Linear Analysis (Fall 2001)

Volker Runde
August 22, 2003
Contents
Introduction 3
1 Basic concepts 4
1.1 Normed spaces and Banach spaces . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Finite-dimensional spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.3 Linear operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.4 The dual space of a normed space . . . . . . . . . . . . . . . . . . . . . . 23
2 The fundamental principles of functional analysis 27
2.1 The HahnBanach theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.2 Applications of the HahnBanach theorem . . . . . . . . . . . . . . . . . . 32
2.2.1 The bidual of a normed space . . . . . . . . . . . . . . . . . . . . . 32
2.2.2 Transpose operators . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.2.3 Quotient spaces and duals . . . . . . . . . . . . . . . . . . . . . . . 34
2.2.4 The dual space of (([0, 1]) . . . . . . . . . . . . . . . . . . . . . . . 35
2.2.5 Runges approximation theorem . . . . . . . . . . . . . . . . . . . 38
2.3 Baires theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.4 The uniform boundedness principle . . . . . . . . . . . . . . . . . . . . . . 44
2.5 The open mapping theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.6 The closed graph theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3 Spectral theory of bounded linear operators 53
3.1 The spectrum of a bounded linear operator . . . . . . . . . . . . . . . . . 53
3.2 Spectral theory for compact operators . . . . . . . . . . . . . . . . . . . . 58
3.3 Hilbert spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.3.1 Inner products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.3.2 Orthogonality and self-duality . . . . . . . . . . . . . . . . . . . . . 65
3.3.3 Orthonormal bases . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.3.4 Operators on Hilbert spaces . . . . . . . . . . . . . . . . . . . . . . 75
3.4 The spectral theorem for compact, self-adjoint operators . . . . . . . . . . 79
1
4 Fixed point theorems and locally convex spaces 82
4.1 Banachs xed point theorem . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.2 Locally convex vector spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.2.1 Weak and weak

topologies . . . . . . . . . . . . . . . . . . . . . . 87
4.3 Schauders xed point theorem . . . . . . . . . . . . . . . . . . . . . . . . 90
4.3.1 Peanos theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.3.2 Lomonosovs theorem . . . . . . . . . . . . . . . . . . . . . . . . . 96
4.4 The MarkoKakutani xed point theorem . . . . . . . . . . . . . . . . . 97
4.5 A geometric consequence of the HahnBanach theorem . . . . . . . . . . . 100
4.6 The KrenMilman theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 102
4.6.1 The StoneWeierstra theorem . . . . . . . . . . . . . . . . . . . . 104
A Point set topology 106
A.1 Open and closed sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
A.2 Continuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
A.3 (Local) compactness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
A.4 Convergence of nets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
A.5 Closedness and continuity via nets . . . . . . . . . . . . . . . . . . . . . . 113
A.6 Compactness via nets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
A.7 Tychonos theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
B Measure and integration 119
B.1 Measure spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
B.2 Denition of the integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
B.3 Theorems about the integral . . . . . . . . . . . . . . . . . . . . . . . . . . 123
2
Introduction
These are the T
E
Xed and polished notes of the course Math 516 (Linear Analysis
1
) as I
taught it in the fall terms 2000 and 2001. The most distinctive feature of these notes is
their complete lack of originality: Everything can be found in one textbook or another.
The book that is probably closest in spirit is
1. J. B. Conway, A Course in Functional Analysis. Springer Verlag, 1985.
Other recommended books are:
2. B. Bollob as, Linear Analysis. An Introductory Course, Second Edition. Cam-
bridge University Press, 1999.
3. N. Dunford and J. T. Schwartz, Linear Operators, I. Wiley-Interscience, 1988.
4. G. K. Pedersen, Analysis Now. Springer Verlag, 1989.
5. W. Rudin, Functional Analysis, Second Edition. McGraw-Hill, 1991.
The notes, which are generally kept in a rather brutal theorem-proof style, are not intended
to replace any of these books, but rather to supplement them by relieving the students
from the necessity of taking notes and thus allowing them to devote their full attention
to the lecture.
Volker Runde, Edmonton August 22, 2003
1
called baby functional analysis by some
3
Chapter 1
Basic concepts
In this chapter, we introduce the main objects of study in this course:
normed linear spaces, in particular Banach spaces, and
the bounded linear maps between them.
1.1 Normed spaces and Banach spaces
All linear spaces considered in these notes are supposed to be over a eld F, which can be
R or C.
Denition 1.1.1 Let E be a linear space. A norm on E is a map | |: E [0, ) such
that
(a) |x| = 0 x = 0 (x E);
(b) |x| = [[|x| ( F, x E);
(c) |x +y| |x| +|y| (x, y E).
A linear space equipped with a norm is called a normed space.
Examples 1. There are several canonical norms on the linear space E := F
N
. For
x = (
1
, . . . ,
N
), let:
|x|
1
:=
N

j=1
[
j
[;
|x|
2
:=
_
_
N

j=1
[
j
[
2
_
_
1
2
;
|x|

:= max[
1
[, . . . , [
N
[.
4
Then | |
1
, | |
2
, and | |

are norms on E which satisfy the inequality

|x|

|x|
1

N|x|
2
N|x|

(x E). (1.1)
2. Let S ,= be a set and dene

(S, F) :=
_
f : S F : sup
sS
[f(s)[ <
_
.
For f

(S, F), dene

|f|

:= sup
sS
[f(s)[.
This turns

(S, F) into a normed space.

3. In fact, there is a norm on any linear space E. Let S be a Hamel basis (see Exercise
1.2 below) for E. Let x E. Then there are (necessarily unique)
1
, . . . ,
n
F
along with s
1
, . . . , s
n
S such that x =

n
j=1

j
s
j
. Dene
|x| :=
n

j=1
[
j
[.
This denes a norm on E.
The last example emphasizes that a normed space is not just a linear space that can
be equipped with a norm, but a linear space equipped with a particular norm.
Exercise 1.1 Justify (1.1).
Exercise 1.2 Let E ,= 0 be a (possibly innite-dimensional) linear space. A Hamel basis for E
is a set S of elements of E with the following properties:
S is linearly independent.
Each element of E is a linear combination of elements of S.
Show that E has a Hamel basis by proceeding as follows:
(i) Let o := T E : T is linearly independent. Show that o , = .
(ii) Let T be a non-empty subset of o which is totally ordered by set inclusion. Show that

T : T T belongs again to o.
(iii) Use Zorns lemma to conclude that o has maximal elements.
(iv) Show that any such maximal element is a Hamel basis for E.
Exercise 1.3 Let p, q (1, ) be such that
1
p
+
1
q
= 1.
5
(i) Show that
x
1
p
y
1
q

x
p
+
y
q
(x, y > 0). (1.2)
(Hint: Apply the logarithm to (1.2) and prove that inequality rst.)
For x = (
1
, . . . ,
N
) F
N
, let
|x|
p
:=
_
_
N

j=1
[
j
[
p
_
_
1
p
.
(ii) Holders inequality. Show that, for x = (
1
, . . . ,
N
), y = (
1
, . . . ,
N
) F
N
, we have
N

j=1
[
j

j
[ |x|
p
|y|
q
.
Which known inequality do you obtain for p = q = 2?
(iii) Minkowskis inequality. Show that
|x +y|
p
|x|
p
+|y|
p
(x, y F
N
).
(iv) Conclude that | |
p
is a norm on F
N
.
Exercise 1.4 Let F be a linear subspace of a normed space E. Show that the closure of F in E
is also a linear subspace of E.
Exercise 1.5 A seminorm on a linear space E is a map p : E [0, ) with the following
properties:
p(x) = [[p(x) ( F, x E);
p(x +y) p(x) +p(y) (x, y E).
What is missing from the denition of a norm?
(i) Show that F := x E : p(x) = 0 is a linear subspace of E.
(ii) For x E, dene |x +F| := p(x). Show that |[ |[ is a norm on E/F.
If (E, | |) is a normed space, then
d: E E [0, ), (x, y) |x y|
is a metric. We may thus speak of convergence, etc., in normed spaces.
Denition 1.1.2 A normed space E is called a Banach space if the corresponding metric
space is complete, i.e. every Cauchy sequence in E converges in E.
Exercise 1.6 Let E be a normed space, and let F be a linear subspace of E. Show that, if F is
a Banach space, then it is closed in F. Conversely, show that, if E is a Banach space and if F is
closed in F, then F is a Banach space.
6
Examples 1. (F
N
, | |
j
) is a Banach space for j = 1, 2, .
2. Let (f
n
)

n=1
be a Cauchy sequence in (

(S, F), | |

). For each s S, we have

[f
n
(s) f
m
(s)[ |f
n
f
m
|

(n, m N).
Hence, (f
n
(s))

n=1
is a Cauchy sequence in F. Dene f : S F by letting
f(s) := lim
n
f
n
(s).
We claim that f

(S, F) and that |f

n
f|

0. We prove both claims

simultaneously. Let > 0, and choose N N such that
|f
n
f
m
| < (n, m N).
It follows that
[f
n
(s) f
m
(s)[ |f
n
f
m
| < (n, m N, s S).
We thus obtain for n N and s S:
[f
n
(s) f(s)[ = lim
m
[f
n
(s) f
m
(s)[
limsup
m
|f
n
f
m
|
. (1.3)
In particular, we obtain
[f(s)[ [f
N
(s)[ + |f
N
|

+,
so that f

(S, F). Taking the supremum over s S in (1.3) yields

|f
n
f|

3. Let X be a topological space, and dene
(
b
(X, F) := f

(X, F) : f is continuous.
By Theorem A.2.4, (
b
(X, F) is a closed subspace of

(X, F) and thus a Banach

space by Exercise 1.4.
4. Let X be a locally compact Hausdor space, and let (
0
(X, F) be dened as in
Denition A.3.13. Let (f
n
)

n=1
be a sequence in (
0
(X, F) converging to f (
b
(X, F)
(with respect to | |

). Let > 0, and choose N N such that

|f
n
f|

<

2
(n N).
7
It follows that
[f
N
(x) f(x)[ |f
N
f|

<

2
and consequently
[f(x)[ [f
N
(x)[ +

2
.
Since f
N
(
0
(X, F), there is a compact set K X such that sup
xX\K
[f(x)[ <

2
.
This implies
sup
xX\K
[f(x)[ < .
Hence, (
0
(X, F) is a closed subspace of the Banach space (
b
(X, F) and thus a Banach
space itself.
5. For N N, dene
(
N
([0, 1], F) := f : [0, 1] F : f is N-times continuously dierentiable.
Dene
f : [0, 1] F, x
_
1
2
x, x
_
0,
1
2

,
x
1
2
, x
_
1
2
, 1

.
Then f belongs to (([0, 1])
1
, but not to (
N
([0, 1]). Since f is piecewise continuously
dierentiable, it can be uniformly approximated by the sequence (s
n
)

n=1
of the
partial sums of its Fourier series. Since s
n
(
N
([0, 1]) for each n N, we obtain
that (
N
([0, 1]) is not closed in ((([0, 1]), | |

). Hence, ((
N
([0, 1]), | |

) is not a
Banach space.
Dene another norm on (
N
([0, 1]):
|f|
N
:=
N

j=1
|f
(j)
|

(f (
N
([0, 1]).
Let (f
n
)

n=1
be a Cauchy sequence in ((
N
([0, 1]), | |
N
). Then (f
(j)
)

n=1
is a Cauchy
sequence in ((([0, 1]), | |

) for j = 0, 1, . . . , N. Since (([0, 1]) is a Banach space,

there are g
0
, g
1
, . . . , g
N
(([0, 1]) such that
|f
(j)
n
g
j
|

0 (j = 0, 1, . . . , N).
We claim that
g
j+1
= g
t
j
(j = 0, . . . , N 1).
1
If the eld F is obvious or irrelevant, we often write C
0
(x), C
N
([0, 1]), etc., without the symbol F.
8
This, however, follows immediately from
g
j
(x) = lim
n
f
(j)
n
(x)
= lim
n
__
x
0
f
(j+1)
(t) dt +f
(j)
n
(0)
_
=
_
x
0
g
j+1
(t) dt +g
j
(0) (x [0, 1], j = 0, . . . , N 1).
Let f := g
0
. Then we obtain inductively that f
(j)
= g
j
for j = 0, 1, . . . , N. This
yields
|f
n
f|
N
=
N

j=0
|f
(j)
n
f
(j)
|

=
N

j=0
|f
(j)
n
g
j
|

0.
Hence, ((
N
([0, 1]), | |
N
) is a Banach space.
The last example shows that a linear space equipped with one norm may fail to be a
Banach space, but can be a Banach space with respect to another norm.
Exercise 1.7 We write shorthand

(F) for

(N, F). Let c(F) denote the subspace of

(F)
consisting of all convergent sequences in F, let c
0
(F) be the subspace of all sequences in F that
converge to zero, and let c
00
(F) consist of all sequences (
n
)

n=1
in F such that
n
= 0 for all but
nitely many n N.
(i) Show that (c(F), | |

) and (c
0
(F), | |

) are Banach spaces.

(ii) Show that (c
00
(F), | |

) is not a Banach space.

Our rst theorem is an often useful characterization of completeness in normed spaces
in terms of convergent series:
Denition 1.1.3 Let (x
n
)

n=1
be a sequence in the normed space E. We say that the
series

n=1
x
n
converges in E if the sequence
_

N
n=1
x
n
_

N=1
of its partial sums converges.
We say that

n=1
x
n
converges absolutely if

n=1
|x
n
| < .
Exercise 1.8 For n N, let
x
n
: N F, m
_
_
_
1
n
, m = n,
0, else.
(i) Show that, for every permutation : N N, the series

n=1
x
(n)
converges in c
0
to the
same limit. Which is it?
(ii) Show that the series

n=1
x
n
is not absolutely convergent.
Theorem 1.1.4 (RieszFischer theorem) Let E be a normed space. Then the follow-
ing are equivalent:
9
(i) E is a Banach space.
(ii) Every absolutely converging series in E converges.
Proof (i) = (ii): This is proven in the same fashion as for series in R.
(ii) = (i): Let (x
n
)

n=1
be a Cauchy sequence. Choose a subsequence (x
n
k
)

k=1
of
(x
n
)

n=1
such that
|x
n
k
x
n
k+1
| <
1
2
k
(k N).
Let
y
k
:= x
n
k
x
n
k+1
(k N).
Then the series

k=1
y
k
converges absolutely and thus converges in E. Since
K

k=1
y
k
= (x
n
1
x
n
2
) + (x
n
2
x
n
3
) + + (x
n
K
x
n
K+1
) = x
n
1
x
n
K+1
(K N),
it follows that (x
n
k
)

k=1
is also convergent, with limit x, say. Let > 0, and choose
K, N N with n
K
N
|x
n
x
m
| <

2
(n, m N) and |x
n
k
x| <

2
(k K).
For n maxN, K, this means
|x
n
x| |x
n
x
n
k
| +|x
n
k
x| < ,
so that x = lim
n
x
n
. .
Example Let (, S, ) be a measure space, let p [1, ), and let
|f|
p
:=
__

[f()[
p
d()
_1
p
for each measurable function f : F. Dene
L
p
(, S, ) := f : F : f is measurable with |f|
p
< .
Using Holders and Minkowskis inequalities (compare Exercise 1.3), it can be shown that
L
p
(, S, ) is a linear space, and | |
p
is a seminorm on it. Let
^
p
:= f L
p
(, S, ) : |f|
p
= 0,
and dene
L
p
(, S, ) := L
p
(, S, )/^
p
.
10
By Exercise 1.5, | |
p
induces a norm on L
p
(, S, ), which we denote by | |
p
as well.
We claim that (L
p
(, S, ), | |
p
) is a Banach space.
We will use Theorem 1.1.4. Let (f
n
)

n=1
be a sequence in L
p
(, S, ) such that

n=1
|f
n
|
p
< . Dene
g : [0, ],
_

n=1
[f
n
()[
_
p
.
We claim that g is integrable. To see this, note that
_
_

_
N

n=1
[f
n
()[
_
p
d()
_
1
p
=
_
_
_
_
_
N

n=1
[f
n
[
_
_
_
_
_
p

n=1
|f
n
|
p
(N N).
By the monotone convergence theorem (Theorem B.3.1), this means
_

g() d() = lim

N
_

_
N

n=1
[f
n
()[
_
p
d()
_

n=1
|f
n
|
p
_
p
< ,
so that g is indeed integrable. Hence, for almost all , the series

n=1
f() converges
absolutely. Dene
f : F,
_

n=1
f
n
(), if g() < ,
0, otherwise.
Then f is measurable with [f[
p
g. For almost all , we have
lim
N

n=1
f
n
() f()

= 0 and

n=1
f
n
() f()

p
g().
For the dominated convergence theorem (Theorem B.3.2) it is then easily inferred that
lim
N
_
_
_

N
n=1
f
n
f
_
_
_
p
= 0, i.e. f =

n=1
f
n
.
The following exercise is the measure theory free version of the preceding example:
Exercise 1.9 Let p [1, ). Let
p
be the set of all sequences (x
n
)

n=1
in F with

n=1
[x
n
[
p
< .
(i) For x = (x
n
)

n=1

p
dene
|x|
p
:=
_

n=1
[x
n
[
p
_1
p
.
Show that | |
p
is a norm on
p
.
(ii) Show that (
p
, | |
p
) is a Banach space.
11
Example Let (, S, ) be a -nite measure space. A measurable function f : F is
called essentially bounded if there is C 0 such
: [f()[ C (1.4)
is a -zero set. Let L

(, S, ) denote the set of all essentially bounded functions on .

For f L

(, S, ) dene
|f|

:= infC 0 : (1.4) is a -zero set.

It is easy to see that | |

is a seminorm on L

(, S, ). Let ^

:= f L

(, S, ) :
|f|

= 0, and dene
L

(, S, ) := L

(, S, )/^

.
Then L

(, S, ) equipped with the norm induced by | |

likewise denoted by | |

is a Banach space.
1.2 Finite-dimensional spaces
We have seen in the previous section, that there may be dierent norms on one linear
space: ((
N
([0, 1]), | |

) is not a Banach space whereas ((

N
([0, 1]), | |
1
) isnt. On the
other hand, the norms | |
1
, | |
2
, and | |

on F
N
are related by (1.1), so that the
resulting topologies are identical. The following denition captures this phenomenon:
Denition 1.2.1 Let E be a linear space. Two norms | |
1
and | |
2
are called equivalent
(in symbols: | |
1
| |
2
) if there is C 0 such that
|x|
1
C|x|
2
and |x|
2
C|x|
1
(x E).
Exercise 1.10 Verify that the equivalence of norms is indeed an equivalence relation, i.e. it is
reexive, symmetric, and transitive.
Exercise 1.11 Let E be a linear space, and let | |
1
and | |
2
be two equivalent norms on E.
Verify in detail that (E, | |
1
) is a Banach space if and only if (E, | |
2
) is a Banach space.
Examples 1. | |
1
, | |
2
, and | |

on F
N
are equivalent by (1.1).
2. for x = (
n
)

n=1
c
00
(F) we have
|x|

n=1
[
n
[ =: |x|
1
.
On the other hand, let
x
n
:= (1, . . . , 1
. .
n-times
, 0, . . . ).
12
Then
|x
n
|

= 1, but |x
n
|
1
= n (n N),
so that | |

, | |
1
.
3. | |

and | |
N
are not equivalent on (
N
([0, 1]). (Why?)
The next theorem shows that there is only one equivalence class of norms on a nite-
dimensional vector space.
Theorem 1.2.2 Let E be a nite-dimensional linear space. Then all norms on E are
equivalent.
Proof Let x
1
, . . . , x
N
E be a basis for E. For x =
1
x
1
+ +
N
x
N
, let
|[x|[ := max[
1
[, . . . , [
N
[.
It is sucient to show that |[ |[ | | for every other norm on E.
Let x E. Then we have:
|x| [
1
[|x
1
| + +[
N
[|x
N
|
|[x|[(|x
1
| + +|x
N
|)
. .
=:C
1
.
It remains to be shown that there is C
2
0 with |[x|[ C
2
|x| for all x E.
Assume otherwise. Then there is a sequence (x
(n)
)

n=1
in E with |[x
(n)
|[ > n|x
(n)
|.
Let
y
(n)
:=
x
(n)
|[x
(n)
|[
(n N).
For each n N, there are unique
(n)
1
, . . . ,
(n)
N
F with y
(n)
=

N
j=1

(n)
j
x
j
. It follows
that
_
_
_
_

(n)
1
, . . . ,
(n)
N
__
_
_

= |[y
(n)
|[ = 1 (n N).
By the HeineBorel theorem, the sequence
__

(n)
1
, . . . ,
(n)
N
__

n=1
has a convergent sub-
sequence
__

(n
k
)
1
, . . . ,
(n
k
)
N
__

k=1
. Let
j
:= lim
k

(n
k
)
j
for j = 1, . . . , n, and dene
y :=

N
j=1

j
x
j
. It follows that
|[y|[ = |(
1
, . . . ,
N
)|

= 1,
so that, in particular, y ,= 0, and
|y
(n
k
)
y| C
1
|[y
(n
k
)
y|[ = C
1
_
_
_
_

(n
k
)
1

1
, . . . ,
(n
k
)
N

N
__
_
_

0.
On the other hand, the choice of (x
(n)
)

n=1
implies n|y
(n)
| < 1 so that lim
n
|y
(n)
| = 0.
But this means that y = 0, which is impossible. .
13
Theorem 1.2.2 does not mean that nite-dimensional normed spaces are uninteresting:
It says nothing about the constant C showing up in Denition 1.2.1. To nd optimal values
for C for concrete norms can be quite challenging. We wont pursue this, however.
Corollary 1.2.3 Every nite-dimensional normed space is a Banach space.
Exercise 1.12 Prove Corollary 1.2.3.
Corollary 1.2.4 Every nite-dimensional subspace of a normed space is closed.
As an immediate consequence, each nite-dimensional normed space can be identi-
ed with F
N
, so that theorems for F
N
carry over to arbitrary nite-dimensional normed
spaces. In particular, a nite-dimensional space has the BolzanoWeierstra property:
Each bounded sequence has a convergent subsequence. As we shall now see, this property
even characterizes the nite-dimensional normed spaces.
Lemma 1.2.5 (Riesz lemma) Let E be a normed space, and let F be a closed, proper,
i.e. F ,= E, subspace of E. Then, for each (0, 1), there is x

E with |x

| = 1, and
|x x

| for all x F.
Proof Let x EF, and let := inf|xy| : y F. Then there is a sequence (x
n
)

n=1
in F with lim
n
|x x
n
| = . If = 0, the closedness of F implies x F, which is a
contradiction. Hence, > 0 must hold. Since (0, 1), we have <

. Choose y F
with 0 < |x y| <

. Let
x

:=
y x
|y x|
,
so that trivially |x

| = 1. For any z F, we then have:

|z x

| =
_
_
_
_
z
y x
|y x|
_
_
_
_
=
_
_
_
_
z
y
|y x|
+
x
|y x|
_
_
_
_
=
1
|x y|
|(|x y|z +y
. .
F
) x|
. .

= .
This completes the proof. .
Theorem 1.2.6 For a normed space E, the following are equivalent:
14
(i) Every bounded sequence in E has a convergent subsequence.
(ii) dimE < .
Proof (ii) = (i) is elementary.
(i) = (ii): Suppose that dimE = . Choose x
1
E with |x
1
| = 1. Suppose that
x
1
, . . . , x
n
have already been chosen such that
|x
j
| = 1 (j = 1, . . . , n)
and
|x
j
x
k
|
1
2
(j, k = 1, . . . , nj ,= k).
Let F := linx
1
, . . . , x
n
. Since dimF < , F is a proper and automatically closed
subspace of E. By Riesz lemma, there is x
n+1
E with
|x
n+1
| = 1 and |x x
n+1
|
1
2
(x F).
Inductively, we thus obtain a sequence (x
n
)

n=1
of unit vectors such that
|x
n
x
m
|
1
2
(n ,= m). (1.5)
By (1.5), (x
n
)

n=1
has no Cauchy subsequence. .
Theorem 1.2.6 is the rst example for the many subtle and often surprising links
between algebra and analysis that surface in this course: A purely algebraic property
a linear space has nite dimension turns out to be equivalent to the purely analytic
BolzanoWeierstra property.
1.3 Linear operators
One of the major topics in linear algebra is the study of linear maps between nite-
dimensional linear spaces. A considerable part of this course will be devoted to the study
of linear maps between (possibly, but not necessarily) innite-dimensional spaces.
Denition 1.3.1 Let E and F be linear spaces. A map T : E F is called linear if
T(x +y) = Tx +Ty (x, y E, , F).
Linear maps are also called linear operators. A linear operator from E to F is called a
linear functional .
15
Examples 1. Let E := F
N
, let F := F
M
, and let A be an M N-matrix. Then
T
A
: E F, x Ax
is a linear operator.
2. Let ,= R
N
be open, and let (
M
() denote the linear space of all functions f :
F for which all partial derivatives of order at most M exist and are continuous.
For each multiindex N
N
0
with [[ :=
1
+ +
N
M, let f

((). Then
D: (
M
() ((), f

[[M
f

f
x

is linear. Operators of this type are called linear (partial) dierential operators.
3. Let k : [0, 1] [0, 1] F be continuous. For f (([0, 1]), dene Tf : [0, 1] F
through
(Tf)(x) :=
_
1
0
f(y)k(x, y) dy (x [0, 1]).
We claim that Tf (([0, 1]). Fix x
0
[0, 1], and let > 0. Since [0, 1] [0, 1] is
compact, k is uniformly continuous. Hence, there is > 0 such that
[k(x, y) k(x
t
, y
t
)[ <

|f|

+ 1
for all (x, y), (x
t
, y
t
) [0, 1] [0, 1] with |(x, y) (x
t
, y
t
)|
2
< . Let x [0, 1] such
that [x x
0
[ < . It follows that
|(x, y) (x
0
, y)|
2
= [x x
0
[ < (y [0, 1]).
We thus obtain:
[(Tf)(x) (Tf)(x
0
)[ =

_
1
0
f(y)[k(x, y) k(x
0
, y)] dy

_
1
0
[f(y)[
. .
|f|

[k(x, y) k(x
0
, y)]
. .
<

f

+1
dy
.
Hence, Tf is continuous.
It is immediately checked that
T : (([0, 1]) (([0, 1]), f Tf
is a linear operator, the Fredholm operator with kernel k. Fredholm operators are
part of the larger class of linear integral operators.
16
The rst examples suggests that we may view linear operators as generalizations of
matrices.
Exercise 1.13 Let E be a linear space with Hamel basis s, let F be another linear space, and
let (y

)
sS
be an arbitrary family of elements of F. Show that there is a unique linear operator
T : E F such that Ts = y
s
for all s S.
There is virtually nothing of substance that can be said on linear operators between
arbitrary linear spaces. We have to conne ourselves to the setting of normed spaces
preferably Banach spaces and continuous linear operators.
Theorem 1.3.2 Let E and F be normed spaces. Then the following are equivalent for a
linear operator T : E F:
(i) T is continuous at 0.
(ii) T is continuous.
(iii) There is C 0 such that |Tx| C|x| for all x E.
(iv) sup|Tx| : x E, |x| 1 < .
Operators satisfying these equivalent conditions are called bounded.
Proof (i) = (ii): Let x E, and let (x
n
)

n=1
be a sequence in E such that x
n
x.
Then
|Tx
n
Tx| = |T(x
n
x
. .
0
)| 0
holds, which proves (ii).
(ii) = (iii): Assume that (ii) holds, but that (iii) is false. Then there is a sequence
(x
n
)

n=1
in E such that |Tx
n
| > n|x
n
| for all n N. Let
y
n
:=
x
n
|Tx
n
|
(n N).
Since 1 > n|y
n
| for all n N, it follows that y
n
0. On the other hand, we have
|Tx
n
| = 1, which is impossible if T is continuous at 0.
(iii) = (iv): Clearly, (iii) implies
sup|Tx| : x E, |x| 1 C.
(iv) = (i): Assume that T is not continuous at 0. Then there is a sequence (x
n
)

n=1
in E such that x
n
0, but := inf |Tx
n
| > 0. Let
y
n
:=
x
n
|x
n
|
(n N),
17
so that |y
n
| = 1 for all n N. On the other hand, we have
|Ty
n
| =
1
|x
n
|
|Tx
n
| >

|x
n
|
,
which contradicts (iv). .
Exercise 1.14 Show that the following are equivalent for a normed space E:
(a) dimE = .
(b) For each normed space F ,= 0, there is an unbounded linear operator T : E F.
(c) There is an unbounded linear functional on E.
Exercise 1.15 Let E be a normed space. Show that the following are equivalent for a linear
functional : E F:
(a) / E

.
(b) (x E : |x| 1) = F.
(c) ker =
1
(0) is dense in E.
Exercise 1.16 Let E and F be Banach spaces, and let T B(E, F) be such that there is C 0
with
|x| C|Tx| (x E).
Show that T is injective and has closed range.
Examples 1. Let E and F be normed spaces, and let T : E F be linear. Suppose
that dimE < . Dene
|[x|[ := max|x|, |Tx| (x E).
Then |[ |[ is a norm on E. By Theorem 1.2.2, |[ |[ and | | are equivalent, so
that there is C 0 with
|Tx| |[x|[ C|x| (x E).
Hence, T is bounded.
2. Let
T : (
1
([0, 1]) (([0, 1]), f f
t
,
and let both (
1
([0, 1]) and (([0, 1]) be equipped with | |

. For n N, dene
f
n
(x) := x
n
(x [0, 1]),
18
so that |f
n
|

= 1 for n N. However, since

f
t
n
(x) =
nx
n1
n
(n N, x [0, 1]),
we have |Tf
n
|

= n, so that T is not bounded. If, however, (

1
([0, 1]) is equipped
with the (
1
-norm | |
1
, T becomes bounded:
|Tf|

= |f
t
|

|f|

+|f
t
|

= |f|
1
(f (
1
([0, 1])).
3. Let k: [0, 1] [0, 1] F be continuous, and let T : (([0, 1]) (([0, 1]) be given by
(Tf)(x) =
_
1
0
f(y)k(x, y) dy (f (([0, 1]), x [0, 1]).
For each f (([0, 1]), we have:
|Tf|

sup
x[0,1]
_
1
0
[f(y)[[k(x, y)[ dy|f|

|k|

.
Hence, T is bounded.
4. Let (, S, ) be a -nite measure space, let p [1, ], and let L

(, S, ).
Dene M

: L
p
(, S, ) L
p
(, S, ) through
M

f := f (f L
p
(, S, )).
It is easy to see that
|M

f|

||

|f|
p
(f L
p
(, S, )).
The rst of these examples shows that the requirement of boundedness is vacuous for
any operator between nite-dimensional spaces.
Given two normed spaces, we shall now see that the collection of all bounded linear
operators between them is again a normed space in a natural manner:
Denition 1.3.3 Let E and F be normed spaces.
(a) The set of all bounded linear operators from E to F is denoted by B(E, F). If
E = F, let B(E, F) =: B(E); if F = F, let B(E, F) =: E

.
(b) For T B(E, F), the operator norm of T is dened as
|T| := sup|Tx| : x E, |x| 1.
Proposition 1.3.4 Let E and F be normed space. Then:
19
(i) B(E, F) equipped with the operator norm is a normed space.
(ii) For T B(E, F), |T| is the smallest number C 0 such that
|Tx| C|x| (x E). (1.6)
(iii) If G is another normed space, then for T B(E, F) and S B(F, G), we have
ST B(E, G) such that
|ST| |S||T|.
Proof (i): It is straightforward to see that B(E, F) is a linear space. We have to check
the norm axioms:
(a) Let T B(E, F) be such that |T| = 0. Let x E 0. Then we have
1
|x|
|Tx| =
_
_
_
_
T
_
x
|x|
__
_
_
_
|T| = 0,
so that Tx = 0. Since x was arbitrary, this means T = 0.
(b) It is routine to see that
|T| = [[|T| ( F, T B(E, F)).
(c) Let x E be with |x| 1. Then we have
|Sx +Tx| |Sx| +|Tx| |S| +|T| (S, T B(E, F))
and thus
|S +T| |S| +|T| (S, T B(E, F)).
(ii): Let x E 0. Since
1
|x|
|Tx| =
_
_
_
_
T
_
x
|x|
__
_
_
_
|T|,
it follows that |Tx| |T||x|. Let C 0 be any other number such that (1.6) holds.
Then
sup|Tx| : x E, |x| 1 C.
(iii): Let x E. Then we have
|STx| |S||Tx| |S||T||x|.
From (ii), it follows that |ST| |S||T|. .
20
Exercise 1.17 Let p [1, ), and let L, R:
p

p
be dened through
L(x
1
, x
2
, x
3
, . . . ) = (0, x
1
, x
2
, . . . )
and R(x
1
, x
2
, x
3
, . . . ) = (x
2
, x
3
, x
4
, . . . ) ((x
1
, x
2
, x
3
, . . . )
p
).
Show that L, R B(
p
) and calculate |L| and |R|.
Exercise 1.18 Let F
N
and F
M
be equipped with | |

, and let A = [a
j,k
] j=1,... ,M
k=1,... ,N
be an MN-
matrix over F. Show that
|T
A
| = max
j=1,... ,M
N

k=1
[a
j,k
[
Theorem 1.3.5 Let E be a normed space, and let F be a Banach space. Then B(E, F)
is a Banach space.
Proof Let (T
n
)

n=1
be a Cauchy sequence in B(E, F). Let x E. Then
|T
n
x T
m
x| |T
n
T
m
||x| (n, m N),
so that (T
n
x)

n=1
is a Cauchy sequence and thus convergent in F. Dene
T : E F, x lim
n
T
n
x.
Clearly, T is linear.
We claim that T B(E, F) and that |T
n
T| 0. Let x E with |x| 1, and let
> 0. There is N N independent of x such that
|T
n
x T
m
x| |T
n
T
m
| < (n, m N).
For n N, this entails that
|T
n
x Tx| = lim
m
|T
n
x T
m
x| limsup
m
|T
n
T
m
| . (1.7)
In particular, we have |Tx| |T
N
|+, so that T B(E, F). Taking the supremum over
x E with |x| 1 in (1.7), we see that |T
n
T| for n N. .
Corollary 1.3.6 For every normed space E, its dual space E

is a Banach space.
As turns out, arbitrary bounded linear operators between Banach spaces are still very
general objects. To obtain stronger results, we have to look at a smaller class of operators:
Denition 1.3.7 Let E and F be normed space. A linear operator T : E F is called
compact if T(x E : |x| 1) is relatively compact in F. The set of all compact
operators from E to F is denoted by /(E, F).
21
Proposition 1.3.8 Let E and F be normed spaces. Then:
(i) /(E, F) is a subspace of B(E, F).
(ii) T B(E, F) is compact if and only if, for each bounded sequence (x
n
)

n=1
in E, the
sequence (Tx
n
)

n=1
has a convergent subsequence.
Proof (i): The set T(x E : |x| 1) is relatively compact and thus bounded in F.
This proves /(E, F) B(E, F). From (ii), it follows easily, that /(E, F) is a subspace of
B(E, F).
(ii): We have:
T is compact T(x E : |x| 1) is relatively compact
rT(x E : |x| 1) is relatively compact for each r > 0
T(x E : |x| r) is relatively compact for each r > 0.
This implies (ii). .
Exercise 1.19 Let E, F, and G, be normed linear spaces, and let T B(E, F) and S B(F, G).
Show that ST /(E, G) if T or S is compact.
Exercise 1.20 Let E be a linear space. A linear operator P : E E is called a projection if
P
2
= P.
(i) Show that a linear map P : E E is a projection if and only if its restriction to PE is the
identity.
(ii) Let E be normed, and let P B(E) be a projection. Show that P has closed range.
(iii) Let E be normed. Show that a projection P B(E) is compact if and only if it has nite
rank.
Exercise 1.21 Is one of the operators L and R from Exercise 1.17 compact?
Examples 1. Let dimE = . Then id
E
: E E is not compact.
2. Let T B(E, F) have nite rank, i.e. dimTE < . Then T is compact. To see this,
let (x
n
)

n=1
be a bounded sequence in E. Then (Tx
n
)

n=1
is a bounded sequence in
TE and thus has a convergent subsequence by Theorem 1.2.2.
3. Let k: [0, 1] [0, 1] F be continuous, and let T B((([0, 1])) be the corresponding
Fredholm operator. Let (f
n
)

n=1
be a bounded sequence, and let C := sup
nN
|f
n
|

.
Let > 0, and choose > 0 such that
[k(x, y) k(x
t
, y
t
)[ <

C + 1
22
whenever |(x, y) (x
t
, y
t
)|
2
< . Let x, x
t
[0, 1] with [x x
t
[ < . It follows that
[(Tf
n
)(x) (Tf
n
)(x
t
)[
_
1
0
[f
n
(y)[
. .
C
[k(x, y) k(x
t
, y
t
)[
. .
<

C+1
dy < .
Hence, (Tf
n
)

n=1
is bounded an equicontinuous. By the Arzel`aAscoli theorem,
(Tf
n
)

n=1
thus has a uniformly convergent subsequence.
Theorem 1.3.9 Let E be a normed space, and let F be a Banach space. Then /(E, F)
is a closed, linear subspace of F.
Proof Let (T
n
)

n=1
be a sequence in /(E, F), and let T B(E, F) be such that |T
n
T|
0. Assume that T / /(E, F). Then there is a bounded sequence (x
n
)

n=1
in E such that
(Tx
n
)

n=1
has no convergent, i.e. Cauchy, subsequence. Passing to a subsequence, we may
thus suppose that
:= inf
n,=m
|Tx
n
Tx
m
| > 0. (1.8)
Let C := sup
nN
|x
n
|, and choose N N so large that |T T
N
| <

3C+1
. For n, m N,
we then have:
|Tx
n
Tx
m
| |Tx
n
T
N
x
n
|
. .
|TT
N
||x
n
|
+|T
N
x
n
T
N
x
m
| +|T
N
x
m
Tx
m
|
. .
|T
N
T||x
m
|
<
2
3
+|T
N
x
n
T
N
x
m
|.
Since T
N
/(E, F), the sequence (T
N
x
n
)

n=1
has a Cauchy subsequence. In particular,
there are n, m N, n ,= m, such that |T
N
x
n
T
N
x
m
| <

3
. It follows that |Tx
n
Tx
m
| <
contradicting (1.8). .
1.4 The dual space of a normed space
We now focus on a particular space of bounded linear operators:
Denition 1.4.1 For a normed space E, the Banach space E

(= B(E, F)) is called the

dual space or dual of E.
We want to give concrete descriptions of some dual spaces.
Denition 1.4.2 Let E and F be normed spaces.
(a) An isomorphism of E and F is a linear map T B(E, F) such that S B(F, E)
exists with ST = id
E
and TS = id
F
. If there is an isomorphism between E and F,
we call E and F isomorphic (in symbols: E

= F).
23
(b) A isometry from E to F is a linear map T : E F such that
|Tx| = |x| (x E).
If there is an isomorphism of E and F which is also an isometry, then E and F are
called isometrically isomorphic (in symbols: E = F).
Exercise 1.22 Let E and F be normed spaces, and let T : E F be an isometry.
(i) Show that T is injective.
(ii) Suppose that T is surjective. Show that there is an isometry S B(F, E) exists with
ST = id
E
and TS = id
F
.
Exercise 1.23 Let c
00
:= c
00
(F) be equipped with the following norms:
|x|

:= sup
nN
[x(n)[ and |x|
1
:=

n=1
[x(n)[ (x c
00
).
Show that the identity map id : (c
00
, | |
1
) (c
00
, | |

) is a continuous bijection, but not an

isomorphism.
Examples 1. Let E and F be normed spaces with dimE = dimF < . Then E

= F.
2. Let p (1, ), and let q (1, ) be such that
1
p
+
1
q
= 1. Dene T :
q
(
p
)

by
letting for x = (x
n
)

n=1

q
and Y = (y
n
)

n=1

p
:
(Tx)(y) :=

n=1
x
n
y
n
.
Since
[(Tx)(y)[ lim
N
N

n=1
[x
n
y
n
[ lim
N
_
N

n=1
[x
n
[
q
_
1
q
_
N

n=1
[y
n
[
p
_
1
p
= |x|
q
|y|
p
,
the map T is well dened and satises |T| 1. Let x
q
with |x|
q
= 1, and
dene (y
n
)

n=1
as follows:
y
n
:=
_
[x
n
[
q
x
n
, if x
n
,= 0,
0, otherwise.
We have

n=1
[y
n
[
p
=

n=1
x
n
=0
[x
n
[
pq
[x
n
[
p
=

n=1
[x
n
[
pqp
= |x|
q
q
= 1,
so that |y|
p
= 1. It follows that
|Tx| [(Tx)(y)[ =

n=1
[x
n
[
q
= |x|
q
q
= 1.
24
For arbitrary x
q
0, we thus have:
|Tx| = |x|
_
_
_
_
T
_
x
|x|
__
_
_
_
|x|.
Hence, T is an isometry.
We claim that T is surjective. Let (
p
)

. For each n N, dene e

(n)

p
through
e
(n)
m
:=
_
1, n = m,
0, n ,= m
Dene x = (x
n
)

n=1
by letting x
n
:= (e
(n)
) for n N. If x
q
, then Tx =
(Why?). For N N, dene x
(N)

q
through
x
(N)
n
:=
_
x
n
, n N,
0, n > N.
For any y = (y
n
)

n=1

p
dene z = (z
n
)

n=1
through
z
n
:=
_
[y
n
x
n
[
x
n
, if n N and x
n
,= 0,
0, otherwise.
If is clear that z
p
with |z|
p
|y|
p
. We now have:
[(Tx
(N)
)(y)[
N

n=1
[x
n
y
n
[
=

n=1
z
n
x
n
=

n=1
z
n
(e
(n)
)
= (z)
= [(z)[
|||z|
p
|||y|
p
.
It follows that
|x|
q
q
= lim
N
N

n=1
[x
n
[
q
= lim
N
|x
(N)
|
q
q
= lim
N
|Tx
(N)
|
q
q
||
q
,
so that x
q
with |x|
q
||.
All in all, we have (
p
)

=
q
.
25
3. Similarly (Exercise 1.24 below), we have (
1
)

and (c
0
)

=
1
.
4. If (, S, ) is any measure space, and if p, q (1, ) are such that
1
p
+
1
q
= 1, we
have an isometric isomorphism T : L
q
(, S, ) L
p
(, S, )

given by
(Tf)(g) :=
_

f()g() d(),
so that L
p
(, S, )

= L
q
(, S, ).
5. For -nite (, S, ), we also have L
1
(, S, )

= L

(, S, ).
Exercise 1.24 Show that (
1
)

and (c
0
)

=
1
.
We now have concrete descriptions of E

for a few normed spaces E. But what can

we say about E

for a general normed space E? So far, the only linear function on E

of which we positively know that its in E

is the zero-functional. Are there any others?

The answer to this question is yes, as we shall see in the next chapter.
26
Chapter 2
The fundamental principles of
functional analysis
In this chapter, we prove four fundamental theorems of functional analysis:
the HahnBanach theorem;
Baires theorem;
the open mapping theorem;
the closed graph theorem.
We illustrate the power of each theorem with application, e.g. to complex variables
and initial value problems.
2.1 The HahnBanach theorem
Given an arbitrary normed space E with dual E

, we cannot tell right now if E

contains
any non-zero elements. This will change in this section: We will prove the HahnBanach
theorem, which implies that there are enough functionals in E

to separate the points of

E.
Roughly speaking, the HahnBanach theorem asserts that, if we have a linear func-
tional on a subspace of a linear space whose growth can somehow be controlled, then
this functional can be extended to the whole space such that the growth remains under
control.
Denition 2.1.1 Let E be a linear space. A map p : E R is called a sublinear
functional if
p(x +y) p(x) +p(y) (x, y E)
27
and
p(x) = p(x) (x E, R, 0).
Lemma 2.1.2 Let E be a linear space over R, let F be a subspace of E, let x
0
E F,
and let p: E R be a sublinear functional. Suppose that : F R is a linear functional
such that
(x) p(x) (x F).
Then there is a linear functional

: F +Rx
0
R which extends and satises

(x) p(x) (x F +Rx

0
).
Proof We need to show that there is R such that
(x) +t p(x +tx
0
) (x F, t R).
If this is done, we can dene

by letting

(x +tx
0
) := (x) +t for all x F and t R.
For any x, y F, we have
(x) +(y) = (x +y) p(x x
0
+x
0
+y) p(x x
0
) +p(x
0
+y)
and thus
(x) p(x x
0
) p(y +x
0
) (y) (x, y F). (2.1)
Let := sup(x) p(x x
0
) : x F. It follows from (2.1) that
(x) p(x x
0
) p(y +x
0
) (y) (x, y F)
and thus
(x) p(x x
0
) (x F) (2.2)
and
(x) + p(x +x
0
) (x F). (2.3)
Let t R, and let x F. If t > 0, we obtain from (2.3):
(x) +t = t
_

_
1
t
x
_
+
_
t p
_
1
t
x +x
0
_
= p(x +tx
0
).
For t < 0, inequality (2.2) yields:
(x) +t = t
_

_
1
t
x
_

_
t p
_
1
t
x x
0
_
= p(x +tx
0
).
This completes the proof. .
28
Theorem 2.1.3 (HahnBanach theorem) Let E be a linear space over R, let F be a
subspace of E, and let p : E R be a sublinear functional. Suppose that : F R is a
linear functional such that
(x) p(x) (x F).
Then there is a linear functional

: E R which extends and satises

(x) p(x) (x E).

Proof Let o be the collection of all pairs (X, ) with the following properties:
X is a subspace of E with F X;
: X R is linear with [
F
= ;
(x) p(x) (x X).
Clearly, (F, ) o.
Dene an order on o:
(X
1
,
1
) (X
2
,
2
) : X
1
X
2
and
2
[
X
1
=
1
.
Let T o be totally ordered. Dene

X :=
_
X : (X, ) T .
Then

X is a subspace of E with F

X. Dene

:

X R by letting

(x) := (x) if
x X for (X, ) T . Then

:

X R is well-dened, linear, extends , and satises

(x) p(x) (x

X).
It follows that (

X,

) o is an upper bound for T . Hence, by Zorns lemma, o has
a maximal element (X
max
,
max
). We claim that X
max
= E. Otherwise, there is x
0

E X
max
. By Lemma 2.1.2, there is a linear extension

max
: X
max
+ Rx
0
R of
max
such that

max
(x) p(x) (x X
max
+Rx
0
).
This contradicts the maximality of (X
max
,
max
). .
Exercise 2.1 A Banach limit on

(R) is a linear functional Lim:

(R) R such that for any

sequence (x
n
)

n=1
(we write Lim
n
x
n
instead of Lim((x
n
)

n=1
)):
(a) liminf
n
x
n
Lim
n
x
n
limsup
n
x
n
;
(b) Lim
n
x
n+k
= Lim
n
x
n
for all k N.
29
What is Lim
n
x
n
if (x
n
)

n=1
converges?
(i) Show that | Lim| = 1.
(ii) Show that Banach limits do exist. (Hint: Let F be the subspace of

(R) consisting of
those sequences (x
n
)

n=1
for which lim
n
1
n

n
k=1
x
n
exists; dene Lim on F to be that
limit, and apply the HahnBanach theorem.)
Exercise 2.2 Let E be a C-linear space.
(i) Let : E C be C-linear. Show that
(x) = Re (x) i Re (ix) (x E).
(ii) Let : E R be R-linear. Show that : E C dened through
(x) := (x) i(ix) (x E)
is C-linear.
We will rarely apply the HahnBanach theorem directly, but rather one of the following
corollaries:
Corollary 2.1.4 Let E be a linear space, let F be subspace of E, and let p: E [0, )
be a seminorm. Suppose that : F F is linear such that
[(x)[ p(x) (x F).
Then has a linear extension

: E F such that
[

(x)[ p(x) (x E).

Proof Suppose rst that F = R. By Theorem 2.1.3, we have a linear extension

: E R
such that

(x) p(x) (x E).

If

(x) 0, then

(x) =

(x) p(x) = p(x),
so that
p(x)

(x) p(x) (x E).
Now consider the case where F = C. Dene : F R through
(x) := Re (x) (x F).
30
By Exercise 2.2(i), we then have
(x) = (x) i(ix) (x F).
By the rst case, has an R-linear extension

: E R with
[

(x)[ p(x) (x E).

Dene

: E R by letting

(x) :=

(x) i

(ix) (x E).
By Exercise 2.2(ii),

is C-linear and clearly extends . Let x E, and choose C
with [[ = 1 such that

(x) = [

(x)[. We obtain:
[

(x)[ =

(x) =

(

x) =

(

x) p(

x) = p(x).
This proves the claim in the complex case. .
Corollary 2.1.5 Let E be a normed space, let F be subspace, and let F

. Then
has an extension

E

with |

| = ||.
Proof Apply Corollary 2.1.4 with p(x) := |||x| for x E. .
Exercise 2.3 Let S ,= be a set, let E be a normed space, and let F be a subspace of E. Show
that any operator T B(F,

(S)) has an extension

T B(E,

(S)) with |

T| = |T|.
Corollary 2.1.6 Let E be a normed space, let F be a closed subspace of E, and let
x
0
E F. Then there is E

with || = 1, [
F
= 0, and (x
0
) = dist(x
0
, F).
Proof Dene
p: E [0, ), x dist(x, F)(:= infx y| : y F)
and
: F +Fx
0
, x +x
0
dist(x
0
, F).
It follows that
[(x)[ dist(x, F) (x F +Fx
0
).
By Corollary 2.1.4, has a linear extension

to all of E with
[

(x)[ dist(x, F) |x| (x E),

so that, in particular, |

| 1. Let > 0. Then there is y F with |x

0
y|
dist(x
0
, F) +. Let z :=
x
0
y
|x
0
y|
, so that |z| = 1. It follows that
[

(z)[ =
(x
0
y)
|x
0
y|
=
(x
0
)
|x
0
y|

dist(x
0
, F)
dist(x
0
, F) +
.
Since > 0 was arbitrary, this means |

| 1. .
31
Corollary 2.1.6 can be used to prove approximation theorems: Let x
0
be an element
of a normed space E and assume that it is not in the closure of a subspace F. Then
Corollary 2.1.6 yields E

which vanishes on F but not in x

0
. Since for many spaces E
we have a concrete description of E

, this may then be used to arrive at a contradiction,

so that x
0
must lie in the closure of F.
Exercise 2.4 A metric space is called separable if it has a countable dense subset (any subset of
a separable metric space is again separable).
(i) Show that c
0
as well as
p
for p [1, ) are separable.
(ii) Show that

is not separable (Hint: Show that the subset of

consisting of those f

such that f(N) 0, 1 is uncountable and conclude that, for this reason, it cannot be
separable.)
Exercise 2.5 Let E be a normed space such that E

is separable. Show that E must be separable

as well. Proceed as follows:
Let
n
: n N be a dense subset of E

: || = 1. For each n N pick x

n
E
with |x
n
| 1 and [
n
(x
n
)[
1
2
.
Use the HahnBanach theorem to show that the linear span of x
n
: n N is dense in E.
Does, conversely, the separability of E imply that E

is separable? Can (

=
1
hold?
Corollary 2.1.7 Let E be a normed space, and let x E. Then there is E

with
|| = 1 and (x) = |x|.
Proof Apply Corollary 2.1.6 with F = 0. .
2.2 Applications of the HahnBanach theorem
We now present several application of the HahnBanach theorem.
2.2.1 The bidual of a normed space
Denition 2.2.1 The bidual E

of a normed space E is dened as (E

.
There is a canonical map J : E E

dened by
(Jx)() := (x) (x E, E

).
By Corollary 2.1.7 we have:
|Jx| = sup[(x)[ : E

, || 1 = |x| (x E).
Hence, J is an isometry and we may identify JE with E. In particular, every normed
space is the subspace of a Banach space.
32
2.2.2 Transpose operators
Denition 2.2.2 Let E and F be normed spaces, and let T B(E, F). The transpose
T

: F

of T is dened through
(T

)(x) := (Tx) (x E, F

)
Exercise 2.6 Let E, F and G be normed spaces, let S, T B(E, F), R B(F, G), and , F.
Show that:
(i) (T +S)

= S

+T

;
(ii) (RT)

= T

.
Theorem 2.2.3 Let E and F be normed spaces, and let T B(E, F). Then T

B(F

, E

) with |T

| = |T|.
Proof For x E and F

, we have
[(T

)(x)[ = [(Tx)[ |||T||x|,

therefore
|T

| |T|||,
and eventually |T

| |T|
Consider T

: E

. We have |T

| |T

|. On the other hand, we have for

x E and F

:
(T

Jx)() = (Jx)(T

) = (T

)(x) = (Tx) = (JTx)().

Hence, T

extends T, so that, in particular, |T

| |T|. .
The following theorem has nothing to do with the HahnBanach theorem, but we will
need it later and it ts into the discussion of transpose operators.
Theorem 2.2.4 (Schauders theorem) Let E and F be normed spaces, and let T
/(E, F). Then T

/(F

, E

).
Proof Let (
n
)

n=1
be a sequence in F

bounded by C 0. Let
K := T(x E : |x| 1).
Then K is a compact metric space. For y, z K and n N, we have
[
n
(y)
n
(z)[ |y z|.
33
Consequently, the sequence (
n
[
K
)

n=1
in ((K) is bounded and equicontinuous. By the
Arzel`aAscoli theorem, there is a subsequence (
n
k
[
K
)

k=1
converging uniformly to some
function in ((K). In particular, (
n
k
[
K
)

k=1
is a Cauchy sequence with respect to the
uniform norm. For k, l N, however, we have:
|
n
k
[
K

n
l
[
K
|

= sup
yK
[
n
k
(y)
n
l
(y)[
sup[
n
k
(Tx)
n
l
(Tx)[ : x E, |x| 1
= |T

n
k
T

n
l
|.
Hence, (T

n
k
)

k=1
is a Cauchy sequence in E

and thus convergent. .

2.2.3 Quotient spaces and duals
Denition 2.2.5 Let E be a normed space, and let F be a closed subspace of F. The
quotient norm on E/F is dened as
|x +F| := inf|x y| : y F (x E).
Theorem 2.2.6 Let E be a normed space, and let F be a closed subspace. Then E/F
equipped with the quotient norm is normed space. If E is a Banach space, then so is E/F.
Proof It is routine to verify that the quotient norm is indeed a norm.
Let (x
n
)

n=1
be a sequence in E such that

n=1
|x
n
+F| < . For each n N, choose
y
n
F such that |x
n
y
n
| <
1
2
n
. It follows that

n=1
|x
n
y
n
| < . Since E is a Banach
space,

n=1
(x
n
y
n
) converges in E to x E. It is clear that x + F =

n=1
(x
n
+ F).
.
Exercise 2.7 Let E be a normed space, and let F be a closed subspace. Let G be another normed
space, and let T B(E, G) vanish on F. Show that

T(x +F) := Tx (x E)
denes

T B(E/F, G) with |

T| = |T|.
Denition 2.2.7 Let E be a normed space. For any subset S of E, we dene
S

:= E

: [
S
= 0.
Theorem 2.2.8 Let E be a normed space, and let F be a closed subspace of E. Then
T : E

, [
F
induces an isometric isomorphism of E

/F

and F

.
34
Proof Clearly, |T| 1 and ker T = F

, so that T, by Exercise 2.7, T induces an injective

map

T : E

/F

with |

T| 1. We claim that this map is surjective and an isometry.

Let F

. By Corollary 2.1.5, there is E

with || = || extending . We
thus have T = and
| +F

| |

T( +F

)| = || = || | +F

|,
which completes the proof. .
Theorem 2.2.9 Let E be a normed space, and let F be a normed subspace. Then T :
F

(E/F)

dened by
(T)(x +F) := (x) ( F

, x E)
is an isometric isomorphism of F

and (E/F)

.
Proof It is routinely checked that T is well-dened and linear.
Let : E E/F be the quotient map. For any (E/F)

, the functional
belongs to F

such that T( ) = . Hence, T is surjective. From Exercise 2.7, it

follows that T is an isometry. .
2.2.4 The dual space of (([0, 1])
Denition 2.2.10 A function : [0, 1] F is said to be of bounded variation if
||
BV
:= sup
_
_
_
n

j=1
[(x
j
) (x
j1
)[ : n N, 0 = x
0
< x
1
< < x
n
= 1
_
_
_
< .
We dene:
BV ([0, 1]) := : [0, 1] F : is of bounded variation
The following are easily checked:
BV ([0, 1]) is a linear space;
| |
BV
is a seminorm on BV ([0, 1]);
||
BV
= 0 is constant.
We let
BV
0
([0, 1]) := BV ([0, 1]) : (0) = 0.
Then | |
BV
is a normed space.
35
Theorem 2.2.11 The linear map T : BV
0
([0, 1]) (([0, 1])

dened by
(T)(f) :=
_
1
0
f(x) d(x)
is an isometric isomorphism.
Proof It is obvious that |T| ||
BV
for all BV
0
([0, 1]).
Conversely, let (([0, 1])

. Let u
0
: 0, and for any x (0, 1], dene u
x
: [0, 1] R
by letting
u
x
(t) :=
_
1, 0 t x,
0, x < t 1.
By Corollary 2.1.5, there is an extension

([0, 1])

of with |

| = ||. Dene
: [0, 1] F, x

(u
x
).
We claim that BV
0
([0, 1]). Let 0 = x
0
< x
1
< < x
n
= 1, and dene, for
j = 1, . . . , n:

j
:=
_
[(x
j
)(x
j1
)[
(x
j
)(x
j1
)
, if (x
j
) ,= (x
j1
),
0, otherwise.
We then have:
n

j=1
[(x
j
) (x
j1
)[ =
n

j=1

j
((x
j
) (x
j1
))
=
n

j=1

j
(

(u
x
j
)

(u
x
j1
))
=

_
_
n

j=1

j
(u
x
j
u
x
j1
)
_
_
. .
||

1
|

|
= ||.
Hence, is a bounded variation such that ||
BV
||.
Next, we claim that T = (this establishes at once that T is a surjective isometry
and thus an isometric isomorphism). For any f (([0, 1]) and any partition T = 0 =
x
0
< x
1
< < x
n
= 1 of [0, 1], let
(T) = sup
j=1,... ,n
[x
j
x
j1
[
36
and
S(f, T) :=
n

j=1
f(x
j
)((x
j
) (x
j1
)).
From the properties of the RiemannStieltjes integral, we know that
lim
(1)0
S(f, T) =
_
1
0
f(x) d(x).
Dene
f
1
:=
n

j=1
f(x
j
)(u
x
j
u
x
j1
)
From the uniform continuity of f, we infer that lim
(1)0
|f
1
f|

= 0. We thus have:
(f) = lim
(1)0

(f
1
)
= lim
(1)0
n

j=1
f(x
j
)(

(u
x
j
)

(u
x
j1
))
= lim
(1)0
n

j=1
f(x
j
)((x
j
) (x
j1
))
= lim
(1)0
S(f, T)
=
_
1
0
f(x) d(x).
This establishes the claim and thus completes the proof. .
This result is only a rather special case of Riesz representation theorem (Theorem
B.3.8).
Exercise 2.8 Let C 0, c
1
, c
2
, . . . F, and f
1
, f
2
, . . . (([0, 1]) be given. Show that the following
are equivalent:
(a) There is BV [0, 1] with ||
BV
C such that
c
n
=
_
1
0
f
n
(t) d(t) (n N).
(b) For all n N, and for all
1
, . . . ,
n
F, we have

k=1

k
c
k

C
_
_
_
_
_
n

k=1

k
f
k
_
_
_
_
_

.
37
2.2.5 Runges approximation theorem
We now use Corollary 2.1.6 to prove an approximation theorem from complex analysis:
Theorem 2.2.12 (Runges approximation theorem) Let K C be compact, and let
E C

K be such that E has at least one point in common with each component of
C

K. Let U C be an open set containing K, and let f : U C be holomorphic.

Then, for each > 0, there is a rational function with poles in E such that
sup
zK
[f(z) r(z)[ < .
Proof Note that f[
K
((K). Let
1
E
(K) := r[
K
: r is a rational function with poles in E.
We need to show that f[
K
1
E
(K).
Assume that this is not true. By Corollary 2.1.6, there is ((K)

with
[
1
E
(K)
= 0 and (f[
K
) ,= 0.
By the Riesz representation theorem (Theorem B.3.8), there is M(K) such that
(g) =
_
K
g(z) d(z) (g ((K)).
Dene
: C K C, w
_
K
d(z)
z w
.
We claim that 0. Let V be a component of C

K, and let p E V .
Case 1: p ,= . Then choose r > 0 such that B
r
(p) V . For xed w B
r
(p), we
then have uniformly in z K:
1
z w
=
1
(z p) (w p)
=
1
z p
1
1
wp
zp
=
1
z p

n=0
_
w p
z p
_
n
=

n=0
(w p)
n
(z p)
n+1
.
Hence, the function
1
zw
of z belongs to 1
E
(K). It follows that (w) = 0. Since w B
r
(p)
was arbitrary, the identity theorem yields [
V
0.
38
Case 2: p = . Choose r > 0 so large that [z[ < r for all z K, and let w C with
[w[ > r. Then we have uniformly in z K:
1
z w
=
1
w
1
z
w
1
=

n=0
z
n
w
n+1
.
It follows again that (w) = 0 for [w[ > r and thus [
V
0.
Let by a cycle in U whose winding number around each point in K is 1 and around
each point in C U is zero. By the Cauchy integral formula, we have
f(z) =
1
2i
_

f(w)
z w
dw (z K).
But this yields
_
K
f(z) d(z) =
1
2i
_
K
__

f(w)
z w
dw
_
d(z)
=
1
2i
_

__
K
f(w)
z w
d(z)
_
dw
=
1
2i
_

f(w)
__
K
1
z w
d(z)
_
dw
=
1
2i
_

f(w) (w)dw
= 0,
which contradicts the choice of . .
The advantage of this functional analytic proof is its brevity and its elegance. The
drawback is that it is not constructive: It depends on the HahnBanach theorem and
therefore on Zorns lemma.
2.3 Baires theorem
Theorem 2.3.1 (Baires theorem) Let X be a complete metric space, and let (U
n
)

n=1
be a sequence of dense open subsets of X. Then

n=1
U
n
is dense in X.
Proof Assume that the theorem is wrong. Then there are x
0
X and > 0 such that
B

(x
0
) X

n=1
U
n
.
Let V
0
:= B

(x
0
). Since U
1
is dense in X, there is x
1
U
1
V
0
. Choose r
1
(0, 1) so
small that
B
r
1
(x
1
) U
1
V
0
.
Let V
1
:= U
r
1
(x
1
). Suppose that open subsets V
0
, V
1
, . . . , V
n
of X have already been
constructed such that
39
V
j+1
U
j+1
V
j
for j = 0, . . . , n 1, and
diam V
j

2
j
for j = 1, . . . , n.
Since U
n+1
is dense in X, there is x
n+1
U
n+1
V
n
. Choose r
n+1

_
0,
1
n+1
_
so small
that
B
r
n+1
(x
n+1
) U
n+1
V
n
,
and let V
n+1
:= B
r
n+1
(x
n+1
). Continue inductively.
Since diam V
n

2
n
for all n N, and since X is complete, there is x

n
n=1
V
n
. By
construction, however,
x
n

n=1
V
n

n=1
U
n
and x
n
_
n=1
V
n
V
0
X

n=1
U
n
,
which is impossible. .
Corollary 2.3.2 Let X be a complete metric space, and let (F
n
)

n=1
be a sequence of
closed subsets of X such that

n=1
F
n
has an interior point. Then at least one F
n
has an
interior point.
Proof Let U
n
:= X F
n
. .
Example Let E be a Banach space with a countable Hamel basis. We claim that dimE <
.
Assume that E has a Hamel basis x
n
: n N. For n N, let
E
n
:= linx
1
, . . . , x
n
.
Then E
n
is a closed subspace of E. Since E =

n=1
E
n
, Corollary 2.3.2 yields that there
is N N such that E
N
has interior points, i.e. there is x
0
E
N
and > 0 such that
B

(x
0
) E
N
. Let x := x
0
+

2
x
N+1
|x
N+1
|
. Then x B

(x
0
), but x / E
N
.
In particular, there is no norm on c
00
turning it into a Banach space.
For our next application of Baires theorem, we need the following approximation
theorem:
Theorem 2.3.3 (Weierstra approximation theorem) Let a, b R, a < b, let f
(([a, b]), and let > 0. Then there is a polynomial p such that |f p|

< .
Proof Without loss of generality, let a = 0, b = 1, and F = R.
For each g (([0, 1]), let
B
n
(g; t) :=
n

k=0
_
n
k
_
t
k
(1 t)
nk
g
_
k
n
_
(t [0, 1]).
40
be its n-th Bernstein polynomial . It is routinely checked that
B
n
(1, t) =
n

k=0
_
n
k
_
t
k
(1 t)
nk
= (t + (1 t))
n
= 1,
B
n
(x, t) =
n

k=0
_
n
k
_
t
k
(1 t)
nk
k
n
=
n

k=1
_
n 1
k 1
_
t
k
(1 t)
nk
=
n1

k=0
_
n 1
k
_
t
k+1
(1 t)
n(k+1)
= t
n1

k=0
_
n 1
k
_
t
k
(1 t)
(n1)k
= t(t + (1 t))
n1
= t,
and
B
n
(x
2
, t) =
n

k=0
_
n
k
_
t
k
(1 t)
nk
_
k
n
_
2
=
n1

k=0
_
n 1
k
_
t
k+1
(1 t)
n(k+1)
k + 1
n
=
t
n
+
n1

k=0
_
n 1
k
_
t
k
(1 t)
(n1)k
k
n
t
=
t
n
+
n 1
n
t
2
=
t(1 t)
n
+t
2
.
Since f is uniformly continuous, there is > 0 such that [f(s) f(t)[ < for all
s, t [0, 1] with [s t[ <

. Let C :=
2|f|

. We claim that
[f(s) f(t)[ +C(t s)
2
(s, t [0, 1]). (2.4)
This is clear if [s t[ <

; if [s t[

, it follows from
+C(t s)
2
> + 2|f|

> [f(s)[ +[f(t)[ [f(s) f(t)[.

Fix t [0, 1], and let f
t
(s) := (t s)
2
. Then (2.4) implies
Cf
t
f f(t) +Cf
t
41
and thus
B
n
( +Cf
t
, ) = B
n
( Cf
t
, ) B
n
(f f(t), ) B( +Cf
t
, ).
Since B
n
(f f(t), ) = B
n
(f, ) f(t), we obtain for p
n
:= B
n
(f, ):
[p
n
(s) f(t)[ B
n
( +Cf
t
, s) = +Ct
2
2Cts +C
_
s(1 s)
n
+s
2
_
(s [0, 1]).
Letting s = t, this yields
[p
n
(t) f(t)[ +C
t(1 t)
n
+
C
n
.
Since > 0 and t [0, 1] are arbitrary, this yields |p
n
f|

0. .
There are proofs of Theorem 2.3.3 that use the HahnBanach theorem, but they are
in no way easier than the one given here.
Example We will now use Baires theorem to prove that there is a continuous function
on [0, 1] which is nowhere dierentiable.
For n N, let
F
n
:=
_
f (([0, 2]) : there is t [0, 1] such that sup
h(0,1)
[f(t +h) f(t)[
h
n
_
.
Obviously, if f (([0, 2]) is dierentiable at some t [0, 1], then sup
h(0,1)
[f(t+h)f(t)[
h
<
, so that f

n=1
F
n
.
Let (f
k
)

k=1
be a sequence in F
n
such that |f
k
f|

0 for some f (([0, 2]). For

each k N, there is t
k
[0, 1] such that
sup
h(0,1)
[f
k
(t
k
+h) f
k
(t
k
)[
h
n.
Suppose without loss of generality that (t
k
)

k=1
converges to some t [0, 1]. Fix h (0, 1)
and > 0, and choose K N so large that
_

_
[f(t +h) f(t
k
+h)[
|f f
k
|

[f(t
k
) f(t)[
_

_
<

4
h (k K).
For k K, this implies
[f(t +h) f(t)[
[f(t +h) f(t
k
+h)[
. .
<

4
h
+[f(t
k
+h) f
k
(t
k
+h)[
. .
<

4
h
+[f
k
(t
k
+h) f
k
(t
k
)[
. .
nh
+[f(t
k
) f
k
(t
k
)[
. .
<

4
h
+[f(t) f(t
k
)[
. .
<

4
h
nh +h,
42
so that
[f(t+h)f(t)[
h
n+. Since h and were arbitrary, this means that f F
n
. Hence,
F
n
is closed.
Assume that every f (([0, 2]) is dierentiable at some point in [0, 1]. Then (([0, 2]) =

n=1
F
n
, so that, by Corollary 2.3.2, there are N N, f (([0, 2]), and > 0 such that
B

(f) F
N
. By Theorem 2.3.3, B

(f) contains at least one polynomial. Without loss of

generality, we may thus suppose that f (
1
([0, 2]).
For m N and j = 0, . . . , m, let t
j
:=
2j
m
. Dene g
m
: [0, 2] F by letting
g
m
(t) :=
_

2
m(t t
j1
), t
_
t
j1
, t
j1
+
1
m

2
m(t
j1
t), t
_
t
j

1
m
, t
j

.
Then g
m
(([0, 2]) with |g
m
| =

2
, but
sup
h(0,1)
[g
m
(t +h) g
m
(t)[
h


2
m (2.5)
holds for any t [0, 1]. Since f +g
m
B

(f) F
N
, there is t [0, 1] such that
sup
h(0,1)
[(f +g
m
)(t +h) (f +g
m
)(t)[
h
N.
This, however, yields
sup
h(0,1)
[g
m
(t +h) g
m
(t)[
h
sup
h(0,1)
[(f +g
m
)(t +h) (f +g
m
)(t)[
h
+ sup
h(0,1)
[f(t +h) f(t)[
h
= N +|f
t
|

,
which contradicts (2.5) if we choose m N so large that

2
m > N +|f
t
|

.
Exercise 2.9 Let (f
k
)

k=1
be a sequence in (([0, 1]) which converges pointwise to a function
f : [0, 1] F.
(i) For > 0 and n N let
F
n
:= t [0, 1] : [f
n
(t) f
k
(t)[ for all k n.
Show that F
n
is closed, and that [0, 1] =

n=1
F
n
.
(ii) Let > 0, and let I be a non-degenerate, closed subinterval of [0, 1]. Show that there is a
non-degenerate, closed subinterval J of I such that
[f(t) f(s)[ (t, s J).
(Hint: Apply (a) with :=

3
and Baires theorem.)
43
(iii) Let I be a non-degenerate, closed subinterval of [0, 1]. Show that there is a decreasing
sequence of non-degenerate, closed subintervals of I such that
the length of I
n
is at most
1
n
, and
[f(t) f(s)[
1
n
for all s, t I
n
.
What can be said about f at all points in

n=1
I
n
?
(iv) Conclude that the set of points in [0, 1] at which f is continuous is dense in [0, 1].
2.4 The uniform boundedness principle
Theorem 2.4.1 (uniform boundedness principle) Let E be a Banach space, let
(F

be a family of normed spaces, and let T

B(E, F

) be such that
sup

|T

x| < (x E). (2.6)

Then sup

|T

| < holds.
Proof For n N, let
E
n
:=
_
x E : sup

|T

x| n
_
.
Then (2.6) implies that E =

n=1
. Let (x
k
)

k=1
be a sequence in E
n
such that x
k
x E.
For any index , we have
|T

x| = lim
k
|T

x
k
| n.
It follows that x E, so that E
n
is closed.
By Corollary 2.3.2 there are thus N N, x
0
E, and > 0 such that U

(x
0
) E
N
.
Let x E be such that |x| 1. It follows that x +x
0
E
N
. Hence, we have for all :
|T

x| = |T

(x)| |T

(x +x
0
)| +|T

x
0
| 2N,
and consequently
|T

x|
2N

.
It follows that sup

|T

|
2N

. .
Examples 1. Let E be a Banach space, let F be a normed space, and let (T
n
)

n=1
be
a sequence in B(E, F) such that lim
n
T
n
x exists in F for each x E. We claim
that T : E F dened through
Tx := lim
n
T
n
x (x E).
44
Since (T
n
x)

n=1
is convergent and thus bounded for each x E, Theorem 2.4.1
implies that C := sup
nN
|T
n
| < . Let x E. Then there is N N such that
|T
N
x Tx| |x|. This yields
|Tx| |T
N
x| +|T
N
x Tx| C|x| +|x| = (C + 1)|x|.
2. For each continuous function f : [, ] R, its Fourier series is
a
0
2
+

k=1
(a
k
cos(kx) +b
k
sin(kx)), (2.7)
where
a
k
=
1

f(x) cos(kx) dx (k N
0
)
and
b
k
=
1

f(x) sin(kx) dx (k N).

We will show that there are f (([, ]) for which (2.7) does not converge in
every point x [, ].
For n N, let
D
n
(x) :=
sin
__
n +
1
2
_
x
_
2 sin
x
2
(x [, ]),
and dene
s
n
: (([, ]) R, f
1

f(x)D
n
(x) dx.
It can be shown that, for any f (([, ]) with Fourier series (2.7), we have
s
n
(f) =
a
0
2
+
n

k=1
a
k
,
i.e. s
n
(f) is the n-th partial sum of (2.7) at x = 0. It is easy to see that (s
n
)

n=1
is
a sequence in (([, ], R)

such that
|s
n
|
1

[D
n
(x)[ dx (n N).
Let > 0, and let x
1
, . . . , x
m
[, ] be the zeros of D
n
in [, ]. Choose > 0
such that
2m|D
n
|

2
.
45
Dene f (([, ]) through
f(x) :=
_
[D
n
(x)[
D
n
(x)
, x /

m
j=1
(x
j
, x
j
+) =: I

,
linear in between.
It is clear that |f|

1. It follows that
|s
n
| [s
n
(f)[
=
1

_
[,]\I

[D
n
(x)[ dx +
_
I

f(x)D
n
(x) dx

_
[,]\I

[D
n
(x)[ dx
1

_
I

[D
n
(x)[ dx
. .

_
[,]\I

[D
n
(x)[ dx

2
,
consequently

2
+|s
n
|
1

_
[,]\I

[D
n
(x)[ dx
=
1

_
[,]
[D
n
(x)[ dx
1

_
I

[D
n
(x)[ dx
. .

_
[,]
[D
n
(x)[ dx

2
,
and nally
+|s
n
|
1

_
[,]
[D
n
(x)[ dx.
Since > 0, this yields
|s
n
| =
1

[D
n
(x)[ dx =
2

_

0
[D
n
(x)[ dx.
Since
_

0

sin
__
n +
1
2
_
x
_

sin
x
2
dx
_

0

sin
__
n +
1
2
_
x
_

x
2
dx
= 2
_
(n+
1
2
)
0
[ sin y[
y
dy
,
it follows that |s
n
| . Hence, Theorem 2.4.1 implies that there is f (([, ])
such that sup
nN
[s
n
(f)[ = . In particular, the Fourier series of f diverges at
x = 0.
46
Exercise 2.10 Prove the BanachSteinhaus theorem: Let E and F be Banach spaces, let T
B(E, F), and let (T
n
)

n=1
be a sequence in B(E, F). Then the following are equivalent:
(a) Tx = lim
n
T
n
x for each x E;
(b) sup
nN
|T
n
| < and Tx = lim
n
T
n
x for all x in some dense subspace of E.
Exercise 2.11 A family (x

of elements in a normed space E is called weakly bounded if

sup

[(x

)[ <
for all E

. Show that a family (x

in a normed space is bounded if and only if it is weakly

bounded.
Exercise 2.12 Let E, F, and G be Banach spaces, and let T : E F G be a bilinear map
which is continuous in each variable. Show that there is C 0 such that
|T(x, y)| C|x||y| (x E, y F).
2.5 The open mapping theorem
Denition 2.5.1 Let E and F be normed spaces. A linear map T : E F is called open
if TU is open in F for every open subset U of E.
Theorem 2.5.2 (open mapping theorem) Let E and F be Banach spaces, and let
T B(E, F) be surjective. Then T is open.
Proof We claim that, for each r > 0, the zero vector 0 is an interior point of TB
r
(0).
Since T is surjective, we have
F =

_
n=1
TB
nr
2
(0).
By Corollary 2.3.2, there is N N such that
TBNr
2
(0) = N TB
r
2
(0)
has an interior point, say x
0
. Hence, there is > 0 such that B

(x
0
) TB
r
2
(0). For any
x B

(0), we then have

x = x +x
0
x
0
TB
r
2
(0) +TB
r
2
(0) TB
r
(0).
This proves the rst claim.
Secondly, we claim that TB
r
2
(0) TB
r
(0) for all r > 0.
Fix y
1
TB
r
2
(0). Since 0 is an interior point of TB
r
4
(0), it follows that
(y
1
TB
r
4
(0)) TB
r
2
(0) ,= .
47
Choose x
1
B
r
2
(0) such that Tx
1
y
1
TB
r
4
(0). Then choose y
2
TB
r
4
(0) such that
Tx
1
= y
1
y
2
. Continuing in this fashion, we obtain sequences (x
n
)

n=1
in E and (y
n
)

n=1
in F such that
_

_
x
n
B
r
2
n
(0),
y
n
TB
r
2
n
(0),
y
n+1
= y
n
Tx
n
.
_

_
(n N).
Since

n=1
|x
n
| < , the series

n=1
x
n
converges to some x E with |x| < r.
Moreover, we have
Tx =

n=1
Tx
n
=

n=1
(y
n
y
n+1
)
= lim
N
N

n=1
(y
n
y
n+1
)
= lim
N
(y
1
y
N+1
)
= y
1
,
so that y
1
TB
r
(0).
Finally, we deduce that T is indeed open.
Let U E be open, and let x U. Choose r > 0 such that B
r
(x) U. Since 0 is
an interior point of TB
r
(0), it follows that Tx is an interior point of TB
r
(x) and thus of
TU. Since x U was arbitrary, this means that TU is open. .
Exercise 2.13 Let E and F be Banach spaces, and let T B(E, F) be such that dimF/TE <
. Show that T has closed range. (Hint: Choose a nite-dimensional subspace G of F with
F = TE +G and TE G = 0, and consider
S: E G F, (x, y) Tx +y.
Apply the open mapping theorem.)
Exercise 2.14 Let E and F be normed spaces, and let T : E F be an open linear map. Show
that T is surjective. (Warning Trick question.)
Corollary 2.5.3 Let E and F be Banach spaces, and let T B(E, F) be bijective. Then
T is an isomorphism, i.e. T
1
B(F, E).
Exercise 2.15 Let E and F be Banach spaces, and let T B(E, T) be surjective. Show that
there is C 0 such that, for each y F, there is x E with |x| C|y| such that Tx = y.
48
Exercise 2.16 Let E be a Banach space. An operator T B(E) is called quasi-nilpotent if
lim
n
n
_
|T
n
| = 0.
Show that a quasi-nilpotent operator can never be surjective unless E = 0. (Hint: Previous
problem.)
Exercise 2.17 Let E and F be Banach spaces. Show that the following are equivalent for T
B(E, F):
(a) T is injective and has closed range.
(b) There is C 0 such that
|x| C|Tx| (x E).
Examples 1. Let E be a Banach space. A closed subspace F is called complemented
in E if there is another closed subspace G of E with E = F +G and F G = 0.
We claim that, if F is complemented, then the canonical projection P : E F with
ker P = G is continuous.
Clearly,

E = F G
with
|(x, y)| = max|x|, |y|| (x F, y G)
is a Banach space. Let :

E F be the projection onto the rst coordinate, and
let
T :

E E, (x, y) x +y.
Then T is continuous and bijective and thus has a continuous inverse. Since P =
T
1
this shows that P is continuous.
2. Given f
0
, f
1
(([0, 1]), the initial value problem
y
tt
+f
1
y
t
+f
0
y = g, y(0) = y
1
, y
t
(0) = y
2
(2.8)
has a unique solution in (
2
([0, 1]) for all g (([0, 1]) and y
1
, y
2
R. Let E =
(
2
([0, 1]) (equipped with | |
2
, and let
F := (([0, 1]) R R
be equipped with
|(f, x
1
, x
2
)| := max|f|

, [x
1
[, [x
2
[ (f (([0, 1]), x
1
, x
2
R).
49
Dene T : E F through
T := (
tt
+f
1

t
+f
0
, (0),
t
(0)) ( (
2
([0, 1])).
Then T is linear such that, for any (
2
([0, 1]):
|T| = max|
tt
+f
1

t
+f
0
|

, [(0)[, [
t
(0)[
max1, |f
1
|

, |f
0
|

j=0
|
(j)
|

. .
=||
2
.
Hence, T is bounded. The existence and uniqueness of the solutions of (2.8) imply
that T is a bijection. By Corollary 2.5.3, T
1
is also continuous. Hence, the solutions
of (2.8) depend continuously on the data g, y
1
. and y
2
.
Exercise 2.18 Show that c
0
is not complemented in

:
(i) Show that there is an uncountable family (S

of innite subsets of N such that S

is nite for all ,= . (Hint: Replace N by Q (you can do that because they have the same
cardinality), use R as your index set, and utilize the fact that every real number is the limit
of a sequence in Q.)
(ii) There is no countable subset of (

/c
0
)

such that for each non-zero f

/c
0
there is
with (f) ,= 0. (Hint: Choose (S

as in (i) and consider the family (f

of the
cosets in

/c
0
of the indicator functions of the sets S

; show that, for xed (

/c
0
)

,
the set f

: (f

) ,= 0 is at most countable.)
(iii) Conclude from (i) and (ii) that c
0
is not complemented in

.
2.6 The closed graph theorem
Denition 2.6.1 Let E and F be normed spaces.
(a) A partially dened operator from E to F is a linear map T : T
T
F, where T
T
is
a subspace of F.
(b) A partially dened operator is called closed if its graph
Gr T := (x, Tx) : x T
T

is closed in E F.
Example Let E = F := (([0, 1]), let T
T
:= (
1
([0, 1]), and let
T : T
T
F, f f
t
.
50
Let ((f
n
, Tf
n
))

n=1
be a sequence in Gr T with (f
n
, Tf
n
) (g, h) E F, i.e.
|f
n
g|

0 and |f
t
n
h|

0.
It follows that (f
n
)

n=1
is a Cauchy sequence in (
1
([0, 1]) with respect to | |
1
. Let
f (
1
([0, 1]) be the limit of (f
n
)

n=1
, i.e.
|f
n
f|

0 and |f
t
n
f
t
|

0.
It follows that g = f and h = f
t
, so that (g, h) Gr T. Consequently, T is closed
(although not continuous).
Theorem 2.6.2 (closed graph theorem) Let E and F be Banach spaces, and let T :
E F be closed. Then T is continuous.
Proof Dene

1
: Gr T E, (x, Tx) x.
Then T is a continuous bijection and thus its inverse
: E Gr T, x (x, Tx)
is continuous as well. Let

1
: Gr T F, (x, Tx) Tx.
Then
2
is continuous, and so is T =
2
. .
Corollary 2.6.3 Let E and F be Banach spaces, and let T : E F be linear with the
following property:
If (x
n
)

n=1
is a sequence in E and y is a vector in F such that x
n
0 and
Tx
n
y F, then y = 0.
Then T is continuous.
Proof Let (x
n
)

n=1
be a sequence in E, and let x E and y F be such that
|x
n
x| 0 and |Tx
n
y|

0.
It follows that
x
n
x 0 and T(x
n
x) = Tx
n
Tx y Tx.
The hypothesis on T implies that y Tx = 0, so that (x, y) Gr, T. Hence, Gr T is
closed, and T is continuous by Theorem 2.6.2. .
51
Exercise 2.19 Let E and F be Banach spaces, and let T : E F be linear. The separating space
of T is dened as
S(T) := y F : there is a sequence (x
n
)

n=1
in E with x
n
0 and Tx
n
y
(i) Show that S(T) is a closed, linear subspace of F.
(ii) Show that S(T) = 0 if and only if T B(E, F).
Example Let X be a locally compact Hausdor space, and let : X F be such that
f (
0
(X) for all f (
0
(X). Dene
M

: (
0
(X) (
0
(X), f f.
Let (f
n
)

n=1
be a sequence in (
0
(X), and let g (
0
(X) be such that
f
n
0 and Tf
n
g.
For all x X, we have that
g(x) = lim
n
(x)f
n
(x) = 0,
so that g = 0. With Corollary 2.6.3, it follows that M

is bounded. (It can then be shown

that (
b
(X).)
Exercise 2.20 Let X be a locally compact Hausdor space. A multiplier of (
0
(X) is a linear
map T : (
0
(X) (
0
(X) such that
T(fg) = fTg (f, g (
0
(X))
Show that T is continuous.
52
Chapter 3
Spectral theory of bounded linear
operators
In this chapter, we develop the basics of the spectral theory of bounded, linear operators
on Banach spaces.
3.1 The spectrum of a bounded linear operator
The spectrum can be thought of as the appropriate innite-dimensional analogue of the
set of eigenvalues of a matrix.
Denition 3.1.1 Let E be a Banach space. We let
Inv B(E) := T B(E) : T is invertible.
Denition 3.1.2 Let E be a Banach space, and let T B(E). Then
(T) := F : T / Inv B(E)
is called the spectrum of T. The complement (T) := C (T) is the resolvent set of T.
Examples 1. Let dimE < . Then:
(T) T is not bijective
T is not injective
there is x E 0 such that Tx = x
is an eigenvalue of T.
2. Let E = R
2
, and let T = T
A
for A =
_
0 1
1 0
_
. Then (T) = .
53
3. Let (([0, 1]) and let
M

: (([0, 1]) (([0, 1]), f f.

Let F ([0, 1]). Then
(x) :=
1
(x)
(x [0, 1])
denes an element of (([0, 1]). We have
M

( M

)f = (M

) M

f =


f = f (f (([0, 1])),
so that M

= ( M

)
1
and / (M

). It follows that (T) ([0, 1]).

Conversely, let ([0, 1]) and assume that / (T). Let := ( M

)
1
1.
Then we obtain
( ) = ( M

) = 1,
which is impossible.
Exercise 3.1 Let E be a Banach space, and let P B(E) be a projection. Show that (P)
0, 1.
Exercise 3.2 Let T B((([0, 1])) be dened through
(Tf)(x) := xf(x) (f (([0, 1]), x [0, 1]).
(i) Show that T has no eigenvalues.
(ii) What is (T)?
Exercise 3.3 Let E be a Banach space, and let T B(T). Show that (T) = (T

).
Exercise 3.4 Let p [1, ], and let L, R:
p

p
be dened through
L(x
1
, x
2
, x
3
, . . . ) = (0, x
1
, x
2
, . . . )
and R(x
1
, x
2
, x
3
, . . . ) = (x
2
, x
3
, x
4
, . . . ) ((x
1
, x
2
, x
3
, . . . )
p
).
(i) Show that every C with [[ < 1 is an eigenvalue of R.
(ii) Conclude that (R) = C : [[ 1.
(iii) Show that L has no eigenvalues, but (L) = C : [[ 1. (Hint: Take adjoints.)
As Exercise 3.2 shows, a bounded linear operator on an innite-dimensional Banach
space may have an empty set of eigenvalues. As we shall see in the remainder of this
section, the spectrum of a bounded, linear operator is a compact set, which is non-empty
if F = C.
54
Lemma 3.1.3 Let E be a Banach space, and let T B(E) be such that |id
E
T| < 1.
Then T Inv B(E).
Proof Let
S :=

n=0
(id
E
T)
n
.
It follows that
S TS = (id
E
T)S
=

n=0
(id
E
T)
n+1
=

n=1
(id
E
T)
n
= S id
E
,
so that TS = id
E
. In a similar way, ST = id
E
is proven. .
Corollary 3.1.4 Let E be a Banach space, and let T B(E). Then (T) is bounded by
|T|.
Proof Let F be such that [[ > |T|. Then
_
_
_
_
1
_
1
T

__
_
_
_
=
_
_
_
_
T

_
_
_
_
< 1,
so that
T =
_
1
T

_
Inv B(E)
by Lemma 3.1.3. This means that (T). .
Corollary 3.1.5 Let E be a Banach space. Then Inv B(E) is open in B(E).
Proof Let T Inv B(E), and let S B(E) be such that |S T| <
1
|T
1
|
. It follows that
|1 T
1
S| = |T
1
(T S)| < 1,
so that T
1
S Inv B(E) by Lemma 3.1.3. .
Corollary 3.1.6 Let E be a Banach space, and let T B(E). Then (T) is closed in F.
55
Proof Let (T), i.e. T Inv B(E). By Corollary 3.1.5, there is > 0 such that
S Inv B(E) whenever | T S| < . For F with [ [ < , we then have
|( T) ( T)| = [ [ < ,
so that T Inv B(E). .
Lemma 3.1.7 Let E be a Banach space, and let (T
n
)

n=1
be a sequence in Inv B(E) which
converges to T Inv B(E). Then T
1
n
T
1
.
Proof We rst show that sup
nN
|T
1
n
| < . Since T
n
T
1
id
E
, there is N N such
that
|id
E
T
n
T
1
| <
1
2
(n N).
In the proof of Lemma 3.1.7, we saw that
(T
n
T
1
)
1
=

k=0
(id
E
T
n
T
1
)
k
(n N),
so that
|TT
1
n
| = |(T
n
T
1
)
1
|

k=0
|id
E
T
n
T
1
|
k
2 (n N).
Consequently,
|T
1
n
| |T
1
||TT
1
n
| 2|T
1
| (n N).
Since
|T
1
n
T
1
| = |T
1
n
(T T
n
)T
1
| |T
1
n
||T T
n
||T
1
| (n N),
it follows that lim
n
T
1
n
= T
1
. .
Theorem 3.1.8 Let E ,= 0 be a Banach space over C, and let T B(E). Then (T)
is a non-empty, compact subset of C.
Proof In view of Corollaries 3.1.4 and 3.1.6, it is clear that (T) is compact (this does
not require that the Banach space be over C).
All we have to show is therefore that (T) ,= . Assume towards a contradiction that
(T) = , i.e. T Inv B(E) for all C. Let B(E)

, and dene
f : C C, (( T)
1
).
56
Let h C 0. Then we have:
f( +h) f()
h
=
1
h
(( +h T)
1
( T)
1
)
=
1
h
(( +h T)
1
[( T) ( +h T)]( T)
1
)
=
1
h
(( +h T)
1
[( T) ( +h T)]( T)
1
)
h0
(( T)
2
).
Hence, f is holomorphic. Moreover, since
f() =
1

((1
1
T)
1
)
[[
0, (3.1)
the function f is also bounded. By Liouvilles theorem, this means that f is constant.
In conjunction with (3.1), this means f 0. In particular, 0 = f(0) = (T
1
). Since
B(E)

was arbitrary, Corollary 2.1.7 yields T

1
= 0, which is impossible. .
Exercise 3.5 Let ,= K C be compact. Show that there are a Banach space E over C and
T B(E) such that (T) = K.
Exercise 3.6 Let p [1, ], and let L, R:
p

p
be dened through
L(x
1
, x
2
, x
3
, . . . ) = (0, x
1
, x
2
, . . . )
and R(x
1
, x
2
, x
3
, . . . ) = (x
2
, x
3
, x
4
, . . . ) ((x
1
, x
2
, x
3
, . . . )
p
).
(i) Show that every C with [[ < 1 is an eigenvalue of R.
(ii) Conclude that (R) = C : [[ 1.
(iii) Show that L has no eigenvalues, but (L) = C : [[ 1. (Hint: Take adjoints.)
Exercise 3.7 Let E be a Banach space over C, and let T Inv B(E).
(i) Show that (T) if and only if
1
(T
1
).
(ii) Suppose further that T is an isometry. Show that (T) C : [[ = 1.
Exercise 3.8 Let E be a Banach space over C. Let T B(E) Inv B(E) be such that there is a
sequence (T
n
)

n=1
in Inv B(E) such that T = lim
n
T
n
. Show that lim
n=1
|T
1
n
| = .
Exercise 3.9 Let E be a Banach space over C. An element C is called an approximate
eigenvalue for T if
inf|( T)x| : x E, |x| = 1 = 0.
Show that
(T) approximate eigenvalues of T (T).
As Exercise 3.5, there is nothing more that can be said about the spectra of bounded,
linear operators on Banach spaces over C except that they are non-empty subsets of C.
To get more detailed information, we need to look at a smaller class of operators.
57
3.2 Spectral theory for compact operators
In this section, all spaces are over C.
Lemma 3.2.1 Let E be a Banach space, let F be a closed subspace of E, let T /(E),
and let C 0 be such that
inf|( T)x| : x F, |x| = 1 = 0.
Then F ker( E) ,= 0.
Proof Let (x
n
)

n=1
be a sequence in F with |x
n
| = 1 for all n N and ( T)x
n
0.
Since T is compact, (Tx
n
)

n=1
has a convergent subsequence (Tx
n
k
)

k=1
. Hence (x
n
k
)

k=1
converges and since ,= 0, so does (x
n
k
)

k=1
. Let x := lim
k
x
n
k
. Then x F with
|x| = 1 belongs to ker( T). .
Proposition 3.2.2 Let E be a Banach space, let T /(E), and let (T)0. Then
the following hold:
(i) dimker( T) < .
(ii) ( T)E is closed and has nite codimension.
(iii) There is n N such that ker(T)
n
= ker(T)
n+1
and (T)
n
E = (T)
n+1
.
Proof For (i), observe that
id[
ker(T)
=
1

T[
ker(T)
is compact. This implies dimker( T) < .
For (ii) choose that a closed subspace F of E such that E = ker(T) F. It follows
that ( T)F = ( T)E. Since F ker( T) = 0, Lemma 3.2.1 implies that
C := inf|( T)x| : x F, |x| = 1 > 0.
Hence,
|( T)x| C|x| (x F),
so that ( T)F = ( T)E is closed.
To see that ( T)E has nite codimension, note that
( T)E

= E

: (( T)x) = 0 for all x E

= E

: (( T))(x) = 0 for all x E

= ker( T

).
58
Since T

is also compact by Theorem 2.2.4, (i) yields dim(T)E

< , so that (T)E

has nite codimension (since it is closed).
We only prove the rst statement of (iii) in detail (the second one is established
analogously). Assume towards a contradiction that
E
n+1
:= ker( T)
n+1
ker( T)
n
=: E
n
(n N).
For each n N choose x
n
E
n+1
such that |x
n
| = 1 and dist(x
n
, E
n
)
1
2
. Let
n > m 2. Since
Tx
n
Tx
m
= x
n
( T)x
n
. .
E
n
+( T)x
m
x
m
. .
E
m
E
n
,
we have
|Tx
n
Tx
m
|
[[
2
.
Hence, (Tx
n
)

n=1
has no convergent subsequence. To prove the second statement, proceed
similarly, rst noting that
( T)
n
=
n

k=0
_
n
k
_

k
(T)
nk
=
n

n1

k=0
(1)
k
(T)
nk
. .
|(E)
(n N),
so that ( T)
n
E is closed for each n N. .
Lemma 3.2.3 Let E be a Banach space, let T /(E), and let (T) 0. Then
is an eigenvalue of T or of T

.
Proof Suppose that is not an eigenvalue of T. By Lemma 3.2.1, this means that
inf|( T)x| = 0 : x E, |x| = 1 > 0.
As in the proof of Proposition 3.2.2, we conclude that T is injective and has closed
range. Since (T), we have ( T)E E. Choose E

0 such that
(T)E

. Again as in the proof of Proposition 3.2.2, we see that ker(T)

. .
Lemma 3.2.4 Let E be a Banach space, let T /(E), and let (
n
)

n=1
be a sequence of
pairwise distinct eigenvalues of T. Then lim
n

n
= 0.
Proof Without loss of generality suppose that
n
,= 0 for all n N. For each n N,
choose x
n
ker(
n
T) 0. Let
E
n
:= linx
1
, . . . , x
n
,
59
so that E
1
E
n
E
n+1
. For each n 2, choose y
n
E
n
such that
|y
n
| = 1 and dist(y
n
, E
n1
)
1
2
.
Let
1
, . . . ,
n
C be such that y
n
=
1
x
1
+ +
n
x
n
. It follows that
(
n
T)y
n
=
1
(
n

1
)x
1
+ +
n1
(
n

n1
)x
n1
.
For n > m 2, this yields:
T(
1
n
y
n
) T(
1
m
y
m
) =
1
n
(
n
T)y
n

1
m
(
m
T)y
m
+y
m
. .
E
n1
y
n
.
Consequently,
|T(
1
n
y
n
) T(
1
m
y
m
)|
1
2
(n ,= m, n, m 2),
so that (T(
1
n
y
n
))

n=1
has no convergent subsequence. Since T is compact, this means
that (
1
n
y
n
)

n=1
has no bounded subsequence, i.e. |
1
n
y
n
| = [
1
n
[ and thus
n
0.
.
Lemma 3.2.5 Let E be a Banach space, and let (T) 0. Then is an isolated
point of (T).
Proof Assume that there is a sequence (
n
)

n=1
in (T) such that
n
. Without
loss of generality, we may suppose that the
n
s are pairwise distinct. By Lemma 3.2.3,
there is a subsequence (
n
k
)

k=1
of (
n
)

n=1
such that
(a) each
n
k
is an eigenvalue of T, or
(b) each
n
k
is an eigenvalue of T

.
However, (a) contradicts Lemma 3.2.4, and (b) leads equally to a contradiction with
Lemma 3.2.4 if we apply that lemma with T

instead of T. .
Theorem 3.2.6 Let E be a Banach space with dimE = , and let T /(E). Then one
of the following holds:
(i) (T) = 0;
(ii) (T) = 0,
1
, . . . ,
n
, where
1
, . . . ,
n
are eigenvalues of T such that dimker(
j

T) < for j = 1, . . . , n.
(iii) (T) = 0,
1
,
2
, . . . , where
1
,
2
, . . . are eigenvalues of T such that dimker(
n

T) < for n N and lim

n

n
= 0.
60
Proof Since dimE = and T /(E), certainly 0 (T). By Lemma 3.2.5, (T) is at
most countably nite, by Lemma 3.2.4,
n
0 whenever (
n
)

n=1
is a sequence of pairwise
distinct eigenvalues of T, and Proposition 3.2.2(i) ensures that dimker(T) < for each
non-zero eigenvalue . We are thus done once we have shown that every (T) 0
is an eigenvalue.
Let (T) 0. By Proposition 3.2.2(iii), there is n N such that ( E)
n
E =
( E)
n+1
E. Since
Proposition 3.2.2(ii) yields that (T)
n
E is closed and has nite codimension. Assume
that is not an eigenvalue of T, so that T is injective. Consequently, (T)[
(T)
n
E
is bijective. Let S B(( T)
n
E) be the inverse of ( T)[
(T)
n
E
, and dene
P : E E, x S
n
( T)
n
x.
It follows that
P
2
= S
n
( T)
n
S
n
( T)
n
= S
n
( T)
n
= P.
We also have:
( T)P = ( T)S
n
( T)
n
= S
n1
( T)
n
= S
n1
S( T)( T)
n
= S
n
( T)
n+1
= P( T).
Since T is not bijective, but T is assumed to be injective, there is x E(T)E,
so that y := x Px ,= 0. On the other hand, we have
( T)
n
y = P( T)
n
y = ( T)
n
Py = ( T)
n
(Px P
2
x) = 0,
so that ker( T)
n
,= 0 and thus ker( T) ,= 0. .
Exercise 3.10 Let

. Show that M

B(

) is compact if and only if c

0
.
Corollary 3.2.7 (Fredholm alternative) Let E be a Banach space, let T /(E), and
let C 0. Then the following are equivalent:
(i) T is bijective.
(ii) T is injective.
(iii) T is surjective.
61
Proof (i) = (ii), (iii) is trivial.
(ii) = (i): clear by Theorem 3.2.6.
(iii) = (i): Assume that T is not bijective, i.e. (T) = (T

). Hence, is
an eigenvalue of T

. Hence, there are non-zero elements in ker( T

) = ( T)E

.
Consequently, ( T)E ,= E must hold, contradicting the surjectivity of T. .
Example Let C 0, and let k (([0, 1] [0, 1]), and consider, for f, g (([01, ])
the integral equations
f(x)
_
1
0
f(y)k(x, y) dy = g(x) (x [0, 1]) (3.2)
and
f(x)
_
1
0
f(y)k(x, y) dy = 0 (x [0, 1]). (3.3)
Then there are the following alternatives:
(i) (3.2) has a unique solution f (([0, 1]) for each g (([0, 1]); in particular, (3.3)
only has the trivial solution f 0.
(ii) There is g (([0, 1]) such that (3.2) has no solution f (([0, 1]). In this case, (3.3)
has non-trivial solutions f (([0, 1]), which form a nite-dimensional subspace of
(([0, 1]).
Since the Fredholm operator on (([0, 1]) with kernel k is compact, this is an immediate
consequence of Corollary 3.2.7.
To make stronger assertions on the spectral theory of compact operators, we need to
leave the general Banach space framework.
3.3 Hilbert spaces
Hilbert spaces are, in a certain sense, the innite-dimensional spaces which behave most
like nite-dimensional Euclidean space.
3.3.1 Inner products
Denition 3.3.1 A semi-inner product on a vector space E is map [, ] : EE F such
that
(a) [x +y, z] = [x, z] +[y, z] (, F, x, y, z E);
(b) [z, x +y] = [z, x] +[z, y] (, F, x, y, z E);
62
(c) [x, x] 0 (x E);
(d) [x, y] = [y, x] (x, y E).
A semi-inner product is called an inner product if
[x, x] = 0 x = 0 (x E).
Example Let (, S, ) be a measure space. For f, g L
2
(, S, ), dene
[f, g] :=
_

f()g() d(). (3.4)

Then [, ] is a semi-inner product.
Proposition 3.3.2 (CauchySchwarz inequality) Let E be a vector space with a se-
mi-inner product [, ] on it. Then
[[x, y][
2
[x, x][y, y] (x, y E)
holds.
Proof Let x, y E. For all F, we have:
0 [x y, x y] = [x, x] [y, x] [x, y] +[[
2
[y, y]. (3.5)
Choose F with [[ = 1 such that [y, x] = [[y, x][. For t R and = t we obtain
from (3.5):
0 [x, x] t[y, x] t[x, y] +t
2
[y, y]
= [x, x] t[[y, x][ t[[x, y][ +t
2
[y, y]
= [x, x] 2t[[y, x][ +t
2
[y, y]
=: q(t).
Then q is a quadratic polynomial in t with at most one zero in R. Hence, the discriminant
of q must be less than or equal to zero:
0 4[[y, x][
2
4[x, x][y, y].
This yields the claim. .
Corollary 3.3.3 Let E be a vector space with a semi-inner product [, ]. Then
|x| := [x, x]
1
2
(x E)
denes a seminorm on E, which is a norm if and only if [, ] is a inner product.
63
Proof Only the triangle inequality needs proof. For x, y E, we have:
|x +y|
2
= [x +y, x +y]
= [x, x] + [x, y] + [y, x]
. .
=[x,y]
+[y, y]
= [x, x] + 2Re [x, y] + [y, y]
[x, x] + 2[[x, y][ + [y, y]
[x, x] + 2[x, x]
1
2
[y, y]
1
2
+ [y, y], by Proposition 3.3.2,
= |x|
2
+ 2|x||y| +|y|
2
= (|x| +|y|)
2
.
Taking roots yields the triangle inequality. .
Exercise 3.11 Let E be a linear space, and let [, ] be a semi-inner product on E.
(i) Show that F := x E : [x, x] = 0 is a linear subspace of E.
(ii) Show that
x +F, y +F) := [x, y] (x, y E)
denes an inner product on E/F.
Denition 3.3.4 A vector space H equipped with an inner product , ) is called a Hilbert
space if H equipped with the norm
|| := , )
1
2
( H)
is a Banach space.
Example Let (, S, ) be a measure space. Then (3.4) induces a inner product on
L
2
(, S, ) turning it into a Hilbert space. In particular, for each index set I ,= ,
the space

2
(I) :=
_
f : I F :

iI
[f(i)[
2
<
_
equipped with
f, g) :=

iI
f(i)g(i) (f, g
2
(I))
is a Hilbert space.
64
3.3.2 Orthogonality and self-duality
Denition 3.3.5 Let H be a Hilbert space. We say that , H are orthogonal in
symbols: if , ) = 0.
Exercise 3.12 Let H be a Hilbert space, and let
1
, . . . ,
n
H be pairwise orthogonal. Show
that
|
1
+ +
n
|
2
= |
1
|
2
+ +|
n
|
2
.
How do you interpret this geometrically?
Lemma 3.3.6 (parallelogram law) Let H be a Hilbert space. Then we have:
| +|
2
+| |
2
= 2(||
2
+||
2
) (, H).
Proof We have
| +|
2
= +, +) = ||
2
+||
2
+ 2Re , )
and
| |
2
= +, +) = ||
2
+||
2
2Re , ).
Adding both equations yields the claim. .
Theorem 3.3.7 Let K be a closed, convex, non-empty subset of a Hilbert space H. Then,
for each H, there is a unique K such that | | = dist(, K).
Proof Let H, and let := dist(, K), so that there is a sequence (
n
)

n=1
in K such
that |
n
| . Note that, for n, m N, we have
_
_
_
_
1
2
(
n

m
)
_
_
_
_
2
=
_
_
_
_
1
2
[(
n
) (
m
)]
_
_
_
_
2
=
1
2
_
|
n
|
2
+|
m
|
2
_

_
_
_
_
1
2
(
n
+
m
)
. .
K

_
_
_
_
2
. .

2
, by Lemma 3.3.6. (3.6)
Let > 0. Choose N N such that
|
n
|
2
<
2
+
1
4

2
(n N).
65
Then (3.6) yields
_
_
_
_
1
2
(
n

m
)
_
_
_
_
2
<
1
2
_
2
2
+
1
2

2
_

2
=
1
4

2
(n, m N)
and thus
|
n

m
| < (n, m N).
Thus, (
n
)

n=1
is a Cauchy sequence which therefore converges to some K. It is
obvious that | | = .
To prove the uniqueness of , let
1
,
2
K be such that |
j
| = for j = 1, 2.
Since
1
2
(
1
+
2
) K, we have

_
_
_
_

1
2
(
1
+
2
)
_
_
_
_
=
_
_
_
_
1
2
(
1
) +
1
2
(
2
)
_
_
_
_

1
2
(|
1
| +|
2
|) = ,
so that
_
_

1
2
(
1
+
2
)
_
_
= as well. This, in turn, implies that

2
=
_
_
_
_
1
2
(
1
) +
1
2
(
2
)
_
_
_
_
2
= 2
_
_
_
_
1
2
(
1
)
_
_
_
_
2
. .
=
1
4

2
+2
_
_
_
_
1
2
(
2
)
_
_
_
_
2
. .
=
1
4

_
_
_
_
1
2
(
1

2
)
_
_
_
_
2
and thus

2
=
2

_
_
_
_
1
2
(
1

2
)
_
_
_
_
2
.
This means that
1
=
2
. .
Lemma 3.3.8 Let H be a Hilbert space, let K be a closed subspace, and let H. Then
the following are equivalent for K:
(i) | | = dist(, K);
(ii) K, i.e. for each K.
Proof (i) = (ii): Let K. Then
| |
2
| ( + )|
2
= | |
2
2Re , ) +| |
2
(3.7)
holds, so that 2Re , ) | |
2
. Choose F such that [[ = 1 and , ) =
[ , )[. Replacing in (3.7) with t for t R thus yields
2Re , t )
. .
=2t[, )[
t
2
| | (t R),
66
i.e.
2[ , )[ t| |
2
(t > 0).
Letting t 0 yields , ) = 0, i.e. .
(ii) = (i): Let K, so that . It follows that
| |
2
= |( ) + ( )|
2
= | |
2
+| |
2
| |
2
,
which proves the claim. .
As you saw in Exercise 2.18, a closed subspace of a Banach space need not be com-
plemented. This situation is dierent for Hilbert spaces:
Theorem 3.3.9 Let H be a Hilbert space, and let K be a closed subspace of H. Then there
is a unique P B(H) with the following properties:
(i) PH = K;
(ii) P
2
= P;
(iii) ker P = K

:= H : K;
(iv) |P| 1.
This map P is called the orthogonal projection onto K.
Proof For H, dene
P := the unique K such that | | = dist(, K).
It is then clear that P : H H satises (i). By Lemma 3.3.8, this means that
P := the unique K such that K.
This yields immediately that P is linear, is the identity on K, i.e. satises (ii), and also
satises (iii).
Since P K for all H, we have
||
2
= |( P) +P|
2
= | P|
2
+|P|
2
|P|
2
( H),
which yields (iv). .
Exercise 3.13 Let H be a Hilbert space, and let K be a closed subspace of H. Show that H

=
K K

.
67
Corollary 3.3.10 Let H be a Hilbert space, and let K be a closed subspace of H. Then K
is complemented in H.
Theorem 3.3.11 Let H be a Hilbert space, and let H

. Then there is a unique H

such that
() = , ) ( H). (3.8)
Moreover, satises || = ||.
Proof The uniqueness of is clear.
For the proof of existence, we may suppose without loss of generality that || = 1. Let
K := ker , and let P B(H) denote the orthogonal projection onto K. Choose
0
HK.
Then
0
P
0
K and
0
P
0
,= 0, so that
:=

0
P
0
|
0
P
0
|
is well-dened. Dene H

as
: H F, , ),
so that ker = K ker . This means that there is F such that = . Since
[()[ = [, )[ | ||| = || ( H)
by the CauchySchwarz inequality, we have
1 || [( )[ = | |
2
= 1,
which implies [[ = 1. Letting := , we obtain (3.8). .
Corollary 3.3.12 Let H be a Hilbert space, and dene, for each H, a functional

by letting

() := , ) ( H).
Then the map
H H

is a conjugate linear isometry.

As an application of this so-called self-duality of Hilbert spaces, we will now give a
Hilbert space theoretic proof of the RadonNikod ym theorem from measure theory.
68
Denition 3.3.13 Let (, S) be a measurable space, and let and be measures on
(, S). We say that is absolutely continuous with respect to in symbols:
if (N) = 0 for every N S such that (N) = 0.
Example Let (, S, ) be any measure space, and let f : [0, ] be measurable. Then
: S [0, ] dened by
(S) :=
_
S
f() d() (S S)
is absolutely continuous with respect to .
Theorem 3.3.14 (RadonNikod ym theorem) Let (, S) be a measurable space, and
let and be nite measures on (, S) such that . Then there is a non-negative
h L
1
(, S, ) such that
(S) =
_
S
h() d() (S S)
Proof Let := +, and dene L
2
(, S, ; R)

by letting
(f) :=
_

f() d() (f L
2
(, S, )).
By Theorem 3.3.11, there is a unique g L
2
(, S, ) such that
(f) :=
_

f()g() d() (f L
2
(, S, )),
so that
_

(1 g())f() d() =
_

g()f() d() (f L
2
(, S, )).
Let A := : g() < 0. It follows that
0
_
A
g() d() =
_

g()
A
() d() =
_

(1 g())
A
() d() 0,
so that (A) = 0 and hence (A) = 0. Let B := : g() 1. Similarly, we have
0
_
B
(1 g()) d() =
_
B
g() d() 0,
so that (B) = (B) = 0 as well. We may therefore suppose without loss of generality
that g() [0, 1) for all .
For S S and n N, dene f
n
:= (1 +g + +g
n
)
S
. It follows that
_
S
(1 g()
n+1
) d() =
_

(1 g())f
n
() d()
=
_

g()f
n
() d()
=
_
S
(g() +g()
2
+ +g()
n+1
) d().
69
Let h :=

n=1
g
n
. Since g() [0, 1), this series converges; for the same reason, we have
1 g
n+1
1 pointwise. All in all, we have:
(S) =
_
S
d()
= lim
n
_
S
(1 g()
n+1
) d(), by dominated convergence,
= lim
n
_
S
(g() +g()
2
+ +g()
n+1
) d()
=
_
S
h() d(), by monotone convergence.
This completes the proof. .
3.3.3 Orthonormal bases
Denition 3.3.15 Let H be a Hilbert space. A family (e

of vectors in H is called
orthonormal if |e

| = 1 and e

for ,= .
Examples 1. Let H = F
N
, and let e
j
:= (0, . . . , 0, 1, 0, . . . , 0) for j = 1, . . . , N, where
the non-zero entry is in the j-th position. Then (e
j
)
N
j=1
is orthonormal.
2. More generally, let H =
2
(I) for I ,= , and dene, for i I, a vector e
i
: I F by
letting e
i
(j) :=
i,j
for j I. Then (e
j
)
jI
is orthonormal.
3. Let H = L
2
([0, 2]; C). For n N, dene e
n
L
2
([0, 2]) by letting
e
n
(x) :=
1

2
e
inx
(x [0, 2]).
Then, clearly, |e
n
| = 1. For n ,= m, we have
e
n
, e
m
) =
1
2
_
2
0
e
i(nm)x
dx =
1
2i(n m)
e
i(nm)x

2
0
= 0.
Hence, (e
n
)

n=1
is orthonormal.
Lemma 3.3.16 (Bessels inequality) Let H be a Hilbert space, and let (e

be an
orthonormal family in H. Then

[, e

)[
2
||
2
( H)
holds.
70
Proof For any nite number of indices
1
, . . . ,
n
, let :=

n
j=1
, e

j
)e

j
. It follows
that e

j
for j = 1, . . . , n. Consequently, we have
||
2
= ||
2
+
_
_
_
_
_
_
n

j=1
, e

j
)e

j
_
_
_
_
_
_
2
= ||
2
+
n

j=1
[, e

j
)[
2

j=1
[, e

j
)[
2
.
Since
1
, . . . ,
n
were arbitrary, this yields the claim. .
Lemma 3.3.17 Let H be a Hilbert space, and let (e

be an orthonormal family in H.
Then

, e

)e

converges for every H.

Proof By Lemma 3.3.16, the set
_
: [, e

)[
1
n
_
is nite for each n N. Hence, there
are only countably many such that , e

) , = 0. Without loss of generality, we can

therefore suppose that we are dealing with a sequence (e
n
)

n=1
.
For n > m, we have:
_
_
_
_
_
n

k=1
, e
k
)e
k

m

k=1
, e
k
)e
k
_
_
_
_
_
2
=
_
_
_
_
_
n

k=m+1
, e
k
)e
k
_
_
_
_
_
2
=
n

k=m+1
[, e
k
)[
2
.
Let > 0. Since

k=1
[, e
k
)[
2
||
2
< by Lemma 3.3.16, there is N N such that
n

k=m+1
[, e
k
)[
2
<
2
(n, m N),
so that
_
_
_
_
_
n

k=1
, e
k
)e
k

m

k=1
, e
k
)e
k
_
_
_
_
_
< (n, m N).
Hence, (

n
k=1
, e
k
)e
k
)

n=1
is Cauchy and thus converges. .
Theorem 3.3.18 Let H be a Hilbert space, and let (e

be an orthonormal family in H.
Then the following are equivalent:
(i) (e

is maximal.
(ii) If e

for all , then = 0.

(iii) H = lin e

: .
71
(iv) If H, then
=

, e

)e

holds.
(v) If , H, then
, ) =

, e

)e

, )
holds.
(vi) Parsevals identity holds for each H:
||
2
=

[, e

)[
2
.
If (e

satises these conditions, it is called an orthonormal basis of H.

Proof (i) = (ii): Assume that there is H 0 such that e

for all . Then

we may add the vector

||
to the family (e

and thus obtain an orthonormal family

strictly larger than (e

.
(ii) = (iii): Assume that K := lin e

: H. By Corollary 2.1.6, there is

H

0 such that [
K
= 0. By Theorem 3.3.11, there is H such that
() = , ) ( H).
It follows that 0 = (e

) = e

, ), so that e

for all and hence = 0. This,

however, contradicts ,= 0.
(iii) = (iv): Let H, and dene :=

, e

)e

, which is well dened by

Lemma 3.3.17. It follows that
, e

) = , e

, e

)e

, e

) = , e

) , e

) = 0
for any index . Since H = lin e

: , this implies , ) = 0 and thus = 0.

(iv) = (v): Let , H. By (iv), we have
=

, e

)e

and =

, e

)e

.
This implies
, ) =
_

, e

)e

, e

)e

_
=

, e

), e

)e

, e

)
=

, e

)e

, ).
72
(v) = (vi): Let = .
(vi) =(i): Let (f

be an orthonormal family such that (e

is a proper subfamily.
Then there is one f

0
such that f

0
e

for all . It follows that

[f

0
, e

)[
2
= 0 ,= 1 = |f

0
|
2
,
which contradicts (vi). .
Corollary 3.3.19 Let H ,= 0 be a Hilbert space. Then H has an orthonormal basis.
Proof Use Zorns lemma to obtain a maximal orthonormal family in H. .
Exercise 3.14 Let H be a Hilbert space, let K be a closed subspace, and let (e

be an orthonor-
mal basis for K. Show that the orthogonal projection P onto K is given by
P =

, e

)e

( H).
Exercise 3.15 Show that every orthonormal basis for a separable, innite-dimensional Hilbert
space is countably innite.
Lemma 3.3.20 Let (e

and (f

be orthonormal bases for a Hilbert space H. Then

(e

and (f

have the same cardinality.

Proof Let be the cardinality of (e

, and let be the cardinality of (f

.
Case 1: is nite.
In this case, dimH < , and (e

is a Hamel basis. It is easy to see that orthonormal

families are always linearly independent. Hence, must be nite, too. Since (f

spans
H, we have that (f

is also a Hamel basis for H. It follows that = dimH = .

Case 2: is innite.
By the rst case, this means that is also innite. For any index , dene B

:= :
e

, f

) , = 0; by Bessels inequality, B

is countable. It follows that

=

0
= .
Similarly, one sees that . .
Theorem 3.3.21 The following are equivalent for two Hilbert spaces H and K:
(i) Every orthonormal basis of H has the same cardinality as every orthonormal basis
of K.
(ii) There are an orthonormal basis of H and an orthonormal basis of K having the same
cardinality.
73
(iii) There is a surjective operator U : H K such that
U, U) = , ) (, H).
If these conditions are satised, H and K are called unitarily equivalent.
Proof (i) = (ii) is obvious, and (ii) = (i) follows with Lemma 3.3.20.
(ii) = (iii): Let (e

and (f

be orthonormal bases of H and K, respectively, with

the same cardinality, i.e. we may use the same index set. Dene
U : H K,

, e

)f

.
The same argument as in the proof of Lemma 3.3.17 shows that this is well-dened. We
then have for , H:
U, U) =
_

, e

)f

, e

)f

_
=

, e

)e

, )f

, f

)
=

, e

)e

, )
= , ).
In particular, |U| = || holds for each H, so that U is an isometry and thus has
closed range.
Assume that U is not surjective. By Corollary 2.1.6 and Theorem 3.3.11, we can then
nd K 0 such that UH. This means in particular that Ue

= f

for all ,
which is impossible.
(iii) = (ii): Let (e

be an orthonormal basis for H. It follows that (Ue

is
orthonormal in K such that K = Ue

: . Hence, (Ue

is an orthonormal basis for K.

Clearly, (e

and (Ue

have the same cardinality. .

Corollary 3.3.22 Let H ,= 0 be a Hilbert space. Then H is unitarily equivalent to
2
(I)
for an appropriate index set I ,= .
Corollary 3.3.23 Up to unitary equivalence, there is only one separable, innite-di-
mensional Hilbert space.
Exercise 3.16 Show that the Hilbert spaces
2
, L
2
(R), are L
2
([0, 1]) all separable and thus uni-
tarily equivalent.
Exercise 3.17 Let H be a Hilbert space, let T /(H), and let (e
n
)

n=1
be an orthonormal
sequence in H. Show that |Te
n
| 0.
74
3.3.4 Operators on Hilbert spaces
As we say in the previous subsection, Hilbert spaces are essentially dull objects. This
dullness, however, forces their bounded linear operators to be much more tractable than
in a general Banach space setting.
Theorem 3.3.24 Let H and K be Hilbert spaces, and let T B(H, K). Then there is a
unique operator T

B(K, H) such that

T, ) = , T

) ( H, K).
The operator T

is called the adjoint of T.

Exercise 3.18 Let H be a Hilbert space, and let T B(H). Show that ker T = (T

H)

.
Proof For xed K, dene H

by letting
() := T, ) ( H).
By Theorem 3.3.11, there is a unique T

H such that
T, ) = () = , T

) ( H).
It is easy to see that K T

is a bounded, linear operator from K to H. .

Remark Despite the use of the same symbol, the adjoint of T is not to be confused with
the transpose dened earlier for operators between Banach spaces. Nevertheless, in many
ways, taking the adjoint operator behaves very much like taking the transpose.
Examples 1. Let H = F
N
, let K = F
M
, and let T = T
A
for
A =
_

_
a
1,1
, . . . , a
1,N
.
.
.
.
.
.
a
M,1
, , a
M,N
_

_
.
Then T

= T
A
, where
A

=
_

_
a
1,1
, . . . , a
M,1
.
.
.
.
.
.
a
1,N
, , a
M,N
_

_
.
2. Let (, S, ) be a -nite measure space, and let L

(, S, ). Then M

= M

.
Proposition 3.3.25 Let H be a Hilbert space, let T, S B(H), and let , F. Then
we have:
75
(i) (T +S)

= T

+S

;
(ii) (ST)

= T

;
(iii) T

= T;
(iv) |T|
2
= |T

|
2
= |T

T|.
Proof (i), (ii), and (iii), are straightforward.
For (iv), let H be such that || 1. It follows that
|T|
2
= T, T)
= T

T, )
= |T

T|||
|T

T|
|T

||T|
and hence
|T|
2
|T

T| |T

||T|.
It follows, in particular, that |T| |T

|. On the other hand, (iii) yields that |T

|
|T

| = |T|. Hence, |T| = |T

| holds and also |T|

2
= |T

T|. .
Denition 3.3.26 Let H be a Hilbert space, and let T B(H). Then:
(a) T is self-adjoint if T = T

.
(b) T is normal if T

T = TT

.
Exercise 3.19 Let H be a Hilbert space. Show that a projection P B(H) is self-adjoint if and
only if it is an orthogonal projection.
Exercise 3.20 Let H be a Hilbert space, and let T : H H be linear such that
T, ) = , T) (, H).
Show that T is bounded and self-adjoint.
Exercise 3.21 Let H be a Hilbert space over C. Show that N B(H) is normal if and only if
|N| = |N

| for all H.
Exercise 3.22 Let H be a Hilbert space over C, let N B(H) be normal, and let H and
C be such that N = . Show that N

=

.
76
Exercise 3.23 Let H be a Hilbert space over C. For T B(H) dene
Re T :=
1
2
(T +T

) and ImT =
1
2i
(T T

).
Show that N B(H) is normal if and only if Re N and ImN commute.
Proposition 3.3.27 Let H be a C-Hilbert space. Then the following are equivalent for
T B(H):
(i) T is self-adjoint.
(ii) T, ) R for H.
Proof (i) = (ii) is clear because
T, ) = , T) = T, ) ( H).
(ii) = (i): Let , H, and let C. Then
T( +), +) = T, ) +T, ) +T, ) +[[
2
T, )
is real and thus equal to its complex conjugate. This, in turn, implies that
T, ) +T, ) = , T) +, T)
= T

, ) +T

, ).
Letting = 1 and = i, respectively, we obtain:
_
T, ) +T, ) = T

, ) +T

, ),
iT, ) iT, ) = iT

, ) +iT

, ).
Dividing the second of those equations by i, we get
_
T, ) +T, ) = T

, ) +T

, ),
T, ) T, ) = T

, ) +T

, ),
and adding them yields 2T, ) = 2T

, ), i.e. T = T

. .
Proposition 3.3.28 Let H be a Hilbert space, and let T B(H) be self-adjoint. Then
|T| = sup[T, )[ : H, || 1. (3.9)
Proof Let M denote the supremum in (3.9). It is clear that M |T|.
Let , H with ||, || 1. It follows that
T( ), ) = T, ) T, ) T, ) +T, )
= T, ) T, ) , T

) +T, )
= T, ) 2Re T, ) +T, ).
77
Subtraction yields
4Re T, ) = T( +), +) T( ), ).
It follows that
4Re T, ) M(| +|
2
+| |
2
)
= 2M(||
2
+||
2
)
4M.
Choose F with [[ = 1 such that T, ) = [T, )[. Replacing in the previous
argument by yields
[T, )[ = T, ) = T(), ) M.
For any H with || 1, we thus have
|T| = sup[T, )[ : H, || 1 M
and therefore |T| M. .
Corollary 3.3.29 Let H be a C-Hilbert space, and let T B(H) be such that T, ) = 0
for all H. Then T is zero.
Proof First, note that
T

, ) = , T) = T, ) ( H).
Let
Re T :=
1
2
(T +T

) and ImT :=
1
2i
(T T

).
Then Re T and ImT are self-adjoint such that T = Re T + iImT and S, ) = 0 for all
H, where S = Re T or S = ImT. By Proposition 3.3.28, this means Re T = ImT = 0
and thus T = 0. .
Remark Proposition 3.3.28 is false for Hilbert spaces over R (take H = R and T = T
A
,
where A =
_
0 1
1 0
_
).
Exercise 3.24 Let H be a Hilbert space, and let T(H) be the family of all nite-dimensional
subspaces of H. For each K T(H), let P
K
be the orthogonal projection onto K. Show that
(P
K
)
KF(H)
is a net such that
|P
K
T T| 0 and |TP
K
T| 0
for all T /(H).
78
3.4 The spectral theorem for compact, self-adjoint opera-
tors
Throughout this section, all spaces are again over C.
In linear algebra, it is shown that a self-adjoint matrix can be diagonalized. This means
that such a matrix can be completely described once its eigenvalues and its eigenspaces
are known. In this section, we extend this theorem on matrices to compact, self-adjoint
operators on Hilbert space.
Lemma 3.4.1 Let H be a Hilbert space, and let T /(H) be self-adjoint. Then |T| or
|T| is an eigenvalue of T.
Proof By Proposition 3.3.28, there is a sequence (
n
)

n=1
in H such that |
n
| = 1 for
all n N and [T
n
,
n
)[ |T|. Passing to a subsequence, we may suppose that
(T
n
,
n
))

n=1
converges to R. It follows necessarily that [[ = |T|. Note that
|( T)
n
|
2
=
2
2T
n
,
n
) +|T
n
|
2
2
2
2T
n
,
n
)
0.
With Lemma 3.2.1, we conclude that is an eigenvalue of T. .
Corollary 3.4.2 Let H be a Hilbert space, and let T /(H) be self-adjoint such that
(T) = 0. Then T = 0.
Lemma 3.4.3 Let H be a Hilbert space, let N /(H) be normal, and let ,= be
eigenvalues of N. Then ker( N) ker( N).
Proof Let ker( N) and ker( N), and note that
, ) = , ) = N, ) = , N

) = , ) = , ).
It follows that , ) = 0. .
Lemma 3.4.4 Let H be a Hilbert space, and let T /(H) be self-adjoint. Then (T) R.
Proof Let (T) 0, so that is an eigenvalue of T. Hence, there is H 0
such that T = . It follows that T = T

= and thus ker( T) ker( T).

By Lemma 3.4.3, this is possible only if = , i.e. R. .
Theorem 3.4.5 (spectral theorem for compact, self-adjoint operators) Let H be
a Hilbert space, let T /(H) be self-adjoint, let
1
,
2
, . . . be the distinct, non-zero
eigenvalues of T, and let P
n
denote the orthogonal projection onto ker(
n
T). Then the
following hold true:
79
(i)
1
,
2
, . . . R;
(ii) P
n
P
m
= P
m
P
n
= 0 (n ,= m);
(iii) T =

n=1

n
P
n
.
Proof (i) follows from Lemma 3.4.4.
For (ii), let H, and note that, by Lemma 3.4.3, for n ,= m
P
n
ker(
n
T) ker(
m
T)

= ker P
m
and thus P
m
P
n
= 0. It follows that P
m
P
n
= 0.
(iii): Choose an eigenvalue
1
of T such that [
1
[ = |T| (this is possible by Lemma
3.4.1). Let H
1
:= ker(
1
T) and let P
1
denote the orthogonal projection onto H
1
. Let
K
2
:= H

1
(= ker P
1
). Let K
2
and H
1
and note that
T, ) = , T) = ,
1
) =
1
, ) = 0.
It follows that TK
2
K
2
. If T[
K
2
= 0, nish. Otherwise since T
2
:= T[
K
2
is also
compact and self-adjoint, there is an eigenvalue
2
of T
2
such that [[ = |T
2
|. Let
H
2
:= ker(
2
T
2
). It is easy to see that H
2
= ker(
2
T). Since H
1
H
2
by denition,
we have
1
,=
2
. Let P
2
be the orthogonal projection onto H
2
, and let K
3
:= (H
1
H
2
)

.
Continue inductively and obtain:
(a) A (possibly nite) sequence
1
,
2
, . . . of distinct eigenvalues of T which satises
[
1
[ [
2
[ .
(b) A sequence of pairwise orthogonal closed subspaces H
n
of H such that
H
n
= ker(
n
T) and [
n+1
[ = |T[
(H
1
H
n
)
|.
Fix n N, and let H
k
with k 1, . . . , n. It follows that
T
n

j=1

j
P
j
=
k

k
= 0.
Let (H
1
H
n
)

. Then P
k
= 0 for k = 1, . . . , n and thus
T
n

j=1

j
P
j
= T (H
1
H
n
)

.
80
It follows that
_
_
_
_
_
_
T
n

j=1

j
P
j
_
_
_
_
_
_
= sup
_
_
_
_
_
_
_
_
_
T
n

j=1

j
P
j

_
_
_
_
_
_
: H, || 1
_
_
_
= sup
_
_
_
_
_
_
_
_
_
T
n

j=1

j
P
j

_
_
_
_
_
_
: (H
1
H
n
)

, || 1
_
_
_
= sup
_
|T| : (H
1
H
n
)

, || 1
_
= [
n+1
[
0
and thus
T =

n=1

n
P
n
.
Assume that T has an eigenvalue which does not occur in the sequence
1
,
2
, . . . .
Since ker( T) ker(
n
T) for n N, this means that
T =

n=1

n
P
n
= 0,
i.e. = 0. .
With the exception of (i), the assertions of Theorem 3.4.5 still hold for compact, normal
operators.
Exercise 3.25 Read paragraphs II.6 and II.7 in Conways book (where the spectral theorem for
compact, normal operators is proven).
Remark The spectral theorem generalizes to arbitrary, not necessarily compact, bounded,
normal operators on Hilbert spaces. Since such operators need no longer have eigenvalues,
the projections onto the eigenspaces have to be replaces by a more general object: Given
a Hilbert space H and a normal operator N B(H), there is a unique so-called spectral
measure E a measure whose values are orthogonal projections on H on (N) such
that
N =
_
(N)
z dE(z).
81
Chapter 4
Fixed point theorems and locally
convex spaces
This last chapter of the lecture notes deals with xed point theorems, i.e. theorems that
guarantee that a certain map of a certain family of maps has a xed point. Fixed point
theorems are important because many problems concerning the solvability of equations
can be formulated as xed point problems. In order to present these xed point theorems
in sucient generality, we develop the theory of locally convex vector spaces to some
extent.
4.1 Banachs xed point theorem
Banachs xed point theorem is one of the most elegant and most widely applicable the-
orems in all of analysis:
Theorem 4.1.1 (Banachs xed point theorem) Let X be a complete metric space,
and let T : X X be a map such that, for some (0, 1),
d(T(x), T(y)) d(x, y) (x, y X).
Then T has a unique xed point in X.
Proof Let x, y X be xed points of X. Since
d(x, y) = d(T(x), T(y)) d(x, y)
and (0, 1), it follows that d(x, y) = 0, i.e. x = y. This proves the uniqueness of the
xed point.
For the proof of existence, choose x
0
X arbitrary, and dene inductively x
n
:=
T(x
n1
) for n N. We claim that the sequence (x
n
)

n=0
converges. An easy induction on
82
n shows that
d(x
n1
, x
n
)
n1
d(x
0
, x
1
) (n N).
For n > m, we have
d(x
m
, x
n
)
n

k=m+1
d(x
k1
, x
k
) d(x
0
, x
1
)
n

k=m+1

k1
. (4.1)
Since (0, 1), the geometric series

n=1

n1
converges. Hence, given > 0, there is
N N such that
n

k=m+1

k1
<

d(x
0
, x
1
) + 1
(n, m N).
Together with (4.1), this shows that d(x
m
, x
n
) < for all n, m N. Hence, (x
n
)

n=0
is a
Cauchy sequence, and x := lim
n
x
n
exists. Since T is clearly (uniformly) continuous,
we have
Tx = lim
n
Tx
n
= lim
n
Tx
n1
= lim
n
x
n
= x,
i.e. x is a xed point of T. .
Remark Banachs xed point theorem not only guarantees that a xed point exsists it
also provides an eective way of computing such a xed point along with error estimates.
Exercise 4.1 Is the following xed point theorem true or not?
Let X be a complete metric space, and let T : X X be a map such that
d(T(x), T(y)) < d(x, y) (x, y X, x ,= y).
Then T has a unique xed point in X.
Give a proof or a counterexample.
Exercise 4.2 Let (X, d) be a complete metric space, and let T : X X be a map such that there
are (0, 1) and n N with
d(T
n
(x), T
n
(y)) d(x, y) (x, y X).
Show that T has a unique xed point.
We apply Banachs xed point theorem to initial value problems:
83
Theorem 4.1.2 (PicardLindelof theorem) Let I = [a, b], and let f : I R R be
continuous such that there is C 0 with
[f(x, y
1
) f(x, y
2
)[ C[y
1
y
2
[ (x I, y
1
, y
2
R).
Then the initial value problem
y
t
= f(x, y), y(a) = y
0
(4.2)
has a unique solution (
1
(I) for each y
0
R.
Proof If (4.2) has a solution , it satises
(x) =
_
x
a
f(t, (t)) dt +y
0
. (4.3)
Conversely, every function ((I) satisfying (4.3) is automatically in (
1
(I) and a solution
of (4.2).
Dene T : ((I) ((I) by letting
(T)(x)
_
x
a
f(t, (t)) dt +y
0
.
Then ((I) solves (4.3) and hence (4.2) if and only if it is a xed point of T.
Let
1
,
2
((I). Then we have
[(T
1
)(x) (T
2
)(x)[ =

_
x
a
(f(t,
1
(t)) f(t,
2
(t))) dt

C
_
x
a
[
1
(t)
2
(t)[ dt
= C(b a)|
1

2
|

(x I),
and therefore
|T
1
T
2
|

C(b a)|
1

2
|

.
The problem that arises at this point is that C(b a) may not belong to (0, 1), so that
Theorem 4.1.1 is not directly applicable. We circumvent this problem by replacing | |

by an equivalent norm | |

for some parameter > 0. Dene for > 0

||

:= sup[(x)[e
x
: x I ( ((I)).
Then, for ((I),
||

||

supe
x
: x I
84
and
||

= sup[(x)[e
x
e
x
: x I ||

supe
x
: x I,
so that | |

and | |

are equivalent. Moreover, we have for

1
,
2
((I) and x I:
[(T
1
)(x) (T
2
)(x)[e
x
= e
x

_
x
a
(f(t,
1
(t)) f(t,
2
(t))) dt

Ce
x
_
x
a
[
1
(t)
2
(t)[ dt
= Ce
x
_
x
a
[
1
(t)
2
(t)[e
t
e
t
dt
Ce
x
|
1

2
|

_
x
a
e
t
dt
Ce
x
|
1

2
|

e
x

=
C

|
1

2
|

.
It follows that
|T
1
T
2
|

|
1

2
|

(
1
,
2
((I)).
Choosing > 0 so large that
C

< 1 and applying Banachs xed point theorem, we obtain

a unique xed point of T and thus a unique solution of (4.2). .
4.2 Locally convex vector spaces
The next xed point theorem we are going to cover Schauders xed point theorem
requires a compactness hypothesis for its domain. Since compactness in normed, innite-
dimensional spaces is rather the exception than the rule, we have to leave the framework of
normed spaces and work in a more general context in order to obtain xed point theorems
of sucient generality.
Denition 4.2.1 A linear space E is called locally convex if it is equipped with a family
T of seminorms on E such that

p1
x E : p(x) = 0 = 0.
Example Let X be a topological space, and let ((X) denote the vector space of all
continuous functions on X. Let / be the collection of all compact subsets of X. For
K /, dene
p
K
(f) := sup[f(x)[ : x K (f ((X)).
Then ((X) equipped with (p
K
)
K|
is a locally convex vector space.
85
Denition 4.2.2 Let E be a locally convex vector space. A subset U of E is dened as
open if, for each x
0
U, there are > 0 and p
1
, . . . , p
n
T such that
_
x E : max
j=1,... ,n
p
j
(x x
0
) <
_
U.
Proposition 4.2.3 Let E be a locally convex vector space. Then the collection of open
subsets of E in the sense of Denition 4.2.2 is a topology, i.e.
(i) and E are open;
(ii) if (U

is a family of open sets, then

is open;
(iii) if U
1
, . . . , U
n
are open, then so is U
1
U
n
.
Proof We only prove (iii).
Let x
0
U
1
U
n
. For each k = 1, . . . , n, there are
k
> 0 and p
(k)
1
, . . . , p
(k)
n
k
T
such that
V
k
:=
_
x E : max
j=1,... ,n
k
p
(k)
j
(x x
0
) <
k
_
U
k
.
Let := min
1
, . . . ,
k
, and note that
_
x E : max
k=1,... ,n
max
j=1,... ,n
k
p
(k)
j
(x x
0
) <
_
V
1
V
n
U
1
U
n
.
By Denition 4.2.2, this means that U
1
U
n
is open. .
Proposition 4.2.4 Let E be a locally convex vector space. Then a net (x

in E con-
verges to x
0
E in the topology if and only if p(x

x
0
) 0 for each p T.
Proof Suppose that x

x
0
in the topology. Fix > 0 and p T. Then U := x
E : p(x x
0
) < is an open neighborhood of x
0
. Hence, there is an index
0
such that
x

U, i.e. p(x

x
0
) < for all ~
0
. Hence, p(x

x
0
) 0.
Conversely, suppose that p(x

x
0
) 0 for all p T. Let U be a neighborhood of
x
0
, i.e. there is an open set V U with x
0
V . By Denition 4.2.2, there are > 0 and
p
1
, . . . , p
n
T such that
_
x E : max
j=1,... ,n
p
j
(x x
0
) <
_
V.
Since p
j
(x

x
0
) 0 for j = 1, . . . , n, there is an index
0
such that
p
j
(x

x
0
) < (j = 1, . . . , n, ~
0
).
This means, however, that x

V U for all ~
0
. .
Exercise 4.3 Let E be a locally convex vector space, and let F be a nite-dimensional subspace.
Show that the relative topology on F is induced by a norm.
86
4.2.1 Weak and weak

topologies
There is a canonical locally convex topology on each normed space:
Denition 4.2.5 Let E be a normed space. Then p

: E

with
p

(x) = [(x)[ (x E, E

)
is a family of seminorms on E such that

E
x E : p

(x) = 0. The corresponding

topology on E is called the weak topology on E.
Lemma 4.2.6 Let E be a linear space, and let ,
1
, . . . ,
n
: E F be linear. Then the
following are equivalent:
(i) there are
1
, . . . ,
n
F such that =
1

1
+ +
n

n
;
(ii)

n
j=1
ker
j
ker .
Exercise 4.4 Prove Lemma 4.2.6.
Theorem 4.2.7 Let E be a normed space. Then the following are equivalent:
(i) dimE < ;
(ii) the weak topology and the norm topology coincide;
(iii) the weak topology is metrizable.
Proof (i) = (ii): If dimE < , then dimE

< . Let
1
, . . . ,
n
E

be a Hamel
basis for E

. Dene
[x[ := max
j=1,... ,n
[
j
(x)[ (x E).
Then E is a norm on E, so that [ [ | |. Let U E be norm open. This means that,
for every x
0
U, there is > 0 such that
x E : [x x
0
[ < =
_
x E : max
j=1,... ,n
[
j
(x x
0
)[ <
_
U.
From the denition of the weak topology, it follows that U is also weakly open. Since the
weak topology is coarser than the norm topology, every weakly open subset of E is norm
open.
(ii) = (iii) is trivial.
(iii) = (i): Suppose that there is a metric d on E which induces the weak topology.
Hence, for all n N,
U
n
:=
_
x E : d(x, 0) <
1
n
_
87
is weakly open. By the denition of the weak topology, there are, for each n N, a
number
n
> 0 and functionals
(n)
1
, . . . ,
(n)
k
n
E

such that
_
x E : max
j=1,... ,k
n

(n)
j
(x x
0
)

<
n
_
U
n
.
Let E

. Then is continuous with respect to the weak topology, i.e. if (x

is
a net in E with x

weakly
0, then (x

) 0. Assume that is unbounded on each U

n
.
This means that, for each n N, an element x
n
U
n
with [(x
n
)[ n. It follows that
d(x
n
, 0)
1
n
0 and thus x
n
weakly
0, whereas (x
n
) , 0. It follows that there is N N
such that sup[(x)[ : x U
N
< . Since
k
N

j=1
ker
(N)
j

_
x E : max
j=1,... ,k
N

(N)
j
(x)

<
n
_
U
N
,
the functional must be bounded on

k
N
j=1
ker
(N)
j
, which, in turn, is possible only if

k
N
j=1
ker
(N)
j
ker . By Lemma 4.2.6, there are thus
1
, . . . ,
k
N
F such that =

(N)
1
+ +
k
N

(N)
k
N
.
It follows that E

is the linear span of the countable set

_

(n)
j
: n N, j = 1, . . . , k
n
_
.
Hence, E

has a countable Hamel basis. Since E

is always a Banach space, this means

that dimE

< and, consequently, dimE < . .

In analogy with the weak topology, one denes an even weaker topology on the dual
of a normed space:
Denition 4.2.8 Let E be a normed space. Then p
x
: x E with
p
x
() = [(x)[ ( E

, x E)
is a family of seminorms on E

such that

xE
E

: p
x
() = 0. The corresponding
topology on E

is called the weak

topology on E

.
Exercise 4.5 Let E be a separable Banach space. Show that the relative topology of the weak

-
topology on the closed unit ball of E

is metrizable. (Hint: Let x

n
: n N be a countable dense
subset of E, and dene
d(, ) :=

n=1
1
2
n
[(x
n
) (x
n
)[
[(x
n
) (x
n
)[ + 1
(, E

).
Show that d is a metric on E

which, on the closed unit ball of E

, denes the weak

-topology.)
Exercise 4.6 Let E be a Banach space. Show that the following are equivalent:
(a) dimE < ;
88
(b) the weak

-topology and the norm topology coincide;

(c) the weak

-topology is metrizable.
How does this go together with Exercise 4.5?
Theorem 4.2.9 (AlaogluBourbaki theorem) Let E be a normed space. Then the
closed unit ball of E

is compact in the weak

topology on E

.
Proof For each x E, let
K
x
:= F : [[ |x|.
Since each K
x
is closed and bounded, it is compact. By Tychonos theorem (Theorem
A.7.4),

xE
K
x
is compact in the product topology. Embed the closed unit ball of E

into

xE
K
x
via
B
1
[0]

xE
K
x
, ((x))
xE
.
Let (

be a net in the closed unit ball of E

; we will show that it has a conver-

gent subnet. By Theorem A.6.6, the net ((

(x))
xE
)

has a subnet ((

(x))
xE
)

that
converges in the product topology, i.e. for each x E, there is
x
K
x
such that

x
= lim

(x).
Dene : E F by letting (x) :=
x
for x E. For x, y E and F, we have
(x +y) =
x+y
= lim

(x +y) = lim

(x) + lim

(y) =
x
+
y
= (x) +(y)
and
(x) =
x
= lim

(x) = lim

=
x
= (x).
Hence, is linear. Moreover, note that, for x E with |x| 1,
[(x)[ = [
x
[ |x| 1
because
x
K
x
. It follows that E

lies in the closed unit ball. From the denition

of the weak

topology, it is clear that

weak

. Theorem A.6.6 eventually yields the

weak

compactness of the closed unit ball of E

. .
Exercise 4.7 Let E be a normed space. Show that there is a compact Hausdor space X and an
isometry : E ((X).
Exercise 4.8 A Banach space E is called reexive if the canonical map J : E E

is an
isomorphism.
89
(i) Let (, S, ) be a measure space, and let p (1, ). Show that L
p
(, S, ) is reexive.
(ii) Conclude that every Hilbert space is reexive.
(iii) Argue that

is not reexive.
Exercise 4.9 Let E be a reexive Banach space. Show that the closed unit ball of E is compact
in the weak topology.
4.3 Schauders xed point theorem
We need the following fact from algebraic topology:
Theorem 4.3.1 There is no continuous map from the closed unit ball of R
N
to S
N1
:=
x R
N
: |x|
2
= 1 whose restriction to S
N1
is the identity.
Theorem 4.3.2 (Brouwers xed point theorem) Let K := x R
N
: |x|
2
1,
and let f : K K be continuous. Then f has a xed point in K.
Proof Assume that the theorem is false, i.e. f(x) ,= x for all x K. Dene : K R
N
by letting
(x) := the unique intersection point of the line from f(x) through x with S
N1
.
Then is continuous with (K) S
N1
and [
S
N1 = id, which is impossible by Theorem
4.3.1. .
Corollary 4.3.3 Let E be a nite-dimensional normed space, let ,= K E be compact
and convex, and let f : K K be continuous. Then f has a xed point in K.
Proof Without loss of generality, let E = R
N
. Choose r > 0 such that
K x R
N
: |x|
2
r =: B.
Dene : B K by letting
(x) := the unique point y K such that |x y|
2
= dist(x, K) (x B).
Then is continuous, and (x) = x for x K. Hence, f is continuous and maps B
into B. By Theorem 4.3.2, there is x B such that f((x)) = x. Since (B) K, we
have x K, so that f(x) = f((x)) = x. .
Exercise 4.10 Use the intermediate value theorem to prove Corollary 4.3.3 in the one-dimensional
case: If f : [a, b] [a, b] is continuous, then f has a xed point.
90
Schauders xed point theorem is the innite-dimensional generalization of Corollary
4.3.3.
Denition 4.3.4 A subset S of a linear space E is called balanced if x : x S,
F, [[ 1 S.
Denition 4.3.5 Let E be a linear space, and let ,= K E be convex and balanced.
Then the Minkowski functional
K
of K is dened by

K
(x) := inft > 0 : x tK (x E).
Proposition 4.3.6 Let E be a locally convex vector space, and let ,= U E be open,
convex, and balanced. Then
U
is a continuous seminorm on E such that
U = x E :
U
(x) < 1. (4.4)
Proof Since U is balanced, we have 0 U. Let x E. Since
1
n
x 0, and since U is
a neighborhood of 0, there is n N such that
1
n
x U. Hence,
U
(x) < . Clearly,

U
(0) = 0. Let x E, and F 0. We have that
x tU x t
1
U
x t
1

[[
U, since

[[
U = U,
x t[
1
[U (t > 0),
and hence

U
(x) = inft > 0 : x t[
1
[U = [[ inft > 0 : x tU = [[
U
(x).
Let x, y E, and let > 0. Choose t, s > 0 such that x tU, y sU, t <
U
(x) +

2
, and
s <
U
(y) +

2
. It follows that
x
t +s

t
t +s
U and
y
t +s

s
t +s
U
and therefore
x +y
t +s

t
t +s
U +
s
t +s
U U
because U is convex. It follows that

U
(x +y) t +s
U
(x) +

2
+
U
(y) +

2
=
U
(x) +
U
(y) +
and therefore

U
(x +y)
U
(x) +
U
(y).
91
All in all,
U
is a seminorm.
Let x U. Since
(0, ) E, t t
1
x
is continuous, and since U is open, there is t (0, 1) such that t
1
x U. It follows that

U
(x) < 1. Since U is balanced,
U
(x) 1 holds trivially for x / U. This proves (4.4).
Since U is open and contains 0, there are > 0 and p
1
, . . . , p
n
T such that
_
x E : max
j=1,... ,n
p
j
(x)
_
U.
Let (x

be a net in E such that x

0, i.e. p
j
(x

) 0 for j = 1, . . . , n. Let > 0.

Then there is
0
such that p
j
(x

) < for j = 1, . . . , n and ~

0
. This means that
x

U for all ~
0
and thus
U
(x

) < for all ~

0
. It it follows that
U
(x

) 0.
Finally, let (x

be a net in E such that x

x E. Since x

x 0, we have
[
U
(x

)
U
(x)[
U
(x

x) 0.
Hence,
U
is continuous. .
Lemma 4.3.7 Let E be a locally convex vector space, let ,= K E be compact, let
f : K K be continuous and suppose that f(x) ,= x for all x K. Then there is an
open, convex, balanced set W such that
((x, f(x)) : x K +W W) (x, x) : x E = .
Proof Let
Gr f := (x, f(x)) : x K and := (x, x) : x E.
Let (x, f(x)) Gr f. Then there are open subset U and V of E with x U, f(x) V ,
and (U V ) = . Without loss of generality, we may suppose that there are > 0
and p
1
, . . . , p
n
, q
1
, . . . , q
m
T such that
U =
_
y E : max
j=1,... ,n
p
j
(x y) <
_
and
V =
_
y E : max
j=1,... ,m
q
j
(f(x) y) <
_
.
Dene
W
x
:=
_
y E : max
j=1,... ,n
k=1,... ,m
p
j
(y), q
k
(y) <
_
.
92
Then W
x
is an open, convex, balanced set with ((x, f(x)) +W
x
W
x
) = . Since K
is compact, there are x
1
, . . . , x
k
K such that
K
k
_
j=1
((x
j
, f(x
j
)) +W
x
j
W
x
j
).
Dene W :=

k
j=1
W
x
j
. .
Exercise 4.11 Let E be a linear space, and let S be a non-empty subset of E. The convex hull
conv S of S is dened as the intersection of all convex subsets of E containing S. Show that conv S
consists of all elements of E of the form

n
j=1
t
j
s
j
, where n N, s
1
, . . . , s
n
S, and t
1
, . . . , t
n
> 0
with

n
j=1
t
j
= 1.
Exercise 4.12 Let E be a locally convex vector space, and let x
1
, . . . , x
n
E. Show that the
convex hull of x
1
, . . . , x
n
is compact.
Theorem 4.3.8 (Schauders xed point theorem) Let E be a locally convex space,
let ,= K E be compact, and let f : K K be continuous. Then f has a xed point
in K.
Proof Assume that the theorem is false, i.e. the hypotheses of Lemma 4.3.7 are satised.
Choose W as specied in Lemma 4.3.7, i.e., in particular,
f(x) / x +W (x K). (4.5)
By Proposition 4.3.6,
W
is a continuous seminorm on E such that W = x E :

W
(x) < 1. Dene
: E R, x max0, 1
W
(x).
Choose x
1
, . . . , x
n
K such that K

n
j=1
x
j
+W. For j = 1, . . . , n, dene
j
: E R
and
j
: K R by letting

j
(x) := (x x
j
) (x E, j = 1, . . . n)
and

j
(x) :=

j
(x)

1
(x) + +
n
(x)
(x K, j = 1, . . . n).
Let F be the linear span of x
1
, . . . , x
n
, and let C := convx
1
, . . . , x
n
. Then C K is
a compact, convex subset of the nite-dimensional (normed) space F. Dene
g : K H, x
n

j=1

j
(x)x
j
.
93
Since g f maps H into itself, Corollary 4.3.3 yields x
0
H such that g(f(x
0
)) = x
0
.
Since
j
(x) = 0 for x / x
j
+W, we have
x g(x) =
n

j=1

j
(x)(x x
j
) W (x K).
In particular,
f(x
0
) x
0
= f(x
0
) g(f(x
0
)) W
holds. This, however, contradicts (4.5). .
Exercise 4.13 Let B be the closed unit ball in
2
, and dene f : B
2
by letting, for x =
(x
n
)

n=1
,
f(x) = ((1 |x|
2
), x
1
, x
2
, . . . ).
Show that f is continuous with f(B) B, but has no xed point.
4.3.1 Peanos theorem
Like Banachs xed point theorem, Schauders xed point theorem can be applied to initial
value problems:
Lemma 4.3.9 (Mazurs theorem) Let E be a Banach space, and let K E be com-
pact. Then conv K is also compact.
Proof Let > 0. Choose x
1
, . . . , x
n
K such that
K
n
_
j=1
B

3
(x
j
).
Let C := convx
1
, . . . , x
n
. Since C is compact, there are y
1
, . . . , y
m
C such that
C
m
_
j=1
B

3
(y
j
).
Let z conv K. Then there is w conv K such that |w z| <

3
. Let t
1
, . . . , t
k
> 0
with t
1
+ +t
k
= 1 and v
1
, . . . , v
k
K such that w =

k
j=1
t
j
v
j
. For each j = 1, . . . , k,
there is (j) 1, . . . , n such that |v
j
x
(j)
| <

3
. It follows that
_
_
_
_
_
_
w
k

j=1
t
j
x
(j)
_
_
_
_
_
_
=
_
_
_
_
_
_
k

j=1
t
j
(v
j
x
(j)
)
_
_
_
_
_
_

j=1
t
j
|v
j
x
(j)
|
<

3
.
94
Since

k
j=1
t
j
x
(j)
C, there is j
0
1, . . . , m such that
_
_
_
_
_
_
k

j=1
t
j
x
(j)
y
j
0
_
_
_
_
_
_
<

3
.
All in all, we have that
|z y
j
0
| |z w| +
_
_
_
_
_
_
w
k

j=1
t
j
x
(j)
_
_
_
_
_
_
+
_
_
_
_
_
_
k

j=1
t
j
x
(j)
y
j
0
_
_
_
_
_
_
<

3
+

3
+

3
= .
Since z conv K was arbitrary, this means that
conv K
m
_
j=1
B

(y
j
),
so that conv K is totally bounded and thus compact. .
Theorem 4.3.10 (Peanos theorem) Let I = [a, b], and let f (
b
(I R). Then the
initial value problem
y
t
= f(x, y), y(a) = y
0
has a solution (
1
(I).
Proof As in the proof of Theorem 4.1.2, we need to nd ((I) such that
(x) =
_
x
a
f(t, (t)) d +y
0
. (4.6)
For the sake of simplicity, suppose that I = [0, 1], y
0
= 0, and |f|

1. Again as in the
proof of Theorem 4.1.2, dene T : ((I) ((I) by letting
(T)(x) :=
_
x
0
f(t, (t)) dt ( ((I), x I).
Let B := ((I) : ||

1. Then T maps B into itself. Note that

[(T)(y) (T)(x)[
_
y
x
[f(t, (t))[ dt y x (y x).
The family T : B is therefore equicontinuous and thus relatively compact. Let
K := conv TB. By Lemma 4.3.9, K is compact and clearly TK K. By Theorem 4.3.8,
T has a xed point in K, i.e. a solution of (4.6). .
Exercise 4.14 Give two dierent solutions of the initial value problem
y

=

y, y(0) = 0.
95
4.3.2 Lomonosovs theorem
We now give a second application of Schauders xed point theorem to a famous open
problem in operator theory.
Denition 4.3.11 Let E be a Banach space, and let T B(E). A closed subspace F of
E is called invariant for T if
(a) 0 F E, and
(b) TF F.
If F is invariant for every S B(E) commuting with T, it is called hyperinvariant
The invariant subspace problem posed by J. von Neumann is the following
question:
Let H be a (separable, innite-dimensional) Hilbert space over C, and let T
B(H). Does T have an invariant subspace?
For operators on Banach spaces, the answer is negative: Counterexamples have been
constructed by P. Eno and C. J. Read. Reads construction even yields an operator
T B(
1
) without invariant subspace.
Exercise 4.15 Let H be a Hilbert space such that either 2 dimH < or that H is not
separable. Show that every bounded linear operator on H has an invariant subspace.
Theorem 4.3.12 (Lomonosovs theorem) Let E be an innite-dimensional Banach
space over C, and let T /(E) 0. Then T has a hyperinvariant subspace.
Proof Assume towards a contradiction that the claim is wrong. Without loss of generality
suppose that |T| = 1. Fix x
0
E with |Tx
0
| > 1, and let B := x E : |x x
0
| 1,
so that
0 / B and 0 / TB.
For any x E 0 dene
F
x
:= Sx : S B(E), ST = TS.
Then F
x
,= 0 is a closed subspace of E with SF
x
F
x
for all S B(E) with ST = TS.
By assumption, this means that F
x
= E for all x E0. Hence, for each y TB, there
is S
y
B(E) commuting with T such that |S
y
y x
0
| < 1. Since K := TB is compact,
there are S
1
, . . . , S
n
B(E) commuting with T such that
K
n
_
j=1
y E : |S
j
y x
0
| < 1.
96
For y K and j = 1, . . . , n, let

j
(y) := max0, 1 |S
j
y x
0
|.
For j = 1, . . . , n, dene

j
: K R, y

j
(y)

1
(y) + +
n
(y)
,
and let
f : B E, x
n

j=1

j
(Tx)S
j
Tx,
so that f is continuous. Let x B, so that Tx K. If
j
(Tx) > 0, then
j
(Tx) > 0
and therefore |S
j
Tx x
0
| < 1, i.e. S
j
Tx B. The convexity of B yields f(B) B. The
compactness of T yields that f(B) is compact. Let C := conv f(B). By Lemma 4.3.9, C
is compact, and clearly f(C) C. By Theorem 4.3.8, there is a xed point y
0
of f in
C B. Let
R :=
n

j=1

j
(Ty
0
)S
j
.
Then T commutes with T and satises RTy
0
= f(y
0
) = y
0
. Since y
0
B E 0, this
means that F
0
:= ker(RT id
E
) ,= 0. Since RT /(E), we also have that dimF
0
<
and thus F
0
E. Clearly, F
0
is invariant form T. Let C be an eigenvalue of T[
F
0
, so
that F := ker( T) ,= 0. If = 0, then F = ker T ,= E, since T ,= 0. If ,= 0, then
dimF < , so that F ,= E as well in this case. Clearly, F is hyperinvariant for T. .
4.4 The MarkoKakutani xed point theorem
We present yet another xed point theorem, this time not for a single function, but for a
whole family of maps.
Denition 4.4.1 Let K be a convex subset of a linear space. A map T : K K is called
ane if
T(tx + (1 t)y) = tTx + (1 t)Ty (x, y K, t [0, 1]).
Theorem 4.4.2 (MarkoKakutani xed point theorem) Let E be a locally con-
vex vector space, let ,= K E be compact and convex, and let o be an abelian semigroup
of continuous ane maps on K, i.e.
(a) each S o is a continuous ane map on K,
97
(b) ST o for all S, T o, and
(c) ST = TS for all S, T o.
Then there is x
0
K such that Sx
0
= x
0
for all S o.
Proof For n N and S o, let
S
n
:=
1
n
n1

k=0
S
k
.
It is easy to see that
S
n
T
m
= T
m
S
n
(n, m N, S, T o).
Let
/ := S
n
K : n N, S o.
Then / consists of non-empty, compact, convex subsets of K. Let n
1
, . . . , n
k
N and
S
(1)
, . . . , S
(k)
o and note that
_
S
(1)
n
1
S
(k)
n
k
_
(K)
. .
,=

j=1
S
(j)
n
j
K.
Hence, every nite family of sets in / has non-empty intersection. Since K is compact,
this means that

S
n
K : n N, S o ,=
Let x
0
be any point in this intersection. Fix S o and n N. Then there is x K such
that
x
0
= S
n
x =
1
n
(x +Sx + +S
n1
x).
It follows that
Sx
0
x
0
=
1
n
(Sx +S
2
x + +S
n
x)
1
n
(x +Sx + +S
n1
x)
=
1
n
(S
n
x x)

1
n
(K K).
Let > 0, and let p
1
, . . . , p
n
T. Let
U :=
_
x E : max
j=1,... ,n
p
j
(x) <
_
.
Since K is compact, so is K K. Hence, there is N N such that
1
n
(K K) U for all
n N. It follows that lim
n
(Sx
0
x
0
) = 0, i.e. Sx
0
= x
0
. .
98
Denition 4.4.3 Let G be a group.
(a) A mean on

(G) is a functional m

(G)

such that m() 0 for 0 and

m(1) = 1.
(b) A mean on

(G) is called left translation invariant if

m(L
g
) = m() (g G,

(G)),
where
(L
g
)(h) := (gh) (g, h G,

(G)).
(c) G is called amenable if there is a left translation invariant mean on

(G).
Examples 1. If G is nite, then it is amenable: Dene
m() :=
1
[G[

gG
(g) (

(G)).
2. The free group F
2
in two generators, say a and b, is not amenable. To see this, let,
for x a, b, a
1
, b
1
,
W(x) := w F
2
: w starts with x,
so that
F
2
= W(a) W(b) W(a
1
) W(b
1
) e, (4.7)
the union being disjoint. In terms of indicator functions, (4.7) becomes
1 =
W(a)
+
W(b)
+
W(a
1
)
+
W(b
1
)
+
e
.
Let w F
2
W(a). Then we have a
1
w W(a
1
) and thus w aW(a
1
). Hence,
we have the union
F
2
= W(a) aW(a
1
)
which means, in terms of indicator functions, that
1
W(a)
+
aW(a
1
)
=
W(a)
+L
a
1
W(a
1
)
;
similarly, we obtain
1 =
W(b)
+L
b
1
W(b
1
)
.
99
Assume now that we have a left translation invariant mean m

(G)

. We have:
1 = m(1)
m(
W(a)
+
W(b)
+
W(a
1
)
+
W(b
1
)
)
= m(
W(a)
) +m(
W(b)
) +m(
W(a
1
)
) +m(
W(b
1
)
))
= m(
W(a)
) +m(
W(b)
) +m(L
a
1
W(a
1
)
) +m(L
b
1
W(b
1
)
))
= m(
W(a)
+L
a
1
W(a
1
)
) +m(
W(b)
+L
b
1
W(b
1
)
))
m(1) +m(1)
= 2.
This is impossible.
3. Let G be an abelian group. Let
K := m

(G)

: |m| 1 and m is a mean.

Clearly, K is convex. Since K is weak

-closed in the closed unit ball of

(G)

, it is
weak

compact by Theorem 4.2.9. Clearly, if m K, then so is L

g
m for all g G.
The family L

g
: g G is an abelian semigroup of continuous ane mappings on
K. By Theorem 4.4.2, there is m
0
K such that L

g
m
0
= m
0
for all g G, i.e.
m
0
() = (L

g
m
0
)() = m
0
(L
g
) (g G,

(G)).
Hence, G is amenable.
4.5 A geometric consequence of the HahnBanach theorem
Lemma 4.5.1 Let E be a locally convex vector space, and let U be a neighborhood of 0.
Then any linear functional : E F with sup[(x)[ : x U < is continuous.
Proof Let > 0, and let C := sup[(x)[ : x U. Then
V :=
_

C + 1
x : x U
_
is a neighborhood of 0 such that [(x)[ < for all x V . Hence, is continuous at 0 and
thus everywhere. .
Lemma 4.5.2 Let E be a locally convex vector space, and let U and K be non-empty
convex subsets of E with U K = and U open. Then there are E

and c R such
that
Re (x) < c Re (y) (x U, y K).
100
Proof Consider rst the case where F = R.
Fix x
0
U and y
0
K. Let z
0
:= y
0
x
0
, and dene
V := U K +z
0
.
Then V is an open, convex neighborhood of 0. Let
V
be the corresponding Minkowski
functional. As in the proof of Proposition 4.3.6, one sees that
V
is a sublinear functional
on E with
V = x E :
V
(x) < 1.
In particular,
V
(z
0
) 1 holds. Let F = Rz
0
, and dene : F R by letting (tz
0
) = t
for t R. It follows that
(tz
0
) =
_
t t
V
(z
0
) =
V
(tz
0
) (t 0)
t < 0
V
(tz
0
) (t < 0).
The HahnBanach theorem (Theorem 2.1.3) then yields : E R with [
F
= and
(x)
V
(x) for all x E. Let x U and y K, and note that
(x) (y) + 1 = (x y +z
0
)
V
(x y +z
0
. .
V
) < 1;
it follows that
(x) < (y) (x U, y K). (4.8)
Since (x) 1 for x V , we have (x) 1 for x V and thus [(x)[ 1 for
x V (V ). By Lemma 4.5.1, this means that E

. By (4.8), (U) and (K) are

disjoint convex subsets of R. Let c := sup
xU
[(x)[. It follows that
(x) c (y) (x U, y K). (4.9)
It is easy to see that (U) R is open. Hence, the rst inequality in (4.9) must be strict.
Next, consider the case where F = C. Find c R and a continuous, R-linear functional

: E R such that

(x) < c

(y) (x U, y K).
Dene E

by letting
(x) =

(x) i

(ix) (x E).
This completes the proof. .
101
Theorem 4.5.3 Let E be a locally convex vector space, let F and K be non-empty, dis-
joint, convex subsets of E such that F is closed and K is compact. Then there are E

and c
1
, c
2
R such that
Re (x) c
1
< c
2
Re (y) (x K, y F).
Proof As in the proof of Lemma 4.3.7, we nd an open, convex, balanced neighborhood
V of 0 such that (K + V ) F = . By Lemma 4.5.2, there are c R and E

such
that
Re (x) < c Re (y) (x K +V, y F).
Let
c
1
:= sup
xK
Re (x) and c
2
:= sup
xK+V
Re (x).
Since (K + V ) and (F) are disjoint convex subsets of R with (K + V ) open and to
the left of (F), we have
c
2
Re (y) (y F).
Since (K) (K +V ) is compact,
(x) c
1
< c
2
(x K)
follows. .
4.6 The KrenMilman theorem
Denition 4.6.1 Let E be a linear space, and let K E be convex. A convex set
,= S K is called an extremal subset of K if tx + (1 t)y S with x, y K and
t (0, 1) only if x, y S. If x K is such that x is an extremal subset of K, the
point x is called an extremal point of K. The set of all extremal points of K is denoted
by ext K.
Lemma 4.6.2 Let E be a locally convex vector space, let ,= K E be compact and
convex, let E

, and let C := sup

xK
Re (x). Then
K

:= x K : Re (x) = C
is an extremal, compact, convex subset of K.
102
Proof Clearly, K

is compact and convex.

Let x, y K and t (0, 1) be such that
Re (tx + (1 t)y) = C.
On the other hand, we have
C = tC + (1 t)C tRe (x) + (1 t)Re (y) = Re (tx + (1 t)y) = C,
which is possible only if Re (x) = Re (y) = C. .
The KrenMilman theorem asserts that under certain conditions extremal points exist
in abundance:
Theorem 4.6.3 (KrenMilman theorem) Let E be a locally convex vector space, and
let ,= K E be compact and convex. Then K := conv(ext K).
Proof Let / be the collection of all extremal, compact, convex subsets of K. By Lemma
4.6.2, / is not empty.
For K
0
/, let /
0
consist of those sets in / contained in K
0
. Let /
0
be ordered by
set inclusion. By Zorns lemma, /
0
has a minimal element, say S. Let x S, and assume
that there is y S x. Choose E

with Re (y) < Re (x). By Lemma 4.6.2,

S

S is an extremal compact, convex subset of S. Since S is an extremal subset of K,

this means that S

is also an extremal subset of K. This contradicts the minimality of S,

so that S = x. In particular, we see that
K
0
ext K ,= . (4.10)
Clearly,

K := conv
_
ext K
_
K
holds, so that

K is compact. Assume that there is x
0
K

K. By Theorem 4.5.3, there
is E

such that
sup
x

K
Re (x) < Re (x
0
) =: C. (4.11)
But then, by Lemma 4.5.2 again,
K

= x E : Re (x) = C
belongs to /. By (4.11), we have K

K = . This, however, contradicts (4.10) (with
K
0
= K

). .
103
4.6.1 The StoneWeierstra theorem
We conclude these notes with a generalization of Theorem 2.3.3 whose proof makes use
of a number of powerful theorems we have encountered:
Theorem 4.6.4 (StoneWeierstra theorem) Let X be a compact Hausdor space,
and let A be a closed subalgebra of ((X) such that:
(a) 1 A;
(b) if f A, then

f A;
(c) if x, y X with x ,= y, there there is f A such that f(x) ,= f(y).
Then A = ((X).
Proof Assume that A ((X). By Corollary 2.1.6, there is ((X)

such that || = 1,
but [
A
= 0. It follows that
K := ((X)

: || 1 and [
A
= 0 ,= 0.
Clearly, K is convex and closed in the weak

topology. By Theorem 4.2.9, this means

that K is weak

compact and Theorem 4.6.3 implies that ext K ,= . Let ext K; it

is easy to see that || = 1. By Theorem B.3.8, there is a unique M(X) such that
(f) =
_
X
f(x) d(x) (f ((X)).
Let X
0
:= supp , and let x
0
X
0
.
We claim that X
0
= x
0
. Let x X x
0
. By (c), there is f
1
A such that
:= f
1
(x) ,= f
1
(x
0
). By (a), we have A and therefore f
2
:= f
1
A. It follows
that f
2
(x
0
) ,= 0 = f
2
(x). Let f
3
= [f
2
[
2
= f
2

f
2
A. Then (c) implies that f
3
A, and it
is clear that f
3
(x) = 0 < f
3
(x
0
). Finally, let
f :=
1
|f
3
|

+ 1
f
3
,
so that
f(x) = 0, f(x
0
) > 0, and 0 f < 1.
Since A is an algebra, we have fg, (1 f)g A for all g A and therefore
0 =
_
fg d =
_
(1 f)g d (g A).
It follows that f, (1 f) K. Let
:= |f| =
_
f d[[.
104
Since f(x
0
) > 0, there are > 0 and a neighborhood U of x
0
such that f on U. It
follows that
=
_
f d[[
_
U
f d[[ [[(U) > 0
because U X
0
,= . In a similar fashion, one shows that < 1. We also have
1 = 1
_
f d[[ =
_
(1 f) d[[ = |(1 f)|.
Hence,
=
f
|f|
+ (1 )
(1 f)
|(1 f)|
holds. Since ext K, this means that
=
f
|f|
,
i.e.
f
|f|
= 1 [[-alsmost everywhere. Since f is continuuos, f equals on X
0
. Since
x
0
X
0
, we have
= f(x
0
) > f(x) = 0,
so that x / X
0
. Consequently, X
0
= x
0
. We thus have F with [[ = 1 such that
=
x
0
. Since
_
1 d(
x
0
) = ,= 0,
this contradicts the choice of . .
Corollary 4.6.5 Let ,= K R
N
be compact, and let f ((K). Then, for each > 0,
there is a polynomial p in N variables such that |f p| < .
Proof Apply Theorem 4.6.4 with
A := p[
K
: p is a polynomial in N variables.
It follows that A = ((K). .
Denition 4.6.6 Let X and Y be compact Hausdor spaces, and let f ((X) and
g ((Y ). Then f g ((X Y ) is dened through
(f g)(x, y) := f(x)g(y) (x X, y Y ).
The tenor product ((X)((Y ) of ((X) and ((Y ) is dened as the linear span in ((XY )
of the set f g : f ((X), g ((Y ).
Corollary 4.6.7 Let X and Y be compact Hausdor spaces, and let f ((XY ). Then
((X) ((Y ) is dense in ((X Y ).
Proof Let A := ((X) ((Y ), and apply Theorem 4.6.4. .
105
Appendix A
Point set topology
We need point set topology in this course for two reasons:
spaces of continuous functions are important examples of Banach spaces;
later in this course, we need to consider topological vector spaces which are not
normed.
This appendix contains the necessary background in point set topology. For most
statements, I have included proofs.
A.1 Open and closed sets
A topological space is a set that has just enough structure, so that we can speak sensibly
of continuous functions on it. The notion of an open set is crucial:
Denition A.1.1 A topological space is a non-empty set X together with a family of
subsets of X such that the following properties are satised:
(i) , X ;
(ii) if | is a family of sets in , then

U : U | ;
(iii) if U
1
, . . . , U
n
, then U
1
U
n
.
The family is called the topology of X.
Examples 1. Let (X, d) be a metric space. Dene U X as open if, for each x U,
there is > 0 such that
B

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

It is well known that the collection of open subsets of X forms a topology. Dierent
metrics can induce the same topology.
106
2. The collection of all subsets of any non-empty set is a topology. This topology is
called the discrete topology of X.
3. For any non-empty set X, the collection , X is a topology. This topology is
called the chaotic topology of X.
Exercise A.1 Verify the statement made in the rst example. Show that, if (X, d) is any metric
space, then
X X [0, ), (x, y)
d(x, y)
1 +d(x, y)
is also a metric and induces the same topology as d.
Denition A.1.2 Let (X, ) be a topological space.
(i) A subset U of X is open if U .
(ii) A subset F of X is closed if X F is open.
Passing to complements, the following is an immediate consequence of Denitions
A.1.1 and A.1.2.
Theorem A.1.3 Let X be a topological space. Then the following are true:
(i) and X are closed;
(ii) if T is a family of closed subsets of X, the

F : F T is closed;
(iii) if F
1
, . . . , F
n
are closed, then F
1
F
n
is closed.
A.2 Continuity
In order to dene continuity for functions between arbitrary topological spaces, we rst
introduce the notion of a neighborhood:
Denition A.2.1 Let X be a topological space, and let x X. A set U X is called a
neighborhood of x if there is an open set V U such that x V .
Exercise A.2 Show that a subset of a topological space is open if and only if its a neighborhood
of each of its points.
Denition A.2.2 Let X and Y be topological spaces. A function f : X Y is continu-
ous at x
0
X if f
1
(U) is a neighborhood of x
0
for each neighborhood U of f(x
0
). If f
is continuous at each x X, we simply say that f is continuous.
107
Exercise A.3 Let X and Y be topological spaces. Show that f : X Y is continuous if and
only if f
1
(U) is open for each open U Y if and only if f
1
(F) is closed for each closed F Y .
For metric spaces, Denition A.2.2 is (equivalent to) the usual denition:
Proposition A.2.3 Let (X, d) and (Y, ) be metric spaces, and let x
0
X. Then f :
X Y is continuous at x
0
in the usual sense if and only if f is continuous at x
0
in the
sense of Denition A.2.2.
Proof Suppose that f is continuous at x
0
in the usual sense. Let U be a neighborhood
of f(x
0
). By the denition of a neighborhood, there is an open set V U such that
f(x
0
) V . From the denition of an open set in a metric space, there is > 0 such that
B

(f(x
0
)) V . From the denition of continuity in the context of metric spaces, there is
> 0 such that, for all x X,
d(x
0
, x) < = (f(x
0
), f(x)) < , (A.1)
i.e.
B

(x
0
) f
1
(B

(f(x
0
))) f
1
(V ) f
1
(U).
Since B

(x
0
) is an open set containing x
0
, this means that f
1
(U) is a neighborhood of
x
0
.
Suppose conversely that f is continuous at x
0
in the sense of Denition A.2.2. Let
> 0. Then B

(f(x
0
)) is a neighborhood of f(x
0
). By Denition A.2.2, f
1
(B

(f(x
0
)))
is a neighborhood of x
0
, so that there is an open set V f
1
(B

(f(x
0
))) with x
0
V .
From the denition of open sets in metric spaces, there is > 0 such that
B

(x
0
) V f
1
(B

(f(x
0
))).
But this just states that (A.1) holds for all x X. .
Exercise A.4 Show that, if X is a non-empty set equipped with the discrete topology, then each
function from X into any topological space is continuous.
Exercise A.5 Let X be a non-empty set equipped with its chaotic topology. Describe the con-
tinuous functions on X into a metric space.
Theorem A.2.4 Let X be a topological space, and let (f
n
)

n=1
be a sequence of F-valued
continuous functions on X that converges uniformly to a function f on X. Then f is
continuous.
108
Proof Let x
0
X be arbitrary, and let U F be a neighborhood of f(x
0
). Without loss
of generality (Why?), we may suppose that U = B

(f(x
0
)) for some > 0. Since f
n
f
uniformly on X, there is N N such that
[f
n
(x) f(x)[ <

3
(x X, n N). (A.2)
This means, in particular, that f
N
(x
0
) U. Let V := B

3
(f
N
(x
0
)). Then V is a neigh-
borhood of f
N
(x
0
). By Denition A.2.2, this means that f
1
N
(V ) is a neighborhood of x
0
.
Hence, there is an open set W f
1
N
(V ) with x
0
W.
For x W, we have
[f(x) f(x
0
)[ [f(x) f
N
(x)[ +[f
N
(x) f
N
(x
0
)[ +[f
N
(x
0
) f(x
0
)[
<

3
+[f
N
(x) f
N
(x
0
)[ +

3
, by (A.2),
<

3
+

3
+

3
, by the choice of W,
= ,
i.e. f(x) B

(f(x
0
)). Hence, W f
1
(B

(f(x
0
))), so that f
1
(B

(f(x
0
))) is a neighbor-
hood of x
0
. .
A.3 (Local) compactness
You probably have already encountered the notion of compactness in the context of metric
spaces. Since we have a notion of openness, in arbitrary topological spaces, we can dene
compact topological spaces through a nite covering property:
Denition A.3.1 A topological space K is called compact if for any family | of open
subsets of K such that K =

U : U | there are U
1
, . . . , U
n
| such that K =
U
1
U
n
.
Exercise A.6 Show that a topological space X is compact if and only if X has the nite inter-
section property, i.e. for each family T of closed subsets of X such that

F : F T = , there
are F
1
, . . . , F
n
T such that F
1
F
n
= .
The following theorem, which we state without proof, characterizes the compact metric
spaces:
Theorem A.3.2 Let X be a metric space. Then the following are equivalent:
(i) X is compact.
(ii) Every sequence in X has a convergent subsequence.
109
Example A subset of F
N
is compact precisely when it is closed and bounded (this is
immediate from the HeineBorel theorem).
We also need to speak of compactness of subsets of arbitrary topological spaces:
Denition A.3.3 Let (X, ) be a topological space, and let Y X be non-empty. Then
[
Y
:= Y U : U
is the relative topology on X induced by .
Exercise A.7 Let (X, ) be a topological space, and let Y X be non-empty. Show that (Y, [
Y
)
is a topological space.
When saying that a certain subset of a topological space is compact (or has some other
topological property), we just mean that it is compact (or has that other property) with
respect to its relative topology.
Theorem A.3.4 Let K be a compact, topological space, let Y be a topological space, and
let f : K Y be continuous. Then f(K) is compact.
Proof Let | be a family of open sets of X such that f(K)

U : U |. Then
f
1
(U) : U | is a family of open subsets of K such that K =

f
1
(U) : U |.
Since K is compact, there are U
1
, . . . , U
n
| such that
K = f
1
(U
1
) f
1
(U
n
)
and hence
f(K) U
1
U
n
.
From the denition of the relative topology on f(K), this means that f(K) is compact
(in its relative topology). .
Corollary A.3.5 Let K be a compact topological space, and let f : K R be continuous.
Then f is bounded and attains both its minimum and its maximum on K.
Exercise A.8 Derive Corollary A.3.5 for Theorem A.3.4.
In a metric space, we can always separate two distinct points through disjoint open
balls. The following denition is a generalization of this property of metric spaces:
Denition A.3.6 A topological space X is called a Hausdor space if, for any x, y X
with x ,= y, there are open sets U, V X with U V = , x U, and y V .
110
Example Every metric space is a Hausdor space.
Exercise A.9 Give an example of a compact topological space which is not a Hausdor space.
Proposition A.3.7 Let K be a compact topological space and let F K be closed. Then
F is compact.
Proof Let | be a family of open sets of K such that F

U : U |. It follows that
K =
_
U : U | (K F).
Since K is compact and K F is closed, there are U
1
, . . . , U
n
| such that
K = U
1
U
n
(K F)
and thus
F U
1
U
n
.
This complete the proof. .
Proposition A.3.8 Let X be a Hausdor space, and let K X be compact. Then K is
closed in X.
Proof Let x X K. For each y K, there are thus open subsets U
y
and V
y
of X with
U
y
V
y
= and x U
y
and y V
y
. Since K

V
y
: y K and K is compact, there
are y
1
, . . . , y
n
K such that K = V
y
1
V
y
n
. Let
W
x
:= U
y
1
U
y
n
.
Then W
x
is open such that x W
x
and U
x
X K.
Since x X K was arbitrary, we can dene W
x
for each such x. Consequently,
X K =
_
W
x
: x X K
is open, so that K is closed. .
Exercise A.10 Does Proposition A.3.8 remain true from compact subsets of non-Hausdor s-
paces.
The following theorem is often useful when it comes to establishing the continuity of
a function:
Theorem A.3.9 Let K be a compact topological space, let X be a Hausdor space, and
let f : K X be continuous and bijective. Then f
1
: X K is also continuous.
111
Proof Let F K be closed. We have to show that (f
1
)
1
(F) = f(F) is closed. By
Proposition A.3.7, F is compact and so is f(F) by Theorem A.3.4. Since X is Hausdor,
this means that f(F) is closed. .
Exercise A.11 Give an example that shows that Theorem A.3.9 becomes false if we drop the
demand that X be Hausdor.
Denition A.3.10 Let X be a topological space, and let S X be arbitrary. Then the
closure of S in X is dened as
S =

F : F X is closed with S F.
Denition A.3.11 Let X be a topological spaces. Then S X is called relatively
compact if S = or if S is compact.
Denition A.3.12 A topological space X is called locally compact if for each point in X
has a relatively compact neighborhood.
Example F
N
is locally compact, but not compact.
Denition A.3.13 Let X be a locally compact Hausdor space. A continuous function
f : X F is said to vanish at innity if, for each > 0, there is a compact subset K of
X such that [f(x)[ < for all x X K. We write (
0
(X, F) for the linear space of all
continuous function on X into F that vanish at innity.
Exercise A.12 Show that for X = R this yields the usual denition of a function vanishing at
innity, i.e.
f (
0
(R, F) lim
t
f(t) = 0.
We state the following theorem without proof; (i) is a consequence of Urysohns lemma,
and (ii) is proved on page 4.
Theorem A.3.14 Let X be a locally compact Hausdor space. Then:
(i) For each compact subset K of X and each closed subset F of X such that KF = ,
there is f (
0
(X, R) such that f[
F
0 and f[
K
1.
(ii) ((
0
(X, F), | |

) is a Banach space.
Exercise A.13 Prove Theorem A.3.14(ii).
112
A.4 Convergence of nets
In metric spaces, topological concepts such as closedness and continuity can be charac-
terized through convergent sequences. This is not possible anymore for topological spaces
(see below), but there is an appropriate substitute:
Denition A.4.1 A non-empty set A is called directed if there is an oder relation on
A such that:
(a) If and , then .
(b) If and , then = .
(c) For any , A, there is A such that and .
Denition A.4.2 A net in a non-empty set S is a function from a directed set into S.
Example All sequences are nets.
If A is a directed set and s: A S is a net, we denote s by (s

)
A
and write s

for
s(); if the index set A is obvious or irrelevant, we often write (s

.
Denition A.4.3 Let X be a topological space, let x X, and let (x

be a net in
X. Then (x

converges to x in symbols x = lim

or x

x if, for each

neighborhood U of x, there is an index such that x

U for each index with .

Proposition A.4.4 Let X a Hausdor space, let x, y X, and let (x

be a net in X
that converges to both x and y. Then x = y.
Proof Assume that x ,= y. Then there are disjoint open sets U, V X such that x U
and y V . Since x = lim

, there is
x
such that x

U for all such that

x
.
Since y = lim

, there is
y
such that x

V for all such that

y
. Choose
such that
x
and
y
. Then x

U V for all ~ , which is impossible since

U V = . .
Exercise A.14 Let X be a nonempty set equipped with the chaotic topology. Show that every
net in X converges to every point of X.
A.5 Closedness and continuity via nets
Theorem A.5.1 Let X be a topological space, and let S be a non-empty subset of X.
Then the following are equivalent for x X:
(i) x S;
113
(ii) there is a net (x

in S such that x = lim

.
Proof (i) = (ii): Let ^
x
denote the collection of all neighborhoods of x. For U, V ^
x
dene:
U V : U V.
Then ^
x
is directed. By the denition of S, there is, for each U ^
x
, an element
x
U
U S. Then (x
U
)
UA
x
is a net in S such that x = lim
U
x
U
.
(ii) = (i): Let (x

be a net in S such that x = lim

, and assume that x U :=

X S. Then there is such that x

U X S for ~ , which is impossible. .

This theorem is wrong for sequences:
Example Let X be an uncountable set, and dene a subset of X as open if it is empty
or has countable complement. It follows that a closed subset of X is the whole space
or countable. Pick x X. Then the only closed set containing S := X x is X, so
that S = X. Assume that there is a sequence (x
n
)

n=1
in S such that x
n
n
x. Then
U := X x
1
, x
2
, . . . is an open neighborhood of x, but x
n
/ U for all n N.
Corollary A.5.2 Let X be a topological space. Then the following are equivalent for a
non-empty subset F of X:
(i) F is closed;
(ii) for each net (x

in F that converges to x X, we have x F.

Proof (i) = (ii): Let (x

be a net in F with limit x X. Assume that x / F, i.e.

x U := X F. Since U is a neighborhood of x, there is such that x

U for ~ .
But this is impossible, since (x

is a net in F.
(ii) = (i): It follows immediately from Theorem A.5.1 that F = F, so that F is
closed. .
Theorem A.5.3 Let X and Y be topological spaces, and let x X. Then the following
are equivalent for a map f : X Y :
(i) f is continuous at x;
(ii) for each net (x

in X such that x

x, we have f(x

)

f(x).
Proof (i) = (ii): Let U be a neighborhood of f(x). Then f
1
(V ) f
1
(U) is open,
so that f
1
(U) is a neighborhood of x. Since x = lim

, there is an index such that

x

f
1
(U) for ~ . But this means that f(x

) U for ~ .
114
(ii) = (i): Let U be a neighborhood of f(x), and assume towards a contradiction
that f
1
(U) is not a neighborhood of x. Hence, V , f
1
(U) for each open subset V of
X with x in V . Let 1
x
denote the collection of all open subsets of X containing X. Then
1
x
is directed in a natural way. By assumptions, we can choose x
V
V f
1
(U) for each
V 1
x
. It is clear that lim
V
x
V
= x (Why?). But since f(x
V
) / U for all V 1
x
, it
follows that f(x
V
) , f(x). .
A.6 Compactness via nets
Denition A.6.1 Let A and B be directed sets. A map : B A is called conal if, for
each A, there is B such that () ~ .
Denition A.6.2 Let X be a non-empty set, and let (x

)
A
and (y

)
B
be nets in X.
Then (y

)
B
is a subnet of (x

)
A
if y

= x
()
for a conal map : A B.
Exercise A.15 Does a subnet of a sequence have to be again a sequence?
Proposition A.6.3 Let X be a topological space, let (x

be a net in X, and let x X

be a limit of (x

. Then each subnet of (x

converges to x.
Exercise A.16 Prove Proposition A.6.3.
Denition A.6.4 Let X be a topological space, and let (x

be a net in X. A point
x X is an cluster point of (x

if, for each and for each neighborhood U of x, there

is ~ such that x

U.
Proposition A.6.5 Let X be a topological space, and let (x

be a net in X. Then the

following are equivalent for x X:
(i) x is an cluster point of (x

;
(ii) there is a subnet of (x

converging to x.
Proof (i) =(ii): Let ^
x
denote the collection of all neighborhoods of x. Let B := A^
x
.
For (
1
, U
1
), (
2
, U
2
) B dene:
(
1
, U
1
)

(
2
, U
2
) :
1

2
and U
1
U
2
.
This turns B into a directed set. Let (, U) B. By the denition of an cluster point,
there is (, U) A with (, U) ~ such that x
(,U)
U. The map : B A is
conal, and the net (x
(,U)
)
(,U)B
converges to x.
(ii) = (i): Clear by denition. .
115
We can now prove the analogue of Theorem A.3.2 for general topological spaces:
Theorem A.6.6 For a topological space X the following are equivalent:
(i) X is compact;
(ii) each net in X has a convergent subnet.
Proof (i) = (ii): Let (x

be a net in X. By Proposition A.6.5, it is sucient to show

that (x

has an cluster point. Assume that (x

has no cluster point. Then, for each

x X, there is a neighborhood U
x
of x (which we can choose to be open) and an index

x
such that x

/ U
x
for ~
x
. The family (U
x
)
xX
is an open cover of X and thus has
a nite subcover U
x
1
, . . . , U
x
n
. Chose an index such that ~
j
for j = 1, . . . , n.
Hence, for ~
x

/ U
x
1
U
x
n
= X,
which is absurd.
(ii) = (i): Assume that K is not compact. Then there is an open cover U which
has no nite subcover. Let T(U) be the collection of all nite subsets of U ordered by set
inclusion. For each | T(U), there is
x
|
X
_
U : U | =

X U : U |
(otherwise, U would have a nite subcover). By hypothesis, the net (x
|
)
|T(U)
has an
cluster point x X. For any U U, and for any open neighborhood V of x, there
is 1 ~ U with x
1
V and thus V X U ,= . Assume that x / X U. Then
x U, so that U would be an open neighborhood of x; by the foregoing, we would have
U X U ,= , which is absurd. It follows that x X U. Since U U is arbitrary, we
have
x

X U : U U = X
_
U : U U = ,
which is again absurd. .
A.7 Tychonos theorem
Tychonos theorem is possibly the deepest theorem in point set topology. It states that
compactness is preserved under arbitrary Cartesian products.
If and are two topologies, the is called coarser than if has fewer open sets
then ( is then called ner than ).
116
Denition A.7.1 Let (X
i
)
iI
be a family of topological spaces, let X :=

iI
X
i
, and
let
i
: X X
i
be the projection onto the i-th coordinate. The product topology on X is
the coarsest topology such that the projections
i
are all continuous.
Lemma A.7.2 Let (X
i
)
iI
be a family of topological spaces, and let X :=

iI
X
i
. Then
the open subsets of X in the product topology are exactly the unions of sets of the form

1
i
1
(U
i
1
)
1
i
n
(U
i
n
), (A.3)
where n N, i
1
, . . . , i
n
I, and U
i
1
X
i
1
, . . . , U
i
n
X
i
n
are open.
Proof The collection of all union of sets of the form (A.3) is a topology that makes the
projections continuous; hence, each open subset of X in the product topology is of that
form.
Conversely, any topology making the projections continuous, must contain the sets of
the form (A.3) and thus their arbitrary unions. .
Proposition A.7.3 Let (X
i
)
iI
be a family of topological spaces, and let X :=

iI
X
i
.
The the following are equivalent for a net (x

in X and a point x X:
(i) x

x in the product topology;

(ii)
i
(x

)


i
(x) for each i I.
Proof (i) = (ii): This is clear by Theorem A.5.3, since the projections are continuous.
(ii) = (i): Let U be a neighborhood of x X. By Lemma A.7.2, there are n N,
i
1
, . . . , i
n
I, and open sets U
i
1
X
i
1
, . . . , U
i
n
X
i
n
with
x
1
i
1
(U
i
1
)
1
i
n
(U
i
n
) U.
By hypothesis, there is such that
i
j
(x

) U
i
j
for ~ and j = 1, . . . , n. This,
however, means that
x

1
i
1
(U
i
1
)
1
i
n
(U
i
n
) U
for ~ . .
Exercise A.17 Let (X
i
)
iI
be a family of Hausdor spaces. Show that

iI
X
i
equipped with
the product topology is also Hausdor.
Theorem A.7.4 (Tychonos theorem) Let (X
i
)
iI
be a family of compact topological
spaces. Then X :=

iI
X
i
equipped with the product topology is also compact.
117
Proof Let (x

be a net in X. Let J I; we call an element x X a J-partial cluster

point of (x

if x[
J
is an cluster point of (x

[
J
)

in

iJ
X
i
. We call x X a partial
cluster point of (x

if it is a J-partial cluster point of for some J I; J is called the

domain of x.
Let T be the set of all partial cluster points of (x

. Let x
1
, x
2
T. We dene
x
1
x
2
: domain of x
1
domain of x
2
and x
2
[
domain of x
1
= x
1
.
Since each X
i
is compact, (x

has i-partial cluster points for each i I.

Let / be a totally ordered subset of T. Let J :=

domain of x : x /. Dene
y

jJ
X
j
by letting y(j) := x(j) if j domain of x. Since / is totally ordered, y is
well dened. We claim that y is a J-partial cluster point of (x

. Let U

jJ
X
j
be a
neighborhood of y. By Lemma A.7.2, we may suppose that
U =
1
j
1
(U
j
1
)
1
j
n
(U
j
n
),
where n N, j
1
, . . . , j
n
J, and U
j
1
X
j
1
, . . . , U
j
n
X
j
n
are open. Clearly, for each
there is ~ such that
x

(j
k
) =
j
(x

) U
j
k
(k = 1, . . . , n),
so that x

U.
By Zorns lemma, T has a maximal element x with domain J. Assume there is i IJ.
There is a subnet (x

of (x

such that
j
(x

)


j
(x) for each j J. Since X
i
is
compact, we may nd a subnet (x

of (x

such that
i
(x

converges to some
x
i
in X
i
. Dene x

jJi
X
j
by letting x[
J
= x and x(i) = x
i
. It follows that x is a
J i-partial cluster point of (x

, which contradicts the maximality of x. .

118
Appendix B
Measure and integration
Like point set topology, measure theory is an important source of examples in functional
analysis.
In this appendix, I have collected the denitions and results we need. Proofs are not
given.
B.1 Measure spaces
Denition B.1.1 Let be a set. A -algebra over is collection S of subsets of such
that the following are satised:
(a) S;
(b) if A S, then A
c
S;
(c) if (A
n
)

n=1
is a sequence in S, then

n=1
A
n
S.
The pair (, S) is called a measurable space.
Examples 1. P() is a -algebra.
2. A : A or A
c
is countable is a -algebra.
3. A : A or A
c
is nite is not a -algebra if is innite.
4. If o P() is arbitrary, there is a smallest -algebra over containing o; this
-algebra is called the -algebra generated by o. If is a topological space, the
-algebra generated by its open subsets is called the Borel -algebra over ; we
denote it by B().
Denition B.1.2 Let (, S) be a measurable space. A (positive) measure on (, S) is
a function : S [o, ) such that:
119
(a) () = 0;
(b) (

n=1
A
n
) =

n=1
(A
n
) for each sequence (A
n
)

n=1
of pairwise disjoint sets in
S.
The triple (, S, ) is called a measure space.
Examples 1. Counting measure: any set; S = T(); (A) := [A[.
2. Dirac measure: any set with xed; S = T(); =

, i.e. (A) = 1 if
A, otherwise (A) = 0.
3. N-dimensional Lebesgue measure: = R
N
; S = B(R
N
); =
N
, i.e. N-dimensio-
nal Lebesgue measure.
Denition B.1.3 A measure space (, S, ) (or rather the measure ) is called:
(a) -nite if there is a sequence (A
n
)

n=1
in S such that =

n=1
A
n
and (A
n
) <
for each n;
(b) nite if () < ;
(c) a probability space (or rather probability measure) if () = 1.
Examples 1. N-dimensional Lebesgue measure is -nite, but not nite.
2. Any Dirac measure is a probability measure.
3. Counting measure is nite if and only if is nite, and -nite if and only if is
countable.
Denition B.1.4 Let (, S, ) be a measure space.
(a) A set N S is called a -zero set if (N) = 0.
(b) The completion of S with respect to is dened as
S : there are A, N S with A S A N and N is a -zero set.
(c) A property is said to hold -almost everywhere (short: -a.e.) on if there is a
-zero set N S such that the property in question holds on N.
Exercise B.1 Show that the completion of a -algebra with respect to a measure is again a
-algebra.
Example The completion of B(R
N
) with respect to
N
is the -algebra of Lebesgue
measurable sets.
120
B.2 Denition of the integral
Denition B.2.1 Let (, S) be a measurable space. A function f : R is called
elementary if there are
1
, . . . ,
n
R and A
1
, . . . , A
n
S such that
f =
n

k=1

A
k
.
Denition B.2.2 Let (, S, ) be a measure space, and let f : R be an elementary
function. The integral of f with respect to is dened as
_
f d :=
n

k=1

k
(A
k
),
where f =

n
k=1

A
k
with
1
, . . . ,
n
R and A
1
, . . . , A
n
S.
Remark It can be shown that the value
_
f d is independent of the representation f =

n
k=1

A
k
.
Denition B.2.3 Let (, S) be a measurable space. A function f : R is
called S-measurable if : f() ) S for each R.
Examples 1. Every elementary function is measurable.
2. If is any topological space, then every continuous function f : R is Borel-
measurable.
3. Every increasing function f : R R is Borel-measurable.
4. Measurability is preserved under taking pointwise suprema, inma, and limits.
Proposition B.2.4 Let (, S) be a measurable space, and let f : [0, ] be S-mea-
surable. Then there is an increasing sequence (f
n
)

n=1
of S-measurable functions on
that converges to f pointwise; in case f is bounded, we can even chose that sequence in
such a way that we have uniform convergence.
Denition B.2.5 Let (, S, ) be a measure space, let f : [0, ] be S-measurable,
and let (f
n
)

n=1
be as in Proposition B.2.4. Then the integral of f with respect to is
dened as
_
f d := lim
n
_
f
n
d.
Remarks 1. It can be shown that the value
_
f d is independent of the choice of the
sequence (f
n
)

n=1
.
121
2. As the limit of an increasing sequence
_
f d always exists, but may be .
3. If
_
f d < , then : f() = is a -zero set.
4. We have always
_
f d 0, and
_
f d = 0 if and only if f = 0 -almost everywhere.
Denition B.2.6 Let (, S, ) be a measure space. A measurable function f :
R is called -integrable if
_
f
+
d < and
_
f

d < . The integral of f with

respect to is dened as
_
f d :=
_
f
+
d
_
f

d.
Remark Treating real and imaginary part separately, one can also dene the integral of
C-valued functions.
Proposition B.2.7 Let (, S, ) be a measure space, and let
L
1
(, S, ) := f : R : f is integrable.
Then:
(i) L
1
(, S, ) is a linear space;
(ii) the integral is linear on L
1
(, S, ), i.e.
_
(f +g) d =
_
f d +
_
g d (, R, f, g L
1
(, S, ));
(iii) the integral is positive, i.e. if f 0 -a.e. for some f L
1
(, S, ), then
_
f d 0.
Examples 1. For (R
N
, B(R
N
),
N
) we get the familiar N-dimensional Lebesgue inte-
gral.
2. For (, P(),

), every function f : R is integrable, and we have

_
f

= f() (f L
1
(, S, )).
3. For (N, P(N), ) with counting measure, a function f : N R is integrable if and
only if the series

n=1
f(n) converges absolutely; in this case, we have
_
f d =

n=1
f(n).
122
B.3 Theorems about the integral
The main advantage the Lebesgue integral has over the Riemann integral is the ease with
which it can be interchanged with pointwise limits. These limit theorems hold in a more
abstract measure theoretic context:
Theorem B.3.1 (monotone convergence theorem) Let (, S, ) be a measure spa-
ce, let (f
n
)

n=1
be an increasing sequence of [0, ]-valued, S-measurable functions on ,
and let f : [0, ] be their pointwise limit. Then
_
f d = lim
n
_
f
n
d.
Theorem B.3.2 (dominated convergence theorem) Let (, S, ) be a measure spa-
ce, let (f
n
)

n=1
be a sequence of R -valued, -integrable functions functions on ,
and let f, g : R be such that:
(a) f = lim
n
f
n
-a.e.;
(b) g is -integrable;
(c) [f
n
[ g -a.e. for all n N.
Then f is -integrable with
_
f d = lim
n
_
f
n
d.
Denition B.3.3 Let (, S) be a measurable space, and let and be measures on
(, S). Then is said to be absolutely continuous with respect to (in symbols: )
if every -zero set is already a -zero set.
Examples 1. Let (, S) be a measure space, let be a measure on (, S), and let
f : [0, ] be S-measurable. Dene : S [0, ] through
(A) :=
_
f
A
d (A S).
Then is a measure on (, S) (which is nite if and only if f is -integrable) such
that .
2. Let BV [a, b]. Then there is a unique measure on ([a, b], B([a, b])) such that
([c, d)) = (d) (c) (c, d [a, b])
(the integral with respect to is just the RiemannStieltjes integral with respect
to ). The measure is absolutely continuous with respect to Lebesgue measure if
and only if is absolutely continuous.
123
Theorem B.3.4 (RadonNikod ym theorem) Let (, S) be a measurable space, and
let and be measures on (, S) such that and is -nite. Then there is a
S-measurable function f : [0, ] such that
(A) :=
_
f
A
d (A S).
Any two such functions must be equal -a.e..
Remark It is necessary that be -nite in the RadonNikod ym theorem: A straight-
forward counterexample for non--nite is (R, B(R)) with counting measure as and
Lebesgue measure as .
Denition B.3.5 Let (, S) be a measurable space. A complex measure on (, S) is a
function : S C such that:
(a) () = 0;
(b) (

n=1
A
n
) =

n=1
(A
n
) for each sequence (A
n
)

n=1
of pairwise disjoint sets in
S.
Remark Every complex measure has a so called Jordan decomposition, i.e. there are
(with some strings attached) uniquely determined nite measures
1
,
2
,
3
, and
4
such
that =
1

2
+ i(
3

4
). A function is said to be -integrable if it is
j
-integrable
for j = 1, . . . , 4. The integral of a -integrable function f is then dened as
_
f d :=
4

j=1
_
g d
j
.
Denition B.3.6 Let (, S) be a measurable space, and let be a complex measure on
(, S). The total variation of is dened as
|| := sup
_
_
_
n

j=1
[(A
j
)[ : n N, A
j
A
k
= for j ,= k, =
n
_
j=1
A
j
_
_
_
.
Denition B.3.7 Let be a locally compact space. A (positive) measure on (, B())
is called regular if
(a) (K) < for all compact K ,
(b) (A) = inf(U) : A U with U open for all A B(), and
(c) (U) = sup(K) : K U is compact for all open U .
124
A complex measure is called regular if all the measures occurring in its Jordan decomposi-
tion are regular. The collection of all regular, complex measures on (, B()) is denoted
by M().
Example N-dimensional Lebesgue measure is regular.
Theorem B.3.8 (Riesz representation theorem) Let be a locally compact space.
Then T : M() (
0
()

with
(T)(f) :=
_
f d ( M(), f (
0
())
is a linear bijection such that
|T| = || ( M()).
Remark The Riesz representation theorem is also valid over R.
125