The following Corollary is interesting in that part (c) tells us that it in checking to
see whether or not T is continuous, it suces to assume that lim
n
Tx
n
exists, and
that we need only verify that the limit is the expected one, namely Tx. Linearity
of T further reduces the problem to checking this condition for x = 0.
9.6. Corollary. Let X and Y be Banach spaces and T : X Y be linear. The
(a) The graph ((T) is closed.
(b) T is continuous.
(c) If lim
n
x
n
= 0 and lim
n
Tx
n
= y, then y = 0.
Proof.
(a) implies (b): This is just the Closed Graph Theorem above.
(b) implies (c): This is clear.
(c) implies (a): Suppose that ((x
n
, Tx
n
))
n=1
is a sequence in ((T) which con-
verges to some point (x, y) X
1
Y. Then, in particular, lim
n
x
n
= x,
and so lim
n
(x
n
x) = 0. Also, lim
n
Tx
n
= y, so lim
n
T(x
n
x) =
y Tx exists. By our hypothesis, y Tx = 0, or equivalently y = Tx. This
in turn says that (x, y) = (x, Tx) ((T), and so the latter is closed.
2
Recall that a two closed subspaces Y and Z of a Banach space X are said to
topologically complement each other if X = YZ.
9.7. Lemma. Two closed subspaces Y and Z of a Banach space X topologically
complement each other if and only if the map
: Y
1
Z X
(y, z) y +z
is a homeomorphism of Banach spaces.
Proof. First note that the norms on Y and on Z are nothing more than the
restrictions to these spaces of the norm on X.
Suppose that Y and Z are topologically complementary subspaces of X. That
is linear is clear. Moreover, since Y and Z are complementary subspaces, it is easy
to see that is a bijection. Hence
|(y, z)| = |y +z|
|y| +|z|
= |(y, z)|,
so that is a contraction. By the Inverse Mapping Theorem 9.4,
1
is continuous,
and so is a homeomorphism.
Conversely, suppose that is a homeomorphism. Now ran = Y+Z = X, since
is surjective, and if w Y Z, then (w, w) ker = (0, 0), so w = 0. Hence X
is the algebraic direct sum of Y and Z. Since Y and Z are closed in X, they are also
topologically complemented.
2
The next result extends our results from Section 3, where we showed that for
a closed subspace / of a Hilbert space H, there exists an orthogonal projection
P B(H) whose range is / (see Remarks 3.7).
9.8. Proposition. Let X a Banach space and let Y and Z be topologically
complementary subspaces of X. For each x X, denote by y
x
and z
x
the unique
elements of Y and Z respectively such that x = y
x
+ z
x
. Dene E : X Y via
Ex = E(y
x
+z
x
) = y
x
for all x X. Then
(a) E is a continuous linear map. Moreover, E = E
2
, ranE = Y, and
ker E = Z.
(b) Conversely, if E B(X) and E = E
2
, then M= ranE and N = ker E are
topologically complementary subspaces of X.
Proof.
(a) From Lemma 9.7 above, we know that there exists a linear homeomorphism
: Y
1
Z X. Consider the map
Y
: Y
1
Z Y
(y, z) y.
It is clear that
Y
is linear and contractive, and so
Y
is continuous. As
such, the map
E :=
Y

1
: X Y
x y
x
is clearly linear (being the composition of linear functions), and
|Ex| = |
Y

1
| |
Y
| |
1
| < ,
so that E is bounded - i.e. E is continuous. That ranE = Y and ker E = Z
are left as exercises.
(b) Since E is assumed to be continuous, N is closed. Now I E is also
continuous, and ranE = ker(I E), so ranE is also closed. If z ranE
ker E, then z = Ew for some w X, so z = E
2
w = Ez = 0. Furthermore,
for any x X, x = Ex + (I E)x ranE + ker E. Hence M and
N are algebraically complemented closed subspaces of X; i.e. they are
topologically complemented.
2
9.9. Remark. A linear map E B(X) is said to be idempotent if E = E
2
.
We point out that the term projection is often used in this context, although in
the Hilbert space setting, the meaning of projection is slightly dierent.
The above Proposition says that a subspace Y of a Banach space X is comple-
mented if and only if it is the range of a bounded idempotent in B(X).
Nobody in the game of football should be called a genius. A genius is

somebody like Norman Einstein.
Joe Theismann
10. Operator Theory
I got kicked out of ballet class because I pulled a groin muscle. It
wasnt mine.
Rita Rudner
10.1. Much of the work that has been done in Banach space theory has focussed
on the study of the geometric structure of Banach spaces. For example, people have
been interested in how close two Banach spaces are to being isomorphic as Banach
spaces (say, in terms of the Banach-Mazur distance between them), and they have
been interested in nding subspaces of a Banach space which have or are close to
having a prescribed geometric structure. Other interesting questions in this area
involve the study of how well nite-dimensional subspaces of one Banach space can
be embedded in a second Banach space, and yet others involve the search for nice
bases for a Banach space (for eg., the search for Schauder bases for quotients of
Banach spaces), or the renorming of Banach spaces by equivalent norms. This list
is anything but inclusive.
In contrast, the study of Hilbert spaces focusses very much on the structure of the
bounded linear operators acting on the space. This is in part because Hilbert spaces
are so well-behaved. This means that one pretends to understand the underlying
space H, and instead focusses upon understanding the more complicated structure,
namely B(H). Because Hilbert spaces are so well-behaved, there is a chance of
getting interesting and deep results about subsets and subalgebras of B(H).
Of course, one can also study the space B(X) of bounded linear operators acting
on a Banach space X. Here one is often interested in how the structure of the
underlying space X determines the operators in B(X). In this Section we shall
examine the notion of compactness of operators on a Banach space. We begin,
however, with the notion of the Banach space adjoint of an operator.
Let X and Y be Banach spaces. Let T B(X, Y). Given y
, the map
x y
(Tx)
is a linear functional on X. Let us denote by T
X
#
the functional on X
determined by this formula, namely:
T
(x) = y
(Tx) for all x X.

Observe that
|T
(x)| = |y
(Tx)| |y
| |T| |x|,
and so |T
| |T| |y
| < , implying that T
. Furthermore, the map

T
is easily seen to be linear (by denition of linear combinations of functionals on X).

10. OPERATOR THEORY 117
10.2. Denition. Let X and Y be Banach spaces and let T B(X, Y). The
map T
dened above is called the Banach space adjoint of T.

Many authors adopt the following notation for the action of a functional on a
vector, namely: given x X and x
, they write x, x
) to denote x
(x). In
this notation, the equation T
(x) = y
(Tx) for all x X, y
becomes:
< x, T
> = < Tx, y
>
for all x X, y
. The reason for using this notation will become apparent

when we look at Hilbert space adjoints below.
10.3. Proposition. Let X, Y, Z be Banach spaces, S, T B(X, Y), and let
R B(Y, Z). Then
(a) for all k
1
, k
2
K, we have (k
1
S +k
2
T)
= k
1
S
+k
2
T
;
(b) (R T)
= T
.
Proof. Let x X, y
, and z
. Then
(a) Observe that
(k
1
S +k
2
T)
(x) = y
((k
1
S +k
2
T)x)
= y
(k
1
(Sx) +k
2
(Tx))
= k
1
y
(Sx) +k
2
y
(Tx)
= k
1
S
(x) +k
2
T
(x)
= [(k
1
S
+k
2
T
)y
](x).
Since this is true for all x X and y
, we conclude that (k
1
S+k
2
T)
=
k
1
S
+k
2
T
.
(b) Now
(R T)
(x) = z
((R T))(x)
= z
(R(Tx))
= R
(Tx)
= T
(x).
Again, since this is true for all x X and then for all z
, we nd that
(R T)
= T
.
2
10.4. Theorem. Let T B(X, Y), where X and Y are Banach spaces. Then
|T
| = |T|.
Proof.
For any y
, we have
|T
| = sup[T
(x)[ : x X, |x| = 1
= sup[y
(Tx)[ : x X, |x| = 1
sup|y
| |Tx| : x X, |x| = 1
= |y
| |T|.
Thus we see that |T
| |T|.
Next, let x X. By the Hahn-Banach Theorem, we can choose y
such
that y
(Tx) = |Tx| and |y
| = 1. Then
|Tx| = y
(Tx)
= T
(x)
|T
| |x|
|T
||x|.
Thus |T| |T
|.
Combining this with the previous estimate, we have that |T
| = |T|.
2
10.5. Proposition. Let X = C
n
and A B(X) M
n
. Then the matrix of the
Banach space adjoint A
of A with respect to the dual basis coincides with A

t
, the
transpose of A.
Proof. Recall that X
X. We then let e
i
n
i=1
be a basis for X and let f
j
n
j=1
be
the corresponding dual basis for X
; that is, f
j
(e
i
) =
ij
, where
ij
is the Kronecker
delta function. Let x X. Dene
j
= f
j
(x), 1 j n.
Writing the matrix of A B(X) as [a
ij
], we have
Ae
j
= [a
ij
]
_
_
0
0
.
.
0
1
0
.
.
0
_
_
=
_
_
a
1j
a
2j
.
.
a
j1 j
a
jj
a
j+1 j
.
.
a
nj
_
_
=
n
k=1
a
kj
e
k
.
Thus a
ij
= f
i
(Ae
j
).
Now A
B(X
) M
n
, and so we can also write the matrix [
ij
] for A
with
respect to f
j
n
j=1
. Paralleling the above computation for [
ij
], we obtain:
A
f
j
=
n
k=1
kj
f
k
,
and thus
ij
= (A
f
j
)(e
i
) = f
j
(Ae
i
) = a
ji
.
In particular, the matrix for A
with respect to f
j
n
j=1
is simply the transpose of
the matrix for A with respect to e
j
n
j=1
.
2
10.6. Denition. Let X and Y be Banach spaces, and T B(X, Y). Then T
is said to be compact if T(X
1
) is compact in Y. The set of compact operators from
X to Y is denoted by /(X, Y), and if Y = X, we simply write /(X).
Recall that a subset K of a metric space L is said to be totally bounded if
for every > 0 there exists a nite cover V
(y
i
)
n
i=1
of K with y
i
K, 1 i n,
where V
(y
i
) = z L : dist (z, y
i
) < .
We leave it as an exercise for the reader to show that if E is a subset of L and
E is totally bounded, then so is E.
10.7. Proposition. Let X and Y be Banach spaces, and T B(X, Y). The
(a) T is compact;
(b) T(F) is compact in Y for all bounded subsets F of X;
(c) If (x
n
)
n
is a bounded sequence in X, then (Tx
n
)
n
has a convergent subse-
quence in Y;
(d) T(X
1
) is totally bounded.
Proof.
(a) implies (b): Suppose that T is compact. Since multiplication by a non-zero
scalar is a homeomorphism in any TVS, TX
r
is compact for all r > 0.
If F X is bounded, then F X
r
for some r > 0, and therefore
TF TX
r
. Since TF is a closed subset of a compact set, it is compact.
(b) implies (c): Suppose that TF is compact for all bounded sets F X. Let
(x
n
)
n
be a bounded sequence in X. By hypothesis, T(x
n
)
n
= (Tx
n
)
n
is
compact in the complete metric space Y. Since compactness and sequen-
tial compactness agree on metric spaces, (Tx
n
)
n
has a subsequence which
converges.
(c) implies (d): If TX
1
is not totally bounded, then there exists > 0 for which
no nite -net exists. Choose y
1
TX
1
. Then V
(y
1
) is not an -net for
TX
1
, and so there exists y
2
TX
1
V
(y
1
). More generally, we can continue
in this manner so that for all n 1, we can nd y
n
TX
1
(
n1
j=1
V
(y
j
)).
The sequence (y
n
)
n
can not have any convergent subsequences, because
|y
n
y
m
| for all m ,= n 1.
On the other hand, y
n
TX
1
, say y
n
= Tx
n
for some x
n
X
1
, n 1,
and so |Tx
n
Tx
m
| for all m ,= n 1. Thus the failure of (d) implies
the failure of (c), or equivalently, (c) implies (d).
(d) implies (a): By the remarks preceding the Theorem, TX
1
is totally bounded
as well. Since it is a closed subset of the complete metric space Y, it is
complete. But a metric space is compact if and only if it is complete and
totally bounded. Thus TX
1
is compact - i.e. T is compact.
2
10.8. Theorem. Let X and Y be Banach spaces. Then /(X, Y) is a closed
subspace of B(X, Y).
Proof. Let k
1
, k
2
K and let K
1
, K
2
/(X, Y). Let (x
n
)
n
be a bounded sequence
in X. Then K
1
generates a convergent subsequence, say (K
1
(x
n(j)
))
j
. Similarly, K
2
generates a convergent subsequence from (x
n(j)
)
j
, say (K
2
(x
n(j(i))
))
i
.
Then ((k
1
K
1
+k
2
K
2
)(x
n(j(i))
))
i
is a convergent subsequence in Y. From part (c)
of Proposition 10.7, k
1
K
1
+k
2
K
2
/(X, Y).
Now we show that /(X, Y) is closed. Suppose K
n
/(X, Y) for n 1 and
lim
n
K
n
= K B(X, Y). We show that K(X
1
) is totally bounded. First let
> 0, and choose N > 0 such that |K
N
K| < /3.
Since K
N
(X
1
) is totally bounded, we can nd y
i
= K
N
(x
i
)
M
i=1
such that
V
/3
(y
i
)
M
i=1
is a nite cover of K
N
(X
1
). Thus for all x X
1
, |K
N
(x) K
N
(x
j
)| <
/3 for some 1 j = j(x) M. Then
|K(x) K(x
j
)| = |K(x) K
N
(x) +K
N
(x) K
N
(x
j
) +
K
N
(x
j
) K(x
j
)|
|K K
N
| |x| +|K
N
(x) K
N
(x
j
)| +
|K
N
K| |x
j
|
(/3) + (/3) + (/3)
= .
Thus K(X
1
) is totally bounded and so K /(X, Y).
2
10.9. Theorem. Let W, X, Y, and Z be Banach spaces. Suppose R B(W, X),
K /(X, Y), and T B(Y, Z). Then TK /(X, Z) and KR /(W, Y).
Proof. Let X
1
denote the unit ball of X. Note that K(X
1
) is compact and T is
continuous, so that T(K(X
1
)) is compact and therefore closed.
T K(X
1
) = T(K(X
1
))
T(K(X
1
))
= T(K(X
1
)).
Since T K(X
1
) is a closed subset of the compact set T(K(X
1
)), it is compact as
well. Thus TK /(X, Z).
Now if W
1
is the unit ball of W, then
KR(W
1
) = K(R(W
1
));
but R(W
1
) is bounded since R is, and so by Proposition 10.7, KR(W
1
) is compact.
Thus KR /(W, Y).
2
10.10. Corollary. If X is a Banach space, then /(X) is a closed, two-sided
ideal of B(X).
10.11. Proposition. Let X and Y be Banach spaces and assume that K
/(X, Y). Then K(X) is closed in Y if and only if dim K(X) is nite.
Proof. K(X) = ranK is easily seen to be a submanifold of Y. Since nite-
dimensional manifolds are always closed, we nd that dim K(X) < implies K(X)
is closed.
Now assume that K(X) is closed. Then K(X) is a Banach space and the map
K
0
: X K(X)
x Kx
is a surjection. By the Open Mapping Theorem, Theorem 9.3, it is also an open
map. In particular, K
0
(int X
1
) is open in K(X) and 0 K
0
(int X
1
). Let G be an
open ball in K(X) centred at 0 and contained in K
0
(int X
1
). Then K
0
(X
1
) = K(X
1
)
is compact, hence closed, and also contains G. Thus G is compact in K(X) and so
dim K(X) is nite - see Corollary 8.5.
2
10.12. Denition. Let X and Y be Banach spaces. Then F B(X, Y) is said
to be nite rank if dim F(X) is nite. The set of nite rank operators from X to
Y is denoted by T(X, Y).
10.13. Proposition. Let X and Y be Banach spaces. Then T(X, Y) /(X, Y).
Proof. Suppose F T(X, Y). Then FX
1
is closed and bounded in ranF, but ranF
is nite dimensional in Y, as F is nite rank. Thus FX
1
is compact in ranF, and
thus compact in Y as well, showing that F is compact.
2
10.14. Proposition. Let X be a Banach space. Then /(X) = B(X) if and only
if X is nite dimensional.
Proof. If dim X < , then B(X) = T(X) /(X) B(X), and equality follows.
If /(X) = B(X), then I /(X), so I(X
1
) = X
1
is compact. In particular, X is
nite dimensional.
2
10.15. Theorem. Let X and Y be Banach spaces and suppose K /(X, Y).
Then K
/(Y
, X
).
Proof. Let > 0. Then K(X
1
) is totally bounded, so we can nd x
1
, x
2
, . . . , x
n
X
1
such that if x X
1
, then |Kx Kx
i
| < /3 for some 1 i n. Let
R : Y
(C
n
, | |
)
y
(y
(K(x
1
)), y
(K(x
2
)), . . . , y
(K(x
n
))).
Then R T(Y
, C
n
) /(Y
, C
n
), and so R(Y
1
) is totally bounded. Thus we can
nd y
1
, y
2
, . . . , y
m
Y
1
such that if y
1
, then |Ry
Ry
j
| < /3 for some
1 j m. Now
|Ry
Ry
j
| = max
1in
[y
(K(x
i
)) y
j
(K(x
i
))[
= max
1in
[K
(y
)(x
i
) K
(y
j
)(x
i
)[.
Suppose x X
1
. Then |Kx Kx
i
| < /3 for some 1 i n, and
[K
(y
)(x
i
) K
(y
j
)(x
i
)[ < /3 for some 1 j m, so
[K
(y
)(x) K
(y
j
)(x)[ [K
(y
)(x) K
(y
)(x
i
)[ +
[K
(y
)(x
i
) K
(y
j
)(x
i
)[ +
[K
(y
j
)(x
i
) K
(y
j
)(x)[
|y
| |Kx Kx
i
| +/3 +|y
j
| |Kx Kx
i
|
< /3 +/3 +/3 = .
Thus |K
j
| and so K
(Y
1
) is totally bounded. We conclude that
K
/(Y
, X
).
2
The fact that a Hilbert space is isometrically isomorphic (via a conjugate-linear
map) to its own dual allows us to dene a separate notion of an adjoint for operators
acting on these spaces.
10.16. Theorem. Let H be a Hilbert space and let T B(H). Then there exists
a unique operator T
B(H), called the Hilbert space adjoint of T, satisfying

Tx, y) = x, T
y)
for all x, y H.
Proof. Fix y H. Then the map
y
: H C
x Tx, y)
is a linear functional and so there exists a vector z
y
H such that
y
(x) = Tx, y) = x, z
y
)
for all x H. Dene a map T
: H H by T
y = z
y
. We leave it to the reader to
verify that T
is in fact linear, and we concentrate on showing that it is bounded.

To see that T
is bounded, consider the following. Let y H, |y| = 1. Then

Tx, y) = x, T
y) for all x H, so
|T
y|
2
= T
y, T
y)
= T T
y, y)
|T| |T
y| |y|.
Thus |T
y| |T|, and so |T
| |T|.
Now T
is unique, for if there exists A B(H) such that Tx, y) = x, T
y) =
x, Ay) for all x, y H, then x, (T
A)y) = 0 for all x, y H, and so (T
A)y = 0
for all y H, i.e. T
= A.
2
10.17. Corollary. Let T B(H), where H is a Hilbert space. Then (T
= T.
It follows that |T| = |T
|.
Proof. For all x, y H, we get
x, (T
y) = T
x, y)
= y, T
x)
= Ty, x)
= x, Ty),
and so (T
= T. Applying Theorem 10.16, we get

|T| = |(T
| |T
| |T|,
and so |T| = |T
|.
2
10.18. Remark. The above proof also shows that for a Hilbert space H and
A, B B(H), we have (AB)
= B
. The adjoint operator

: B(H) B(H)
is an example of an involution on a Banach algebra. Namely, for all , C and
A, B B(H), we obtain
(i) (A)
= A
;
(ii) (A+B)
= A
+B
; and
(iii) (AB)
= B
.
(iv) (A
= A.
10.19. Proposition. Let H = C
n
and T B(H) M
n
. Then the matrix of
T
with respect to any orthonormal basis is the conjugate transpose of that of T.

Proof. Let e
i
n
i=1
be an orthonormal basis for H. With respect to this basis, T
has a matrix [t
ij
]
1i,jn
and T
has a matrix [r
ij
]
1i,jn
.
But t
ij
= Te
j
, e
i
) = e
j
, T
e
i
) = T
e
i
, e
j
) = r
ij
, completing the proof.
2
10.20. Proposition. Let H be a Hilbert space and T B(H). Then (ranT)
=
ker T
. In particular, therefore:
(i) ranT = (ker T
;
(ii) ranT is not dense in H if and only if ker T
,= 0.
Proof. Let y H. Then
y ker T
if and only if for all x H, 0 = (x, T
y),
if and only if for all x H, 0 = (Tx, y),
if and only if y (ranT)
.
The second statement is now obvious.
2
10.21. The set of compact operators acting on a Hilbert space is more tractable
in general than the set of compact operators acting on an arbitrary Banach space.
One of the reasons for this is the characterization given below. Recall that the set
of nite rank operators acting on a Banach space X is denoted by T(X).
10.22. Theorem. Let H be a Hilbert space and let K B(H). The following
are equivalent:
(i) K is compact;
(ii) K
is compact;
(iii) There exists a sequence F
n
n=1
T(H) such that K = lim
n
F
n
.
Proof.
(i) (iii) Let B
1
denote the unit ball of H, and let > 0. Since K(B
1
) is
compact, it must be separable (i.e. it is totally bounded). Thus /= ranK
is a separable subspace of H, and thus possesses an orthonormal basis
e
n
n=1
.
Let P
n
denote the orthogonal projection of H onto spane
k
n
k=1
. Set
F
n
= P
n
K, noting that each F
n
is nite rank. We now show that K =
lim
n
F
n
.
Let x H and consider y = Kx /, so that lim
n
|P
n
y y| = 0.
Thus lim
n
|F
n
x Kx| = lim
n
|P
n
y y| = 0. Since K is com-
pact, K(B
1
) is totally bounded, so we can choose x
k
m
k=1
B
1
such that
K(B
1
)
m
k=1
B(Kx
k
, /3), where given z H and > 0, B(z, ) = w
H : |w z| < .
If |x| 1, choose i such that |Kx
i
Kx| < /3. Then for any n > 0,
|Kx F
n
x|
|Kx Kx
i
| +|Kx
i
F
n
x
i
| +|F
n
x
i
F
n
x|
< /3 +|Kx
i
F
n
x
i
| +|P
n
| |Kx
i
Kx|
< 2/3 +|Kx
i
F
n
x
i
|.
Choose N > 0 such that |Kx
i
F
n
x
i
| < /3, 1 i m for all n > N.
Then |Kx F
n
x| 2/3 + /3 = . Thus |K F
n
| < 3 for all n > N.
Since > 0 was arbitrary, K = lim
n
F
n
.
(iii) (ii) Suppose K = lim
n
F
n
, where F
n
is nite rank for all n 1.
Note that F
n
is also nite rank (why?), and that |K
n
| = |K F
n
|
for all n 1, which clearly implies that K
= lim
n
F
n
, and hence that
K
is compact.
(ii) (i) Since K compact implies K
is compact from above, we deduce that

K
compact implies (K
= K is compact, completing the proof.

2
We can restate the above Theorem more succinctly by saying that /(H) is the
norm closure of the set of nite rank operators on H. This is an extraordinarily
useful result.
10.23. Remark. Contained in the above proof is the following interesting
observation. If K is a compact operator acting on a separable Hilbert space H, then
for any sequence P
n
n=1
of nite rank projections tending strongly (i.e. pointwise)
to the identity, |K P
n
K| tends to zero. By considering adjoints, we nd that
|K KP
n
| also tends to zero.
Let > 0, and choose N > 0 such that n N implies |K KP
n
| < /2 and
|K P
n
K| < /2. Then for all n N we get
|K P
n
KP
n
| |K KP
n
| +|KP
n
P
n
KP
n
|
|K KP
n
| +|K P
n
K| |P
n
|
< /2 +/2 = .
It follows that if H has an orthonormal basis indexed by the natural numbers,
say e
n
n=1
, then the matrix for K with respect to this basis comes within of
the matrix for P
N
KP
N
. In other words, K virtually lives on the top left-hand
corner.
Alternatively, if Hhas an orthonormal basis indexed by the integers, say f
n
nZ
,
and we let P
n
denote the orthogonal projection onto spane
k
n
k=n
, then the matrix
for K with respect to this basis can be arbitrarily well estimated by a suciently
large but nite central block.
10.24. Example. Let H be a separable Hilbert space with orthonormal basis
e
n
n=1
. Let d
n
n=1
be a bounded sequence and consider the diagonal operator
D B(H) dened locally by De
n
= d
n
e
n
and extended to all of H by linearity and
continuity.
Then D /(H) if and only if lim
n
d
n
= 0.
10.25. Example. Let H = L
2
([0, 1], dx), and consider the function k(x, t)
L
2
([0, 1][0, 1], dm), where dm represents Lebesgue planar measure. Then we dene
a Volterra operator
V : L
2
([0, 1], dx) L
2
([0, 1], dx)
(V f)(x) =
_
1
0
f(t) k(x, t) dt.
(The classical Volterra operator has k(x, t) = 1 if x t, and k(x, t) = 0 if x < t.)
Now for f L
2
([0, 1], dx) we have
|V f|
2
=
_
1
0
[V f(x)[
2
dx
=
_
1
0
_
1
0
f(t) k(x, t)dt
2
dx
_
1
0
__
1
0
[f(t) k(x, t)[dt
_
2
dx
_
1
0
|f|
2
2
_
1
0
[k(x, t)[
2
dtdx by the Cauchy-Schwartz Inequality
= |f|
2
2
|k|
2
2
,
so that |V | |k|
2
.
Let / denote the algebra of continuous functions on [0, 1] [0, 1] which can be
resolved as g(x, t) =

n
i=1
u
i
(x) w
i
(t). Then / is an algebra which separates points,
contains the constant functions, and is closed under complex conjugation. By the
Stone-Weierstra Theorem, given > 0 and h (([0, 1] [0, 1]), there exists g /
such that |h g|
2
|h g|
< . But since (([0, 1] [0, 1]) is dense (in the

L
2
-topology) in L
2
([0, 1] [0, 1], dm), / must also be dense (in the L
2
-topology) in
L
2
([0, 1] [0, 1], dm).
Let > 0. For k as above, choose g / such that |k g|
2
< . Dene
V
0
: L
2
([0, 1], dx) L
2
([0, 1], dx)
V
0
f(x) =
_
1
0
f(t) g(x, t)dt.
From above, we nd that |V V
0
| |k g|
2
< .
To see that V
0
is nite rank, consider the following; rst, g(x, t) =

n
i=1
u
i
(x) w
i
(t).
If we set /= span
1in
u
i
, then /is a nite dimensional subspace of L
2
([0, 1], dx).
Moreover,
V
0
f(x) =
_
1
0
f(t) g(x, t)dt
=
n
i=1
__
1
0
f(t) w
i
(t)dt
_
u
i
(x),
so that V
0
f /.
Thus V can be approximated arbitrarily well by elements of the form V
0

T(L
2
([0, 1], dx), and so V is compact.
11. APPENDIX TOPOLOGICAL BACKGROUND 127
11. Appendix topological background
A child of ve could understand this. Fetch me a child of ve.
Groucho Marx
11.1. At the heart of analysis is topology. A thorough study of topology is
beyond the scope of this course, and we refer the reader to the excellent book [Wil70]
General Topology, written by my former colleague Stephen Willard. The treatment
of topology in this section borrows heavily from his book.
We shall only give the briefest of overviews of this theory - assuming that the
student has some background in metric and norm topologies. We shall only cover
the notions of weak topologies and nets, which are vital to the study of Functional
Analysis.
11.2. Denition. A topology T on a set X is a collection of subsets of X,
called open sets, which satisfy the following:
(i) X, T - i.e. the entire space and the empty set are open;
(ii) If G
T , then
T - i.e. arbitrary unions of open sets are

open;
(iii) If n 1 and G
k
n
k=1
T , then
n
k=1
G
k
T - i.e. nite intersections of
open sets are open
A set F is called closed if XF is open. We call (X, T ) (or more informally, we
call X) a topological space.
It is useful to observe that the intersection of a collection T
of topologies
on X is once again a topology on X.
11.3. Example.
(i) Let X be any set. Then T = , X is a topology on X, called the trivial
topology on X.
(ii) At the other extreme of the topological spectrum, if X is any non-empty
set, then T = T(X), the power set of X, is a topology on X, called the
discrete topology on X.
(iii) Let X = a, b, and set T = , a, a, b. Then T is a topology on X.
(iv) Let (X, d) be a metric space. Let
T = G X : for all g G there exists > 0 such that
V
(g) := y X : d(x, y) < G.

Then T is a topology, called the metric topology on X induced by d.
This is the usual topology one thinks of when dealing with metric spaces,
but as we shall see, there can be many more.
(v) Let X be any non-empty set. Then
T
cf
= Y X : XY is nite
is a topology on X, called the co-nite topology on X.
11.4. Denition. Let (X, T ) be a topological space, and x X. A set U
is called a neighbourhood (abbreviated nbhd) of x if there exists G T so that
x G U. The reader is cautioned that some authors require nbhds to be open - we
do not. The neighbourhood system at x is |
x
:= U X : U is a nbhd of x.
The following result from [Wil70] illustrates the importance of nbhd systems.
11.5. Theorem. Let (X, T ) be a topological space, and x X. Then:
(a) If U |
x
, then x U.
(b) If U, V |
x
, then U V |
x
.
(c) If U |
x
, there exists V |
x
such that U |
y
for each y V .
(d) If U |
x
and U V , then V |
x
.
(e) G X is open if and only if G contains a nbhd of each of its points.
Conversely, if in a set X a non-empty collections |
x
of subsets of X is assigned
to each x X so as to satisfy conditions (a) through (d), and if we use (e) to dene
the notion of an open set, the result is a topology on X in which the nbhd system at
x is precisely |
x
.
Because of this, it is clear that if we know the nbhd system of each point in X,
then we know the topology of X.
There are a number of natural separation axioms that a topological space might
satisfy.
11.6. Denition. Let (X, T ) be a topological space.
(i) (X, T ) is said to be T
0
if for every x, y X such that x ,= y, either there
is a neighbourhood U
x
of x with y , U
x
or there is a neighbourhood U
y
of
y with x , U
y
.
(ii) (X, T ) is said to be T
1
if for every x, y X such that x ,= y, there are
neighbourhoods U
x
of x and U
y
of y with y , U
x
and x , U
y
.
(iii) (X, T ) is said to be T
2
(or Hausdor) if for every x, y X such that
x ,= y, there are neighbourhoods U
x
of x and U
y
of y with U
x
U
y
= .
We say that two subsets A and B of X can be separated by T if there exist U, V T
with A U, B V and U V = .
(iv) (X, T ) is said to be regular if whenever F X is closed and x , F,
F and x can be separated.
(v) (X, T ) is said to be normal if whenever F
1
, F
2
X are closed and disjoint,
then F
1
and F
2
can be separated.
(vi) (X, T ) is said to be T
3
if it is T
1
and regular.
(vii) (X, T ) is said to be T
4
if it is T
1
and normal.
11.7. Theorem. Let (X, d) be a metric space. Then X, equipped with the
metric topology, is T
4
.
11.8. Denition. Let (X, T ) be a topological space. A neighbourhood base
B
x
at a point x X is a collection B
x
|
x
so that U |
x
implies that there exists
B B
x
so that B U. We refer to the elements of B
x
as basic nbhds of the point
x.
The importance of neighbourhood bases is that all open sets can be constructed
from them, as we shall soon see.
11.9. Example. Consider (X, d) be a metric space equipped with the metric
topology T . For each x X, x a sequence r
n
(x)
n=1
of positive real numbers
such that lim
n
r
n
(x) = 0 and consider B
x
= V
rn
(x) : n 1. Then B
x
is a nbhd
base at x for each x X.
11.10. Denition. Let (X, T ) be a topological space. A base for the topology
is a collection B T so that for every G T there exists ( B so that G = B :
B (. That is, every open set is a union of elements of B. Note that if ( is empty,
then B : B ( is also empty, so we do not need to include the empty set in our
base. A subbase for the topology is a collection o T such that the collection B of
all nite intersections of elements of o forms a base for T .
As we shall see in the Assignments, any collection ( of subsets of X serves as a
subbase for some topology on X, called the topology generated by (.
11.11. Example. Let (X, T ) be a topological space, and for each x X,
suppose that B
x
is a neighbourhood base at x. Then B :=
xX
B
x
is a base for the
topology T on X.
11.12. Example. Consider R with the usual topology T . The collection B =
(a, b) : a, b R, a < b is a base for T . (You might remember from Real Analysis
that every open set in R is a disjoint union of open intervals - although the fact the
union is disjoint in this setting is a luxury item which we have not built into the
denition of a base in general.)
The collection o = (, a) : a R (b, ) : b R is a subbase for the
usual topology, but is not a base for T .
11.13. Denition. Let (X, T ) be a topological space. A directed set is a set
with a relation that satises:
(i) for all ;
(ii) if
1

2
and
2

3
, then
1

3
; and
(iii) if
1
,
2
, then there exists
3
so that
1

3
and
2

3
.
The relation is sometimes called a direction on .
A net in X is a function P : X, where is a directed set. The point P()
is usually denoted by x
, and we often write (x
to denote the net.

A subnet of a net P : X is the composition P , where : M is a
increasing conal function from a directed set to ; that is,
(a) (
1
) (
2
) if
1

2
(increasing), and
(b) for each , there exists M so that () (conal).
For M, we often write x
for P (), and speak of the subnet (x
.
11.14. Denition. Let (X, T ) be a topological space. The net (x
is said to
converge to x X if for every U |
x
there exists
0
so that
0
implies
x
U.
We write lim
= x, or lim
= x.
This mimics the denition of convergence of a sequence in a metric space.
11.15. Example.
(a) Since N is a directed set under the usual order , every sequence is a net.
Any subsequence of a sequence is also a subnet. The converse to this is
false, however. A subnet of a sequence need not be a subsequence, since
its domain need not be N (or any countable set, for that matter).
(b) Let A be a non-empty set and denote the power set of all subsets of A,
partially ordered with respect to inclusion. Then is a directed set, and
any function from to R is a net in R.
(c) Let T denote the set of all nite partitions of [0, 1], partially ordered by
inclusion (i.e. renement). Let f be a continuous function on [0, 1]; then
to P = 0 = t
0
< t
1
< < t
n
= 1 T, we associate the quantity
L
P
(f) =

n
i=1
f(t
i1
)(t
i
t
i1
). The map f L
P
(f) is a net (T is a
directed set), and from Calculus, lim
PP
L
P
(f) =
_
1
0
f(x)dx.
11.16. Denition. Let (X, T
X
) and (Y, T
Y
) be topological spaces. We say that
a function f : X Y is continuous if f
1
(G) is open in X for all G T
Y
.
That this extends our usual notion of continuity for functions between metric
space is made clear by the following result:
11.17. Proposition. If (X, d
X
) and (Y, d
Y
) are metric spaces with metric space
topologies T
X
and T
Y
respectively, then the following are equivalent for a function
f : X Y :
(a) f is continuous on X, i.e. f
1
(G) T
X
for all G T
Y
.
(b) lim
n
f(x
n
) = f(x) whenever (x
n
)
n=1
is a sequence in X converging to
x X.
As we shall see in the Assignments, sequences are not enough to describe conver-
gence, nor are they enough to characterize continuity of functions between general
topological spaces. On the other hand, nets are sucient for this task, and serve
as the natural replacement for sequences. (The following result also admits a local
version, which we shall also see in the Assignments.)
11.18. Theorem. Let (X, T
X
) and (Y, T
Y
) be topological spaces. Let f : X Y
be a function. The following are equivalent:
(a) f is continuous on X.
(b) Whenever (x
is a net in X which converges to x X, it follows that

(f(x
))
is a net in Y which converges to f(x).

The notion of a weak topology on a set X generated by a family of functions
f
from X into topological spaces (Y
, T
) is of crucial importance in the study of

topological vector spaces and of Banach spaces. It is also vital to the understanding
of the product topology on a family of topological spaces, which we shall see shortly.
11.19. Denition. Let ,= X be a set and (Y
, T
be a family of topo-
logical spaces. Suppose that for each there exists a function f
: X Y
. Set
T = f
.
If o = f
1
(G
) : G
, , then o T(X) and as noted above o

is a subbase for a topology on X, denoted by (X, T), and referred to as the weak
topology on X induced by T.
The main and most important result concerning weak topologies induced by a
family of functions is the following:
11.20. Proposition.
(a) If T is a topology on X and if f
: (X, T ) (Y
, T
) is continuous for all

, then (X, T) T . In other words, (X, T) is the weakest topology
on X under which each f
is continuous.
(b) Let (Z, T
Z
) be a topological space. Then g : (Z, T
Z
) (X, (X, T)) is
continuous if and only if f
g : Z Y
is continuous for all .

11.21. Denition. Let (X
, T
be a collection of topological spaces. The

Cartesian product of the sets X
is
= x :
[ x() X
for each .
As with sequences, we write (x
for x.
The map
: X
(x) = x
is called the th projection map.

The product topology on
is the weak topology on
induced by the
family
. As we shall see in the Assignments, this is the topology which has

as a base the collection B =
, where
(a) U
for all ; and

(b) for all but nitely many , U
= X
.
It should be clear from the denition that in (a), it suces to ask that we take
U
, where B
is a xed base for T
, .
Observe that if U
and U
= X
for all except for

1
,
2
, ...,
n
, then
=
1
1
(U
1
)
1
n
(U
n
).
From this it follows that
1
(U
) : U
, is a subbase for the product

topology, where B
is a xed base (or indeed even a subbase will do) for the topology
on X
.
It is perhaps worth pointing out that it follows from the Axiom of Choice that
if for all we have X
,= , then X ,= .
We leave it to the reader to verify that the product topology on R
n
=
n
k=1
R is
just the usual topology on R
n
.
Bibliography
[Dav88] K.R. Davidson. Nest algebras. Triangular forms for operators on a Hilbert space, volume
191 of Pitman Research Notes in Mathematics. Longman Scientic and Technical, Harlow,
1988.
[KR83] R.V. Kadison and J.R. Ringrose. Fundamentals of the theory of operator algebras I: Ele-
mentary theory. Academic Press, New York, 1983.
[Lin71] J. Lindenstrauss. The geometric theory of the classical banach spaces. Actes du Congr`es
Intern. Math., 1970, Paris, 2:365372, 1971.
[Sch27] J. Schauder. Zur Theorie stetiger Abbildungen in funktionalraumen Math. Z., 26:4765,
1927.
[Tsi74] B.S. Tsirelson. Not every Banach space contains an imbedding of
p
or c0. Functional
Anal. Appl., 8:138141, 1974.
[Whi66] R. Whitley. Projecting m onto c0. The Amer. Math. Monthly, 73:285286, 1966.
[Wil70] Stephen Willard. General Topology. Addison-Wesley Publishing Co., Reading, Mass.-
London-Don Mills, Ont., 1970.
133
Index
C
-algebra, 104
-direct sum, 4
(N), 2
p
-direct sum, 4
p
(N), 2
-norm, 3
c0, 2
p-norm, 2
absolutely summable series, 6
adjoint
Banach space, 116
Hilbert space, 117
ane hyperplane, 83
ane manifold, 83
ane subspace, 83
algebraic dual, 70
algebraically complemented, 34, 80
annihilator, 100
Axiom of Choice, 132
Baire Category Theorem, 93, 111
balanced, 45
Banach space, 1, 3
reexive, 25
Banach space adjoint, 116
Banach-Alaoglu Theorem, 97
Banach-Mazur compactum, 43
Banach-Mazur distance, 31, 43, 116
Banach-Steinhaus Theorem, 94
Bankead, T., 88
Barros, A.B., 110
base
neighbourhood, 129
base for a topology, 129
basic neighbourhoods, 129
basis
orthonormal, 36
Schauder, 22, 36, 40
basis, Hamel, 36
Berra, Yogi, 31
Bessels Inequality, 37
Beurlings spectral radius formula, 42
Beurling, Arne, 42
Bierce, Ambrose, ii
bilateral weighted shift operator, 20
block-disjoint, 13
bounded operator, 17
bounded variation
sequences of, 13
canonical embedding, 25, 81
Cauchy complete, 47
Cauchy net, 47
Cauchy-Schwarz Inequality, 31, 35, 64
Close Graph Theorem, 112
closed set, 127
co-nite topology, 128
compact
locally, 104
compact operator, 116, 119
compactum
Banach-Mazur, 43
compatible, 44
complement
orthogonal, 34
complemented
algebraically, 34, 80
topologically, 34, 79
complete
normed linear space, 1
family of pseudo-metrics, 68
metric, 68
complex-valued Borel measure, 15
condition J, 14
consecutively supported, 13
continuous linear functionals, 22
convergence
pointwise, 91
135
136 INDEX
weak, 90
convex, 56
convexity, 56
diagonal operator, 20
dierentiation operator, 21
dimension, 39
direct sum
p
, 4
directed set, 130
direction, 130
discrete topology, 127
distance
Banach-Mazur, 31, 43
double dual space, 22
dual
algebraic, 70
topological, 70
dual pair, 90
dual space, 10, 22
equivalent norms, 4
Ext(C), 107
extension, 53
extreme point, 107
F. Johns Theorem, 43
face, 107
Fields, W.C., 69
Finite Intersection Property, 96, 109
nite measure, 15
Frechet space, 68
function of bounded variation, 25
functional, 22, 70
gauge, 56
Minkowski, 14, 56, 82
positive linear, 104
functional calculus, 42
gauge functional, 56
Goldstines Theorem, 98
Goldwyn, Samuel, ii
Gram-Schmidt Orthogonalisation Process,
37
H olders Inequality, 10, 12
H olders Inequality., 11
Hadas, Moses, ii
Hahn-Jordan Decomposition Theorem, 15
half-space
closed, 83
open, 83
Hamel basis, 36, 74
Hausdor topology, 128
Hilbert space, 3, 31, 64
Hilbert space adjoint, 117
Hilbert space dimension, 39
hyperplane, 73
ane, 83
idempotent, 34, 115
Inequality
Bessels, 37
Cauchy-Schwarz, 31
inner product spaces, 31
integral
Riemann-Stieltjes, 25
integral operator, 26
Inverse Mapping Theorem, 112
invertible
linear operator, 42
involution, 123
isomorphism of Hilbert spaces, 39
iterated dual spaces, 22
James space, 14
kernel
of an integral operator, 26
Krein-Milman Theorem, 109
Kronecker delta function, 40, 118
Laurel, S., 102
LCS, 58
Lebesgue Dominated Convergence Theorem,
96
Lec, Stanislaw J., ii
Levenson, Sam, 104
linear functional, 70
locally compact, 104
locally convex space, 58, 61
manifold
ane, 83
Martin, Steve, 30, 111
Marx, Groucho, ii, 55, 127
measurable partition, 15
measure
complex-valued, Borel, 15
nite, 15
regular, 15
metric
complete, 68
translation invariant, 68
Milligan, Spike, 16
Minkowski functional, 14, 56, 61, 82
Minkowskis Inequality, 10, 11
multiplication operator, 19
INDEX 137
nbhd, 128
neighbourhood, 128
balanced, 45
neighbourhood base, 129
neighbourhood system, 128
net, 130
Cauchy, 47
convergence of, 130
Newhart, Bob, 1
norm, 1
Euclidean, 45
operator, 17, 21
total variation, 75
uniform, 3
norm topology, 1
normal topological space, 128
normed linear space, 1
norms
equivalent, 4
Novak, Ralph, ii
open map, 9, 48
Open Mapping Theorem, 111
open set, 127
operator
bilateral weighted shift, 20
bounded, 17
compact, 116, 119
diagonal, 20
dierentiation, 21
multiplication, 19
positive, 104
Volterra, 42
weighted shift, 20
operator norm, 17, 21
operators
unitary, 39
orthogonal, 31
orthogonal complement, 34
orthogonal projection, 34
orthonormal basis, 36
orthonormal set, 36
Parallelogram Law, 32, 43
Parcevals Identity, 38
Parker, Dorothy, ii
partition
measurable, 15
Philips, Emo, 70
Phillips Theorem, 34
positive linear functional, 104
positive operator, 104
pre-annihilator, 100
product topology, 132
projection, 115
pseudo-metric
complete family of, 68
pure states, 104
Pythagorean Theorem, 32
quasinilpotent, 42
quotient topology, 48
reexive Banach space, 25
regular measure, 15
regular topological space, 128
Riemann-Stieltjes integral, 25
Riesz Representation Theorem, 35, 95
Riesz-Markov Theorem, 95
Rudner, Rita, 44, 116
Schatten p-class, 28
Schauder basis, 22, 36, 40, 116
standard, 23
seminorm, 1, 55, 57
seminorms
separating family of, 59
separated, 83, 128
strictly, 83
separating
family of functions, 63
separating family, 89
separating family of seminorms, 59
separation axioms, 128
series
absolutely summable, 6
Shandling, Garry, 43
space
Banach, 3
Hilbert, 3, 64
James, 14
Tsirelson, 13
spectral radius, 42
spectral radius formula
Beurlings, 42
spectrum, 42
standard Schauder basis, 23
state, 104
state space, 104
states
pure, 104
strictly separated, 83
strong operator topology, 64
subbase for a topology, 129
sublinear functional, 56
subnet, 130
138 INDEX
subspace
ane, 83
summable
unconditionally, 37
Theismann, J., 115
Theorem
Baire Category, 93, 111
Banach-Alaoglu, 97
Banach-Steinhaus, 94
Closed Graph, 112
F. John, 43
Goldstine, 98
Hahn-Jordan Decomposition, 15
Inverse Mapping, 112
Krein-Milman, 109
Lebesgue Dominated Convergence, 96
Open Mapping, 111
Phillips, 34
Pythagorean, 32
Reisz-Markov, 95
Riesz Representation, 95
Riesz Representation (for Hilbert spaces),
35
Tychono, 96
Whitleys proof of Phillips, 34
topological dual, 70
topological space
T0, 128
T1, 128
T3, 128
T4, 128
Hausdor, 128
normal, 128
regular, 128
separation axioms, 128
topological vector space, 44
topologically complemented, 34, 79, 113
topology, 127
base for a, 129
co-nite, 128
compatible, 44
discrete, 127
generated by a subbase, 129
norm, 1
product, 132
quotient, 48
strong operator, 64
subbase for a, 129
trivial, 127
weak, 90, 131, 132
weak operator, 65
weak
, 91
total variation, 13
total variation norm, 75
trace functional, 104
translation invariant metric, 68
trivial topology, 127
Tsirelson space, 13, 14
Tsirelson, B.S., 13
TVS, 44
Twain, Mark, 89
Tychonos Theorem, 96
unconditionally summable, 37
uniform boundedness principle, 92, 93
uniform continuity on a subset, 51
uniform norm, 3
uniform space, 69
uniformity, 50
uniformly continuous function, 50
unilateral backward weighted shift operator,
20
unilateral forward weighted shift operator,
20
unitary operators, 39
variation, 15, 25
bounded, 25
total, 13
Volterra operator, 26, 42
weak convergence, 90
weak operator topology, 65
weak topology, 89, 131
weak topology (induced by the continuous
dual), 90
weak topology induced by a family of
functions, 131
weak
-topology, 91
weighted
2
-space, 32
weighted shift operator, 20
Willard, Stephen, 127
Wright, Steven, 17
Youngman, Henny, 54

An Introduction To Functional Analysis

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

An Introduction To Functional Analysis

Uploaded by

Copyright:

Available Formats

An Introduction to Functional Analysis

) is a normed linear space. Again, we abbreviate this to

= max[f(z)[ : z [0, 1].

-direct sum of the

are equivalent norms on X. As we shall later see,

are inequivalent norms on

(x)) W. That is, V

(x) > 0 for x (0, 1) and f

(x) < 0 for x (1, ),

1; i.e. B is contained in the unit ball of c

I have the body of an eighteen year old. I keep it in the fridge.

= sup[f(x)g(x)[ : x [0, 1] |f|

(X, d) and let

1/n. Since n 1 was arbitrary,

(N). As we are wont to

(N). It follows that M

(N) and we dene the linear map V : H H via

(Z), dene the linear map

). Dene the map

= 1 for each n 1 and Dp

are referred to as continuous linear

is again a Banach space.

, n 3. All of these are Banach spaces.

= 1 for all n 1, and

= 1. Since, for all n 1,

, where as always q is the

) = f(0) and f(1

T is a hermitian matrix, and as such, has positive eigenvalues. Denote by

. For each n 1, let z

is a norm-closed subspace of H. In particular,

is not topologically complemented in

. This result is known as

, so that if x H, then we may

in a unique way. Consider

spano. In fact, if we let

. Suppose that there exists 0 ,= x

, x , /. Then x H, and so we can write x = m

. Consider the vector w =

, and thus w = Px. That is, Px =

. Then there exists a unique vector y H so

. If = 0, then = (0). Otherwise, let / = ker , so that

, by Remark 3.7 (d), for all x H, we have

of a Hilbert space H is said to be or-

| = 1 for all , and ,= implies that e

is any orthonormal set in H, then an easy application of Zorns

of vectors in X, let T denote the

, and we say that x

= y if for all > 0 there exists F

, |z| = 1. But then c z is an orthonormal set, contradicting

, which exists by Lemma 3.16. Then

) = 0 for all , so y x is orthogonal to / = spanc = H.

for all , then

be an orthonormal basis for H

is an orthonormal basis for H

| = 1 for all , and

f()g(). The proof that

) = (x, y). Then there exists

y| < /2. But then

y| < . In other words, (x

) tends to (x, y) and so is

) = (k, x). But then we can nd

[ < [k[+1. Next we can nd

x| < /2([k[ + 1). Choosing