Functional Analysis

Analysis
Series

M. Winklmeier

Chigüiro Collection
Contents

1 Banach spaces 7
1.1 Metric spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2 Normed spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.3 Hölder and Minkowski inequality . . . . . . . . . . . . . . . . . . 22

2 Bounded maps; the dual space 25

2.1 Bounded linear maps . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2 The dual space and the Hahn-Banach theorem . . . . . . . . . . 31
2.3 Examples of dual spaces . . . . . . . . . . . . . . . . . . . . . . . 36
2.4 The Banach space adjoint and the bidual . . . . . . . . . . . . . 38

3 Linear operators in Banach spaces 45

3.1 Baire’s theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.2 Uniform boundedness principle . . . . . . . . . . . . . . . . . . . 46
3.3 The open mapping theorem . . . . . . . . . . . . . . . . . . . . . 55
3.4 The closed graph theorem . . . . . . . . . . . . . . . . . . . . . . 57
3.5 Projections in Banach spaces . . . . . . . . . . . . . . . . . . . . 61
3.6 Weak convergence . . . . . . . . . . . . . . . . . . . . . . . . . . 63

4 Hilbert spaces 67
4.1 Hilbert spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.2 Orthogonality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.3 Orthonormal systems . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.4 Linear operators in Hilbert spaces . . . . . . . . . . . . . . . . . 80
4.5 Projections in Hilbert spaces . . . . . . . . . . . . . . . . . . . . 84
4.6 The adjoint of an unbounded operator . . . . . . . . . . . . . . . 87

5 Spectrum of linear operators 97

5.1 The spectrum of a linear operator . . . . . . . . . . . . . . . . . 97
5.2 The resolvent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
5.3 The spectrum of the adjoint operator . . . . . . . . . . . . . . . . 106
5.4 Compact operators . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.5 Hilbert-Schmidt operators . . . . . . . . . . . . . . . . . . . . . . 121
5.6 Polar decomposition . . . . . . . . . . . . . . . . . . . . . . . . . 124

3
4 CONTENTS

A Lp spaces 129
A.1 A reminder on measure theory . . . . . . . . . . . . . . . . . . . 130
A.2 Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
A.3 Lp spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

B Exercises 139

References 152
CONTENTS 5

Notation
The letter K usually denotes either the real field R or the complex field C. The
positive real numbers are denoted by R+ := (0, ∞).
Chapter 1. Banach spaces 7

Chapter 1

Banach spaces

1.1 Metric spaces

We repeat the definition of a metric space.

Definition 1.1. A metric space (M, d) is a non-empty set M together with a

map

d:M ×M →R

such that for all x, y, z ∈ M :

(i) d(x, y) = 0 ⇐⇒ x = y,
(ii) d(x, y) = d(y, x),
(iii) d(x, y) ≤ d(x, z) + d(z, y).

The last inequality is called triangle inequality. Usually the metric space (M, d)
is denoted simply by M .

Note that the triangle inequality together with the symmetry of d implies

d(x, y) ≥ 0, x, y ∈ M,

since 0 = d(x, x) ≤ d(x, y) + d(y, x) = 2d(x, y).

It is easy to check that

|d(x, y) − d(y, z)| ≤ d(x, z), x, y, z ∈ M.

A subset N ⊆ M is called bounded if

diam N := sup{d(x, y) : x, y ∈ N } < ∞.

8 1.1. Metric spaces

Let r > 0 and x ∈ M . Then

Br (x) := {y ∈ M : d(x, y) < r} =: open ball with centre x and radius r,

Kr (x) := {y ∈ M : d(x, y) ≤ r} =: closed ball with centre x and radius r,
Sr (x) := {y ∈ M : d(x, y) = r} =: sphere with centre x and radius r.

Examples. • R with the d(x, y) = |x − y| is a metric space.

• Let X be a set and define d : X × X → R by d(x, y) = 0 for x = y
and d(x, y) = 1 for x 6= y. Then (X, d) is a metric space. d is called the
discrete metric on X.

Let (M, d) be a metric space. Recall that the metric d induces a topology on
M : a set U ⊆ M is open if and only if for every p ∈ U there exists an ε > 0
such that Bε (p) ⊆ U . In particular, the open balls are open and closed balls are
closed subsets of M . Let x ∈ M . A subset U ⊆ M is called a neighbourhood of
x if there exists an open set Ux such that x ∈ Ux ⊆ U .
It is easy to see that the topology generated by d has the Hausdorff property,
that is, for every x 6= y ∈ M there exist neighbourhoods Ux of x and Uy of y
with Ux ∩ Uy = ∅.
Recall that a set N ⊆ M is called dense in M if N = M , where N denotes the
closure of N .

Definition 1.2. A sequence (xn )n∈N ⊆ M converges to x ∈ M if and only if

lim d(xn , x) = 0, that is,
n→∞

∀ε > 0 ∃N ∈ N : n ≥ N =⇒ d(xn , x) < ε.

The limit x is unique. A sequence (xn )n∈N is a Cauchy sequence in M if and

only if

∀ε > 0 ∃N ∈ N : m, n ≥ N =⇒ d(xn , xm ) < ε.

Definition 1.3. A metric space in which every Cauchy sequence is convergent,

is called a complete metric space.

Definition 1.4. Let (X, OX ) and (Y, OY ) be topological spaces.

(i) A function f : X → Y is called continuous if and only if f −1 (U ) is open
in X for every U open in Y .
(ii) An bijective function f : X → Y is called a homeomorphism if and only
if f and f −1 are contiunous.

The following lemma is often useful.

Lemma 1.5. Let (M, d) be a complete metric space and N ⊆ M . Then N is

closed in M if and only if (N, d|M ) is complete.
Chapter 1. Banach spaces 9

Remarks. • Every convergent sequence is a Cauchy sequence.

• Every Cauchy sequence is bounded. Recall that a sequence (xn )n∈N is
bounded if the set {xn : n ∈ N} is bounded.

Not every metric space is complete, but every metric space can be completed in
the following sense.

Definition 1.6. Let (M, dM ) and (N, dN ) be metric spaces. A map f : M → N

is called an isometry if and only if dN (f (x), f (y)) = dM (x, y) for all x, y ∈ M .
The spaces M and N are called isometric if there exists a bijective isometry
f : M → N.

Note that an isometry is necessarily injective since x 6= y implies f (x) 6= f (y)

because d(f (x), f (y)) = d(x, y) 6= 0, and that every isometry is continuous.

Theorem 1.7. Let (M, d) be a metric space. Then there exists a complete
metric space (M b and an isometry ϕ : M → M
c, d) c such that ϕ(M ) = M
c. M
c is
called completion of M ; it is unique up to isometry.

Proof. Let

CM := {(xn )n∈N ⊆ M : (xn )n∈N is a Cauchy sequence in M }

be the set of all Cauchy sequences in M . We define the equivalence relation ∼

on CM by

x∼y ⇐⇒ d(xn , yn ) → 0, n → ∞

for all x = (xn )n∈N , y = (yn )n∈N ∈ CM . It is easy to check that ∼ is indeed a
equivalence relation (reflexivity and symmetry follow directly from properties (i)
and (ii) of the definition of a metric and transitivity of ∼ is a consequence of
the triangle inequality).
Let Mc := CM / ∼ the set of all equivalence classes. The equivalence class
containing x = (xn )n∈N is denoted by [x]. On M c we define

c×M
db : M c → R, d([x],
b [y]) = lim d(xn , yn ). (1.1)
n→∞

We have to show that db is well-defined.

Let (xn )n∈N ∈ [x] and (yn )n∈N ∈ [y]. Then

|d(xn , yn ) − d(xm , ym )| ≤ |d(xn , yn ) − d(xm , yn )| + |d(xm , yn ) − d(xm , ym )|

≤ d(xn , xm ) + d(yn , ym ) → 0, m, n → ∞.

Since (d(xn , yn ))n∈N is a Cauchy sequence in the complete space R, the limit in
(1.1) exists.
10 1.1. Metric spaces

Moreover, for (x̃n )n∈N ∈ [x] and (ỹn )n∈N ∈ [y] it follows that

|d(xn , yn ) − d(x̃n , ỹn )| ≤ |d(xn , yn ) − d(x̃n , yn )| + |d(x̃n , yn ) − d(x̃n , ỹn )|

≤ d(xn , x̃n ) + d(yn , ỹn ) → 0, n → ∞.

Hence db is well-defined.
Let

ϕ:M →M
c, ϕ(x) = [(x)n∈N ].

We will show that (M c, d)

b is a complete metric space, that ϕ is an isometry and
that ϕ(M ) = Mc in several steps.
Step 1: (M
c, d)
b is a metric space.
Proof. Let [x], [y], [z] ∈ M
c. Then

• 0 = d([x],
b [y]) = lim d(xn , yn ) ⇐⇒ x∼y ⇐⇒ [x] = [y].
n→∞
• d([x],
b [y]) = lim d(xn , yn ) = lim d(yn , xn ) = d([y],
b [x]).
n→∞ n→∞
• d([x],
b [y]) = lim d(xn , yn ) ≤ lim d(xn , zn ) + d(zn , yn ) = d([x],
b [z]) +
n→∞ n→∞
d([z],
b [y]).

Step 2: ϕ is an isometry.
Proof. This follows immediately from the definition.
Step 3: ϕ(M ) = M c.
Proof. Let (xn )n∈N ∈ [x] ∈ M c and ε > 0. Then there exists an N ∈ N such that
ε
d(xn , xm ) < 2 , m, n ≥ N . Let z := xN ∈ M . Then
ε
d(ϕ(z),
b [x]) = lim d(xN , xn ) ≤ < ε.
n→∞ 2

Next we show that (Mc, d)

b is complete. Let (x̂n )n∈N be a Cauchy sequence in
M c there exists a sequence z = (zn )n∈N ⊆ M such
c. Since ϕ(M ) is dense in M
that

b xn , zn ) < 1 ,
d(b n ∈ N.
n
The sequence z is a Cauchy sequence in M because

d(zn , zm ) = d(ϕ(zn ), ϕ(zm )) ≤ d(ϕ(zn ), x̂n ) + d(x̂n , x̂m ) + d(x̂m , ϕ(zm ))

b b b b
1 b n , x̂m ) + 1 → 0,
< + d(x̂ m, n → ∞.
n m
The sequence (x̂n )n∈N converges to [z] because

b n , z) ≤ d(x̂ 1
d(x̂ b n , ϕ(zn )) + d(ϕ(z
b n ), z) < + lim d(zn , zm ) → 0, n → ∞.
n m→∞
Chapter 1. Banach spaces 11

We have shown that ϕ(M ) is a dense subset of the complete metric space (M
c, d)
b
and that ϕ is an isometry.
Finally, we have to show that Mc is unique (up to isometry). Let (N, dN ) be
complete metric space and ψ : M → N an isometry such that ψ(M ) = N . Then
the map
T : ϕ(M ) → ψ(M ), T (ϕ(x)) = ψ(x)
c → N by
can be extended to a surjective isometry T : ϕ(M ) = M
T x = T ( lim xn ) := lim T xn
n→∞ n→∞

for x = lim xn with xn ∈ ϕ(M ), n ∈ N.

n→∞

Examples. • Cn with d(x, y) = max{|xj − yj | : j = 1, . . . , n} is a complete

metric space.
p
• Cn with d(x, y) = |x1 − y1 |2 + · · · + |xn − yn |2 is a complete metric
space.
• Let C([a, b]) be the set of all continuous functions on the interval [a, b].
For f, g ∈ C([a, b]) let
d1 (f, g) := max{|f (x) − g(x)| : x ∈ [a, b]},
Z b
d2 (f, g) := |f (x) − g(x)| dx.
a

Then d1 and d2 are metrics on C([a, b]). (C([a, b]), d1 ) is complete, (C([a, b]), d2 )
is not complete.

Remark. The completion of (C([a, b]), d2 ) is L1 (a, b) (the set of all Lebesgue
integrable functions on (a, b)).

Definition 1.8. A metric space is called separable if it contains a countable

dense subset.

Proposition 1.9. Let (M, d) be a separable metric space and N ⊆ M . Then

N is separable.

Proof. We have to show that there exists a countable set B ⊆ N such that B ⊇
N where the closure is taken with respect to the metric on M . By assumption
on M there exists a countable set A := {xn : n ∈ N} ⊆ M such that A = M .
1
Let J := {(n, m) ∈ N × N : ∃y ∈ N with d(xn , y) < m }. For every (n, m) ∈ J
choose a yn,m ∈ N and define B := {yn,m : (n, m) ∈ J}. Obviously, B is a
countable subset of N . To show that B is dense in N it suffices to show that for
every y ∈ N and k ∈ N there exists a b ∈ B such that d(b, y) < k1 . By definition
1
of A there exists a xn ∈ A such that d(xn , y) < 2k . In particular, (n, 2k) ∈ J.
It follows that d(yn,2k , y) ≤ d(yn,2k , xn ) + d(xn , y) < k1 .
12 1.2. Normed spaces

1.2 Normed spaces

Definition 1.10. Let X be a vector space over K. A norm on X is a map

k·k:X →R

such that for all x, y ∈ X, α ∈ K

(i) kxk = 0 ⇐⇒ x = 0,
(ii) kαxk = |α| kxk,
(iii) kx + yk ≤ kxk + kyk.

Remarks. • Note that the implication ⇐ in (i) follows from (ii) because
k0k = k2 · 0k = 2k0k.

• Note that kxk ≥ 0 for all x ∈ X because 0 = kx − xk ≤ 2kxk. The last

inequality follows from the triangle inequality (iii) and (ii) with α = −1.

Remark. A function [ · ] : X → R which satisfies only (ii) and (iii) of Defi-

nition 1.10 is called a seminorm. As seen in the remark above for norms, a
seminorm is non-negative and satisfies [0] = 0.

Remark. A norm on X induces a metric on X by setting

d(x, y) := kx − yk, x, y ∈ X.

Hence a norm induces a topology on X via the metric and we have the concept
of convergence etc. on a normed space.

Definition 1.11. A complete normed space is called a Banach space.

Obviously, every subspace of a normed space is a normed space by restriction

of the norm. A subspace of a Banach space is a Banach space if and only if it
is closed.

Proposition 1.12. Let X be a normed space. Then the following is equivalent:

(i) X is complete.
(ii) Every absolutely convergent series in X converges in X.

Proof. Exercise 1.2.

Example 1.13 (Quotient space). Let X be a Banach space and M ⊆ X a

closed subspace. On X we have the equivalence relation

x∼y ⇐⇒ x − y ∈ M.
Chapter 1. Banach spaces 13

For x ∈ X we denote the equivalence class of X/M containing x by [x]. Then

X/M is a vector space if we set
[x] + [y] := [x + y], α[x] := [αx], x, y ∈ X, α ∈ K.
For x ∈ X let dist(x, M ) := inf{kx − mk : m ∈ M }.
• (X/M, k · k∼ ) is a normed space with
k · k∼ : X/M → R, k[x]k∼ := dist(x, M ).

Proof. First we show that k · k∼ is well-defined. For x, y ∈ X with x − y ∈

M we find
∈M
z }| {
dist(x, M ) = inf{kx − mk : m ∈ M } = inf{ky − (y − x +m)k : m ∈ M }
= inf{ky − mk : m ∈ M } = dist(y, M ).
Property (ii) in the definition of a norm is easily checked. For property
(iii) let [x], [y] ∈ X/M . Then
k[x] + [y]k∼ = k[x + y]k∼ = inf{kx + y − mk : m ∈ M }
= inf{kx − mx + y − my k : mx , my ∈ M }
≤ inf{kx − mx k : mx ∈ M } + inf{ky − my k : my ∈ M }
= k[x]k∼ + k[y]k∼ .
It is clear that [x] = 0 implies k[x]k∼ = 0. Now assume that k[x]k∼ = 0.
We have to show that x ∈ M . By definition of dist there exists a sequence
(mn )n∈N such that kx − mn k → 0, that is, (mn )n∈N converges to x. Since
M is closed, it follows that x ∈ M .

• Let X be a Banach space and M a closed subspace. Then X/M is Banach

space with the norm defined in Example 1.13.

Proof. We already saw that X/M is normed space. It remains to prove

completeness. Let ([xn ])n∈N be a Cauchy sequence.
First we show that we can assume k[xn ] − [xm ]k∼ ≤ 2−n for all m ≥
n: Choose N1 ∈ N such that k[xN1 ] − [xm ]k∼ ≤ 2−1 for all m ≥ N1 .
Next choose N2 > N1 such that k[xN2 ] − [xm ]k∼ ≤ 2−2 for all m ≥
N2 . Continuing this process, we obtain a subsequence with the desired
property. Since a Cauchy sequence converges if and only if it contains a
convergent subsequence, it suffices to prove convergence of the subsequence
constructed above.
By definition of the quotient norm we can assume that kxn − xn+1 k ≤
k[xn − xn+1 ]k∼ + 2−n < 21−n . Then (xn )n∈N is Cauchy sequence in X
because for all n > m
n−1 n−1 n−1
X X X
kxn − xm k = xn+1 − xn ≤ kxn+1 − xn k < 2 2−j .
j=m j=m j=m
14 1.2. Normed spaces

Therefore x := lim xn exists and

n→∞

k[xn ] − [x]k∼ = k[xn − x]k∼ ≤ kxn − xk → 0, n → ∞.

Remark 1.14. (i) In the proof above we used that, by definition of k · k∼ ,

for every x ∈ X and every ε > 0 there exists an x
e ∈ [x] such that ke
xk <
k[x]k∼ + ε. Equivalently, there exists an m ∈ M such that kx + mk <
k[x]k∼ + ε.
(ii) Obviously, kxk ≥ k[x]k∼ for every x ∈ X.

Examples 1.15. (i) Finite dimensional normed spaces. Cn and Rn are com-
plete normed spaces with
k · k∞ : Kn → R, kxk∞ = max{|xj | : j = 1, . . . , n}.

Let 1 ≤ p < ∞. Then Cn and Rn are complete normed spaces with

n
X  p1
k · kp : Cn → R, kxkp = |xj |p .
j=1

The triangle inequality kx+ykp ≤ kxkp +kykp is called the Minkowski inequality
(see Section 1.3).
(ii) Let T be a set and define
`∞ (T ) := {x : T → K bounded map}.
Obviously, `∞ (T ) is a vector space. Let
kxk∞ := sup{|x(t)| : t ∈ T }, x ∈ `∞ ,
be the supremum norm. Then (`∞ (T ), k · k∞ ) is a Banach space.

Proof. Exercise 1.4.

(iii) Sequence spaces.

• `∞ := `∞ (N) is a Banach space.

• For 1 ≤ p < ∞ let
n ∞
X o
`p := `p (N) := (xn )n∈N ⊆ K : |xn |p < ∞
n=1

and
∞
X  p1
kxkp := |xn |p , x ∈ `p .
n=1

With the usual component-by-component addition and multiplication with

a scalar, `p is a vector space and (`p , k · kp ) is a Banach space.
Chapter 1. Banach spaces 15

Proof. First we show that `p is a vector space. For α ∈ K and x, y ∈ `p we

have
∞
X ∞
X
|αxn |p = |α|p |xn |p < ∞
n=1 n=1

and
∞ ∞ ∞
X X
p X
p
|xn + yn |p ≤ 2 max{|xn |, |yn |} = 2p max{|xn |, |yn |}
n=1 n=1 n=1
∞
X
≤ 2p |xn |p + |yn |p = 2p (kxkpp + kykpp ) < ∞.
n=1

Hence `p is a K-vector space. Properties (i) and (ii) in the definition of a

norm are easily verified. The triangle inequality is the Minkowski inequality
(see Section 1.3).
To show that (`p , k · kp ) is complete, let (xn )n∈N be a Cauchy sequence in
`p . Set xn = (xn,m )m∈N . Then the sequence of the m-th components is a
Cauchy sequence in K because

|xn,m − xk,m | < kxn − xk kp , m ∈ N.

Since K is complete, the limit ym := lim xn,m exists. Let y := (ym )m∈N .
n→∞
k·kp
We will show that y ∈ `p and that xn −−−→ y. Let ε > 0 and N ∈ N such
that kxn − xk k < ε for all k, n ≥ N . For every M ∈ N
M
X
|xn,j − xk,j |p ≤ kxn − xk kpp < εp .
j=1

Taking the limit k → ∞ on the left hand side yields

M
X
|xn,j − yj |p < εp .
j=1

Taking the limit M → ∞ on the left hand side finally gives

∞
X
|xn,j − yj |p ≤ εp < ∞,
j=1

in particular, xn − y ∈ `p . Since `p is a vector space, we obtain y =

xn + (y − xn ) ∈ `p and kxn − ykp ≤ ε. That (xn )n∈N converges to y follows
from the inequality above since ε can be chosen arbitrarily.

(iv) Lp spaces: See measure theory.

16 1.2. Normed spaces

(v) Subspaces of `∞ . Let

d := {x = (xn )n∈N ⊆ K : xn 6= 0 for at most finitely many n},

c0 := {x = (xn )n∈N ⊆ K : lim xn = 0},
n→∞
c := {x = (xn )n∈N ⊆ K : lim xn exists},
n→∞

Obviously, the inclusions d ( c0 ( c ( `∞ hold. Moreover, it can be shown that

c0 and c are closed subspaces of `∞ and that d is a non-closed subspace of `∞ .
In particular, (c0 , k · k∞ ) and (c, k · k∞ ) are Banach spaces, (d, k · k∞ ) is not a
Banach space (see Exercise 1.5).
(vi) Spaces of continuous functions. For metric space T (e. g. an interval in R)
let

C(T ) := {f : T → K : f is continuous},
B(T ) := {f : T → K : f is bounded},
BC(T ) := C(T ) ∩ B(T ).

For f ∈ B(T ) let

kf k∞ := sup{|f (t)| : t ∈ T }.

In Analysis 1 it was shown that (B(T ), k · k∞ ) and (BC(T ), k · k∞ ) are Banach

spaces. Note that C(T ) = BC(T ) for a compact metric space T .
(vii) Spaces of differentiable functions. Let [a, b] a real interval. We can define
several norms on the vector space

C 1 ([a, b]) := {f : [a, b] → K : f is continuously differentiable}.

• (C 1 ([a, b]), k · k∞ ) is not a Banach space.

1
Proof. For n ∈ N let fn : [−1, 1] → K, fn (t) := (t2 + n−2 ) 2 . Then
the fn converge to g : [−1, 1] → K, g(t) = |t| in the k · k∞ -norm. But
g∈/ C 1 ([a, b]). Hence C 1 ([a, b]) is not closed as a subspace of the Banach
space C([a, b]), so it is not a Banach space.

• For f ∈ C 1 ([a, b]) let

kf k(1) := kf k∞ + kf 0 k∞ .

Then (C 1 ([a, b]), k · k(1) ) is a Banach space. Note that the right hand side
is finite because by assumption f 0 is continuous.

Proof. Let (xn )n∈N be a Cauchy sequence in (C 1 ([a, b]), k · k(1) ). Then
there exist x, y ∈ C([a, b]) such that xn → x and x0n → y in the supremum
norm. A well-known theorem in analysis implies x0 = y, hence xn → x in
k · k(1) .
Chapter 1. Banach spaces 17

In the following, C 1 ([a, b]) will always be considered to be equipped with

the norm k · k(1) unless stated otherwise.

Theorem 1.16. Let X be a Banach space, Y a closed subspace and N a finite

dimensional subspace of X. Then Y + N is a closed subspace. In particular,
every finite-dimensional subspace is closed.

Proof. Obviously, Y + N is a subspace of X. To proof that it is closed, we pro-

ceed by induction. Therefore we can assume without restriction that dim N = 1.
Let z ∈ X such that N = {λz : λ ∈ K} and (xn )n∈N = (yn + an z)n∈N a Cauchy
sequence in Y + N .
C ase 1. (an )n∈N is bounded. Then it contains a convergent subsequence
(ank )k∈N . Then the sequence (ynk )k∈N = (xnk − ank z)k∈N converges because
it is the sum of two convergent sequences.
C ase 2. (an )n∈N is unbounded. Then there exists a subsequence (ank )k∈N with
lim |ank | = ∞. Since (xnk )k∈N is bounded, it follows that
k→∞

1 1
z + y nk = xnk → 0, n → ∞.

ank ank

Hence d(z, Y ) = 0. Since Y is closed, this implies z ∈ Y , therefore N + Y = Y

is closed in X.
Finally, choosing Y = {0} shows that every finite-dimensional subspace is closed.

Note that the sum of two closed subspaces is not necessarily closed, see as the
following example shows. Another example can be found in [Hal98, § 15].

Example. In `1 consider the subspaces

U := {(xn )n∈N ∈ `1 : x2n = 0, n ∈ N}

V := {(xn )n∈N ∈ `1 : x2n−1 = nx2n , n ∈ N}.

Obvioulsy, U and V are closed subspaces of `1 . Let en be the nth unit vector
1
in `1 . Let m ∈ N. Then e2m−1 ∈ U ⊆ V + U and e2m = (e2m + m e2m−1 ) −
1
m e2m−1 ∈ V + U . Since span{en : n ∈ N} is a dense subset of ` 1 , it follows
that V + U = `1 .
Now we will show that V + U 6= `1 . Let
(
1
x = (xn )n∈N , xn = n2 , n even,
0, n odd.

Clearly x ∈ `1 . Suppose that there exist v = (vn )n∈N ∈ V , u = (un )n∈N ∈ U

18 1.2. Normed spaces

such that x = v + u. It follows for all m ∈ N

1
= x2m = v2m + u2m = v2m ,
(2m)2
1
0 = x2m−1 = v2m−1 + u2m−1 = mv2m + u2m−1 = + u2m−1 ,
4m2
1
impliying that u2m−1 = − 4m 2 , m ∈ N, hence u 6= `1 . Therefore x 6= V + U .

Definition 1.17. Let X be a normed space and k · k1 and k · k2 be norms on

X. They are called equivalent norms if there exist m, M > 0 such that

mkxk1 ≤ kxk2 ≤ M kxk1 , x ∈ X. (1.2)

Theorem 1.18. Let k · k1 and k · k2 be norms on a vector space X. The the

following are equivalent:

(i) k · k1 and k · k2 are equivalent.

(ii) A sequence (xn )n∈N ⊆ X converges with respect to k · k1 if and only if

it converges with respect to k · k2 and in this case the k · k1 -limit and the
k · k2 -limit are equal.

(iii) A sequence (xn )n∈N ⊆ X converges to 0 with respect to k · k1 if and only

if it converges with respect to k · k2 .

Proof. (i) =⇒ (ii) =⇒ (iii) is clear.

“(iii) =⇒ (i)”: Obviously it suffices to show the existence of M ∈ R such
that (1.2) is true. Assume no such M exists. Then there exists a sequence
(xn )n∈N ⊆ X such that kxn k1 = 1 and kxn k2 > nkxn k1 = n. Let yn := n−1 xn .
k·k1 k·k2
Then yn −−→ 0, so by assumption also yn −−→ 0. This contradicts kyn k2 > 1
for all n ∈ N.

The theorem above implies in particular, that the topologies generated by equiv-
alent norms coincide. Moreover, the identity map id : (X, k · k1 ) → (X, k · k2 ) is
uniformly continuous for equivalent norms.

Example 1.19. On C 1 ([a, b]) define the norm

kf k(2) := sup{max{|x(t)|, |x0 (t)|} : t ∈ [a, b]}.

and let k · k(1) be as in Example 1.15 (7). It is not hard to see that

kxk(1) ≤ kxk(2) ≤ 2kxk(1) , x ∈ C 1 ([a, b]).

Chapter 1. Banach spaces 19

Theorem 1.20. All norms on Kn are equivalent.

Pn
Proof. Let {e1 , . . . , en } be a basis of Kn . For x = j=1 αn en define
n
X  21
kxk2 := |αj |2 .
j=1

Obviously, k · k2 is a norm on X and it suffices to show thatPevery norm on X

n
is equivalent to k · k2 . Let k · k be a norm on X and x = j=1 αn en . Using
triangle inequality for k · k and Hölder’s inequality, we obtain
Xn X n n
X n
 12 X  21
kxk = αj ej ≤ |αj |k ej k ≤ |αj |2 k ej k2 = C kxk2

j=1 j=1 j=1 j=1
(1.3)
P  12
n
with constant C := j=1 k ej k2 independent of x.
Note that k · k2 : X → R is continuous, hence S := {x ∈ X : kxk2 = 1} is
closed being the preimage of the closed set {1} in R. In addition, S is bounded,
therefore S is compact by the theorem of Heine-Borel. Now consider the map
T : (X, k · k2 ) → R, T x = kxk. By (1.3), T is uniformly continuous, so its
restriction to the compact set S has a minimum m and a maximum M . Since
k · k is a norm, m > 0 (otherwise there would exist an x ∈ S with kxk = 0, thus
x = 0 but 0 ∈ / S). Therefore

mkxk2 = m ≤ kxk ≤ M = M kxk2 , x ∈ S,

and by the homogeneity of the norms

mkxk2 ≤ kxk ≤ M kxk2 , x ∈ X.

The theorem above implies that all norms a a finite-dimensional K-vector space
are equivalent. Moreover, it follows that every finite normed space is complete
because Kn with the Euclidean norm is complete and that a subset of a finite
dimensional normed space is compact if and only if it is bounded and closed
(Theorem of Heine-Borel for Kn with the Euclidean metric). In particular, the
unit ball in a finite dimensional space is compact.
This is no longer true in infinite dimensional normed spaces. In fact, the unit
ball is compact if and only if the dimensions of the space is finite. For the
proof we use the following theorem which is also of independent interest, as it
shows that in a certain sense quotient spaces can work as a substitute for the
orthogonal complement in inner product spaces (see 4.2).

Theorem 1.21 (Riesz’s lemma). Let X be a normed space, Y ⊆ X a closed

subspace with Y 6= X and ε > 0. Then there exists an x ∈ X such with kxk = 1
and dist(x, Y ) > 1 − ε.
20 1.2. Normed spaces

Proof. If Y = {0} or ε ≥ 1, the assertion is clear. Now assume 0 < ε < 1. Note
1
that in this case 1−ε > 1. Since Y is closed and different from X, the quotient
space X/Y is not trivial. Hence there exists an ξ ∈ X such that k[ξ]k∼ = 1. By
Remark 1.14 there exists y ∈ Y such that
1
1 = k[ξ]k∼ ≤ kξ + yk < .
1−ε
Let x = kξ + yk−1 (ξ + y). Obviously, kxk = 1 and for every z ∈ Y

kx − zk = kξ + yk−1 ξ + y − kξ + ykz ≥ kξ + yk−1 k[ξ]k∼ = kξ + yk−1 > 1 − ε.

| {z }
∈Y

Hence d(x, Y ) = inf{kx − zk : z ∈ Y } > 1 − ε.

Theorem 1.22. For a normed space X the following are equivalent:

(i) dim X < ∞,
(ii) BX := {x ∈ X : kxk ≤ 1} is compact.
(iii) Every bounded sequence in X contains a convergent subsequence.

Proof. “(i) =⇒ (ii)” follows from Theorem 1.20.

“(ii) =⇒ (i)”: Assume that BX is compact. Then there are x1 , . . . , xn ∈ X
with kxj k ≤ 1, j = 1, . . . , n, such that
n
[
BX ⊆ B 12 (xj ). (1.4)
j=1

Let U = span{x1 , . . . , xn }. If U 6= X, then, by Riesz’s lemma, there exists

an x ∈ X such that kxk = 1 and dist(x, U ) > 21 , in contradiction to (1.4).
Therefore dim X = dim U ≤ n.
“(ii) =⇒ (iii)”: If BX is compact, then obviously for every α ≥ 0 also αBX :=
{αx : x ∈ BX } is compact. Since every bounded sequence is a subset of some
αBX , it must contain a convergent subsequence.
“(iii) =⇒ (i)”: Assume that dim X = ∞. Choose x1 ∈ X with kx1 k = 1
and set U1 := span{x1 } = 6 X. By Riesz’s lemma there exists an x2 ∈ X
with kx2 k = 1 and dist(x2 , U1 ) > 21 , in particular kx1 − x2 k > 12 . Set U2 :=
span{x1 , x2 } =
6 X. Continuing this way, we obtain a sequence x = (xn )n∈N ⊆ X
with kxn − xm k > 12 for all n, m ∈ N with n 6= m. Therefore, the sequence x
does not contain a convergent subsequence, hence BX is not compact (Recall
that a compact metric space is sequentially compact).

Let X be a vector space and Λ a set. A family (xλ )λ∈Λ ⊆ X is called linearly
independent if every finite subset is linearly independent. A Hamel basis (or an
Chapter 1. Banach spaces 21

elgebraic basis) of X is a family (xλ )λ∈Λ ⊆ X that is linearly independent and

such that every element x ∈ X is a (finite!) linear combination of the xλ . The
existence of a Hamel basis can be shown with Zorn’s lemma.

Definition 1.23. Let X be a normed space. A family (xn )n∈N is a Schauder

basis of X if every x ∈ X can be written uniquely as
∞
X
αn xn with αn ∈ K.
n=1

Definition 1.24. Let (X, k · k) be a normed space over K. A subset Y ⊆ X is

said to be a total subset of X if

span(Y ) = X,

that is, if the set of all linear combinations of elements of Y is dense in X.

Theorem 1.25. A normed space (X, k · k) is separable if and only if it contains

a countable total subset.

Proof. Let A be a dense countable subset of X. Then obviously span A = X,

that is, A is a total subset of X.
Now assume that A is countable total subset of X. Let B := {λan : n ∈ N, λ ∈
e where Q
Q} e := Q if X is a R-vector space and Q
e := Q + iQ if X is a C-vector
space. In both cases B is countable. We will show that B = X. Let x ∈ X
and ε > 0. Since A is a total subset of X, there exist a1 , . . . , an ∈ A and
λ1 , . . . , λn ∈ K such that
n
X ε
kx − λj aj k < .
j=1
2

e is dense in K, there exist µ1 , . . . , µn ∈ Q

Since Q e such that
n −1
ε X
|µj − λj | < kaj k , j = 1, . . . , n.
2 j=1
Pn
Then y := j=1 µj aj ∈ span A and
Xn Xn ε X n
kx − yk ≤ x − λj aj + y − λj aj < + |µj − λj |aj

j=1 j=1
2 j=1
n
ε n X ε ε
≤ + max |µj − λj | kan k < + = ε.
2 j=1
j=1
2 2

Note that every normed space with a Schauder basis is separable, but not every
separable normed space has a Schauder basis.
22 1.3. Hölder and Minkowski inequality

Examples 1.26. (i) `p is separable for 1 ≤ p < ∞.

Proof. Let en := (0, . . . , 0, 1, 0 . . . ) be the nth unit vector in `p . We will show

that {en : n ∈ N} is a total subset of `p . Let x = (xn )n∈N ⊆ `p . Then
n
X ∞
X
x − xj ej = xj ej → 0, n → ∞.

p p
j=1 j=n+1

(ii) `∞ is not separable.

Proof. Recall that the set A := {(xn )n∈N : xn ∈ {0, 1}} i s not countable.
Obviously, A ⊆ `∞ . Let B be a dense subset of `∞ . Then for every x ∈ A there
exists an bx ∈ B such that kx − bx k∞ < 21 . Since kx − yk∞ = 1 for x 6= y ∈ A, it
follows that B has at least the cardinality of A, that is, there exists no countable
dense subset of `p .

(iii) C[a, b] is separable since by the theorem of Weierstraß the set of polyno-
mials

{[a, b] → R, x 7→ xn : n ∈ N}

is a total subset of C[a, b].

1.3 Hölder and Minkowski inequality

In this section we prove Hölder’s inequality and Minkowski’s inequality. For the
proof we use Young’s inequality.

Theorem 1.27. Let p, q ∈ (1, ∞) such that

1 1
+ = 1.
p q

Then for all a, b ≥ 0:

1 p 1 q
ab ≤ a + b . (1.5)
p q

Proof. If ab = 0, then inequality (1.5) is clear. Now assume ab > 0. Since the
logarithm is concave and p1 + 1q = 1 is follows that
1 1 q 1 1
ln ap + b ≥ ln(ap ) + ln(bq ) = ln(a) + ln(b) = ln(ab).
p q p q

Application of the monotonically increasing function exp : R → R yields (1.5).

Chapter 1. Banach spaces 23

p
Theorem 1.28 (Hölder’s inequality). Let 1 ≤ p ≤ ∞ and q = p−1 , i. e.,

1 1
+ =1
p q
1
(setting ∞ = 0). If x ∈ `p and y ∈ `q , then z = (xn yn )n∈N ∈ `1 and

kzk1 ≤ kxkp kykq . (1.6)

Proof. If x = 0 or y = 0 then the inequality (1.6) clearly holds. Also the cases
p = 1 and p = ∞ are clear.
Now assume x, y 6= 0 and 1 < p < ∞. The Young inequality (1.6) with

|xj | |yj |
a= , b=
kxkp kykq

yields

|xj | |yj | 1 |xj |p 1 |yj |q

≤ + .
kxkp kykq p kxkpp q kykqq

Taking the sum over gives

∞ ∞ ∞
1 X 1 1 X 1 1 X 1 1
|xj yj | ≤ p |xj |p + q |yj |q = + = 1.
kxkp kykq j=1 p kxkp j=1 q kykq j=1 p q
| {z } | {z }
=kxkp
p =kykqq
| {z } | {z }
=1 =1

In the special case p = q = 2 we obtain the Cauchy-Schwarz inequality.

Corollary 1.29 (Cauchy-Schwarz inequality). For x = (xn )n∈N , y =

(yn )n∈N ∈ `2 the Hölder inequality implies
X∞ 
|hx , yi| :=  xj yj  ≤ kxk2 kyk2 .
 
j=1

Theorem 1.30 (Minkowski inequality). For 1 ≤ p ≤ ∞ and x, y ∈ `p

Minkowski’s inequality holds:

kx + ykp ≤ kxkp + kykp . (1.7)

Proof. If x + y = 0 then (1.7) clearly holds. Also the cases p = 1 and p = ∞

are easy to check.
Now assume x + y 6= 0 and 1 < p < ∞. Let q ∈ (1, ∞) such that p1 + 1q = 1.
The triangle inequality in K and Hölder’s inequality (1.6) yield for all M ∈ N:
24 1.3. Hölder and Minkowski inequality

M
X M
X
p
|xj + yj | = |xj + yj | · |xj + yj |p−1
j=1 j=1
M
X M
X
p−1
≤ |xj | |xj + yj | + |yj | |xj + yj |p−1
j=1 j=1
p p
M
X M
1 X z }| {  1 M
X M
 p1  X z }| {  1
p q (p−1)q q
≤ |xj |p |xj + yj |(p−1)q + |yj |p |xj + yj |
j=1 j=1 j=1 j=1
 M
 X 1
p q
≤ kxkp + kxkp |xj + yj | .
j=1

P 1
M p q
Note that j=1 |x j + yj | 6= 0 for M large enough. Hence the above
inequality yields
M
X  p1
|xj + yj |p ≤ kxkp + kxkp
j=1

p 1


using p − q =p 1− q = 1. Taking the limit M → ∞ finally proves (1.7).
Chapter 2. Bounded maps; the dual space 25

Chapter 2

Bounded maps; the dual

space

2.1 Bounded linear maps

Definition 2.1. Let X, Y be normed spaces over the same field K. The set of
all linear continuous maps X → Y is denoted by L(X, Y ), i. e.,

L(X, Y ) = {T : X → Y : T linear and continuous}

and L(X) := L(X, X).

Recall that the following is equivalent:

(i) T : X → Y is continuous
(ii) lim T xn = T limn→∞ xn for every convergent sequence (xn )n∈N ∈ X
n→∞

(iii) ∀ x0 ∈ X ∀ ε > 0 ∃ δ > 0 : kx − x0 k < δ =⇒ kT x − T x0 k < ε

(iv) U ⊆ Y open =⇒ T −1 (U ) = {x ∈ X : f (x) ∈ U } open in X.

Definition 2.2. Let X, Y be normed spaces over the same field K. For a linear
map T : X → Y define the operator norm

kT k := sup{kT xk : x ∈ X, kxk = 1}.

If kT k < ∞ then T is called a bounded linear operator and kT k is the operator

norm of kT k.

Remark 2.3. (i) For a continuous linear map T : X → Y

kT xk ≤ kT k kxk, x ∈ X.
26 2.1. Bounded linear maps

Proof. The inequality is obvious for x = 0 or kxk = 1. For x ∈ X \ {0} let

e = kxk−1 x. By definition of kT k we find kT xk = kxk kT x
x ek ≤ kxk kT k.
Note that the inequality is also true if T is unbounded and x 6= 0.

(ii) The following is easy to check:

kT k = sup{kT xk : x ∈ X, kxk = 1}
= sup{kT xk : x ∈ X, kxk ≤ 1}
n kT xk o
= sup : x ∈ X, x 6= 0
kxk
= inf{M ∈ R : ∀ x ∈ X kT xk ≤ M kxk}.

Remark 2.4. (i) For S, T ∈ L(X, Y ) and λ ∈ K we define

(λT + S) : X → Y, (λT + S)x := λT x + Sx.

Since the sum and composition of continuous functions is continuous, and

(λT + S obviously is linear, L(X, Y ) is a vector space.
It will be shown in Theorem 2.6 that k · k is indeed a norm. Note the the
operator norm depends on the norms on X and Y . This is can be made
explicit using the notation kT kL(X,Y ) , or similar notation.
(ii) Let X, Y, Z be normed spaces and T ∈ L(X, Y ), S ∈ L(Y, Z). Then

ST : X → Z, ST x := S(T x).

Obviously, ST ∈ L(X, Z) as composition of continuous linear functions

and kST k ≤ kSk kT k because by Remark 2.3

kST xk ≤ kSk kT xk ≤ kSk kT k kxk, x ∈ X.

In particular, L(X) is an algebra.

Theorem 2.5. Let X, Y be normed spaces, T : X → Y linear. The following

is equivalent:
(i) T is continuous.
(ii) T is continuous in 0.
(iii) T is bounded.
(iv) T is uniformly continuous.

Proof. The implications (iii) =⇒ (iv) =⇒ (i) =⇒ (ii) are obvious.

“ (ii) =⇒ (iii)”: Assume that T is not bounded. Then there exists a sequence
(xn )n∈N ⊆ X such that kxn k = 1 and kT xn k > n for all n ∈ N. Let yn := n−1 xn .
Then yn → 0 but kT yn k > 1 for all n ∈ N in contradiction to the continuity of
T in 0.
Chapter 2. Bounded maps; the dual space 27

Theorem 2.6. Let X, Y be normed spaces.

(i) L(X, Y ) is a normed space.
(ii) If Y is Banach space, then L(X, Y ) is a Banach space.

Proof. (i) In Remark 2.4 we have seen that L(X, Y ) is a vector space. From
definition of the operator norm it is clear that kT k = 0 if and only if T = 0
and that kλT k = |λ| kT k for all λ ∈ K. To prove the triangle inequality let
S, T ∈ L(X, Y ) and x ∈ X.

k(S + T )xk = kSx + T xk ≤ kSxk + kT xk ≤ kSk + kT k.

Taking the supremum over all x ∈ X with kxk = 1 yields kS + T k ≤ kSk + kT k.

(ii) Let (Tn )n∈N be a Cauchy sequence in L(X, Y ). For x ∈ X, the sequence
(Tn x)n∈N is a Cauchy sequence in Y because

kTn x − Tm xk ≤ kTn − Tm k kxk.

Since Y is complete, we can define

T : X → Y, T x := lim Tn x.
n→∞

It is easy to check that T is linear. That T is bounded and Tn → T follows as

in Example 1.13(2): For ε > 0 exists an N ∈ N such that
ε
kTn − Tm k < , n, m ≥ N.
2
In particular, for all x ∈ X it follows for n ≥ N that
ε
kT x − Tn xk ≤ kT x − Tm xk + kTm x − Tn xk ≤ kT x − Tm xk + , m ∈ N.
2
(2.1)

Taking the limit m → ∞ on the right hand side yields kT x − Tn xk ≤ 2ε < ε. It

follows that T − Tn is a bounded linear map. Since L(X, Y ) is a vector space,
also T = Tn + (T − Tn ) is a bounded linear map. In addition, (2.1) shows that
Tn → T , n → ∞.

Examples 2.7. In the following examples, the linearity of the operator under
consideration is easy to check.

(i) Let X be a normed space. Then the identity id : X → X is bounded and

k id k = 1.
(ii) Let 1 ≤ p ≤ ∞. The left shift and the right shift on `p are defined by

R : `p → `p , (x1 , x2 , x3 . . . )n∈N 7→ (0, x1 , x2 , . . . ),

L : `p → `p , (x1 , x2 , x3 . . . )n∈N 7→ (x2 , x3 , . . . ).
28 2.1. Bounded linear maps

Obviously, R and L are well-defined and linear. Moreover, R is an isome-

try because kRxkp = kxkp ; in particular kRk = 1.
The left shift is not an isometry because, e. g., kL(1, 0, 0, . . . )kp = k0kp =
0 < 1 = k(1, 0, 0, . . . )kp . It is easy to see that kLxkp ≤ kxkp , x ∈ `p ,
implying that kLk ≤ 1. Since kL(0, 1, 0, 0 . . . )kp = k(1, 0, 0 . . . )kp =
k(0, 1, 0, 0 . . . )kp we also have kLk ≥ 1, so that altogether kLk = 1.
Note that LR = id`p but RL 6= id`p .

(iii) T : C 1 ([0, 1], k · kC 1 ) → C([a, b], k · k∞ ), T x = x0 with kxkC 1 := kxk∞ +

kx0 k∞ . The operator T is bounded and kT k = 1.

Proof. The operator T is bounded with kT k ≤ 1 because kT xk∞ =

kx0 k∞ ≤ kxk∞ + kx0 k∞ ≤ kxkC 1 for all x ∈ X.
To proof that kT k ≥ 1 let xn : [0, 1] → R, xn (t) := n1 exp(−nt). Ob-
viously, xn ∈ C 1 ([0, 1]), kxn kC 1 = n1 + 1 and kT xn k∞ = 1. It follows
that

kT k = sup kT
 xk∞ 1
 kT xn k∞
kxkC 1 : x ∈ C ([0, 1]) \ {0} ≥ sup kxn kC 1 : n ∈ N

= sup 1+1 1 : n ∈ N = 1.

n

(iv) T : C 1 ([0, 1], k · k∞ ) → C([a, b], k · k∞ ), T x = x0 is not bounded.

1
Proof. As in the example above let xn : [0, 1] → R, xn (t) := n exp(−nt).
It follows that
 kT xk∞  kT xn k∞
sup kxk∞ : x ∈ C 1 ([0, 1]) \ {0} ≥ sup :n∈N
kxn k∞
= sup 11 : n ∈ N = ∞

n

Lemma 2.8. let X, Y be normed spaces, X finite-dimensional. Then every

linear map T : X → Y is bounded.

Proof. Let e1 , . . . , en be a basis of X. Since on X all norms are equivalent, we

can assume that
Xn X n
αj ej = |αj |.


j=1 j=1

Let M := max{kTPnej k : j = 1, . . . , n}. Then T is bounded with kT k ≤ M

because for x = j=1 αj ej ∈ X

Xn n
X n
X
kT xkY = αj T ej ≤ |αj | kT ej kY ≤ M |αj | = M kxkX .

Y
j=1 j=1 j=1
Chapter 2. Bounded maps; the dual space 29

Theorem 2.9. Let X, Y be normed spaces, Y a Banach space. Let D ⊆ X be a

dense subspace of X and T ∈ L(D, Y ). Then there exists exactly one continuous
extension Tb : X → Y of T . The extension is bounded with kTbk = kT k.

Proof. For x ∈ X choose a sequence (xn )n∈N ⊆ D which converges to x. The

sequence is a Cauchy sequence in D, hence, by the uniform continuity of T ,
(T xn )n∈N is a Cauchy sequence in Y , and therefore it converges in Y because Y
is complete. Let (ξn )n∈N be another Cauchy sequence in D which converges to
x. By what was said before, (T ξn ) converges in Y . Then lim kT xn − T ξn k =
n→∞
lim kT (xn − ξn )k ≤ lim kT k k(xn − ξn )k = kT k lim k(xn − ξn )k = 0, the
n→∞ n→∞ n→∞
following operator is well defined:

Te : X → Y, Tex := lim T xn for any (xn )n∈N ⊆ D which converges to x.

n→∞

It is not hard to see that Te is a linear extension of T and that kTek ≥ kT k. To

see that indeed equality holds, we only need to observe that by definition of Te

{kT xk : x ∈ D, kxk = 1} = {kTexk : x ∈ X, kxk = 1},

hence the suprema of both sets without the closure are equal (and equal to
the supremum of the closed sets). Since Te is linear and bounded by kT k, it is
continuous.
Assume that S is an arbitrary continuous extension of T . For x ∈ X and a
sequence (xn )n∈N ⊆ D which converges to x we find

Sx = lim Sxn = lim T xn = lim Tbxn = Tbx.

n→∞ n→∞ n→∞

Therefore, Tb is the unique continuous extension of T .

Finally we give a criterion for the invertibility of a bounded linear operator.

TheoremP2.10 (Neumann series). Let X be a normed space and T ∈ L(X)

∞
such that n=0 T n converges. Then id −T is invertible in L(X) and
∞
X
(id −T )−1 = T n. (2.2)
n=0

In particular, if X is a Banach space and kT k < 1, then id −T is invertible and

k(id −T )−1 k ≤ (1 − kT k)−1 .

Proof. The proof is analogous to the proofPfor the convergence of the geometric
m
series. We define the partial sums Sm := n=0 T n , m ∈ N0 . Then

(id −T )Sm = Sm (id −T ) = id −T m+1 , m ∈ N0 . (2.3)

Note that:
30 2.1. Bounded linear maps

P∞
(i) T m → 0 for m → ∞ because m=0 T m converges.
P∞
(ii) Sm → n=0 T n for m → ∞ by assumption.
(iii) For fixed R ∈ L(X) the maps L(X) → L(X), S 7→ RS and S 7→ SR
respectively are continuous.

Hence taking the limit m → ∞ in (2.3) gives

∞
X ∞
X 
(id −T ) Tn = T n (id −T ) = id
n=0 n=0

implying that id −T is invertible and that (2.2) holds.

P∞
Now assume that X is a Banach space and that kT k < 1. Then n=0 T n
Pm
converges in norm because kT n k ≤ kT kn . In particular, j=0 T
j
is
m∈N
a Cauchy sequence in L(X). Since L(X) is complete by assumption on X
and Theorem 2.6 the series converges. By the first part of the proof, id −T is
invertible and formula (2.2) holds.

Application 2.11 (Volterra integral equation). Let k ∈ C([0, 1]2 ) and

y ∈ C([0, 1]). We ask if the equation
Z s
x(s) − k(s, t)x(t) dt = y(s), s ∈ [0, 1]. (2.4)
0

has solution x ∈ C([0, 1]). If a solution exists, is it unique? Can the norm of
the solution be estimated in terms of y?

Solution. Note that equation (2.4) can be written as an equation in the Banach
space C([0, 1]):

x − Kx = y

where
Z s
K : C([0, 1]) → C([0, 1]), (Kx)(s) := k(s, t)x(t) dt, s ∈ [0, 1].
0

Obviously, K is a well-defined linear operator and for all x ∈ C([0, 1]), s ∈ [0, 1]
Z s  Z s
 
|Kx(s)| =  k(s, t)x(t) dt ≤ |k(s, t)| |x(t)| dt ≤ s kkk∞ kxk∞ ,
0 0
Z s Z t  Z sZ t
2
k(t, t1 )x(t1 ) dt1 dt ≤ kkk2∞ kxk∞
 
|K x(s)| =  k(s, t) dt1 dt
0 0 0 0
s2
= kkk2∞ kxk∞ .
2
Chapter 2. Bounded maps; the dual space 31

Repeating this process, it follows that

sn
|K n x(s)| ≤ kkkn∞ kxk∞ , s ∈ [0, 1], x ∈ C([0, 1]), n ∈ N,
n!
P∞
which shows that kK n k ≤ kkk ∞
n! . In particular, n=0 K
n
converges so that
id −K is invertible by Theorem 2.10. Hence equation (2.4) has exactly one
solution x ∈ C([0, 1]), given by
∞
X
x= K n y.
n=0
P
∞ P∞ P∞ kkk∞
Moreover, kxk∞ = n=0 K n y ≤ kK n k kyk∞ ≤ n! kyk∞ =

n=0 n=0
∞
e kkk∞ kyk∞ .

2.2 The dual space and the Hahn-Banach theo-

rem
Definition 2.12. Let X be a normed space. X 0 := L(X, K) is the dual space
of X; elements in the dual space are called functionals.

Note that in general the algebraic dual space, i. e., the space of all linear maps
X → K in general is larger than the topological dual space defined above.
Theorem 2.6 implies immediately:

Proposition 2.13. The dual space of a normed space X with the norm

kx0 k = sup{|x0 (x)| : x ∈ X, kxk ≤ 1}, x0 ∈ X 0 ,

is a Banach space.

Definition 2.14. Let X be normed space. p : X → R is a a seminorm if

(i) p(λx) = |λ|p(x), λ ∈ K, x ∈ X,

(ii) p(x + y) ≤ p(x) + p(y), x, y ∈ X.

A seminorm p is called bounded if there exists an M ∈ R such that

p(x) ≤ M kxk, x ∈ X.

If p satisfies

(i’) p(λx) = λp(x), λ ≥ 0, x ∈ X

instead of (i), then it is called a sublinear functional.

32 2.2. The dual space and the Hahn-Banach theorem

For example, every norm on X is sublinear. If ϕ : X → K is linear, then

X → R, x 7→ |ϕ(x)|

is sublinear.

Remark. Observe that p(x) ≥ 0 for every x ∈ X and every sublinear functional
p. Moreover, note that every seminorm is a sublinear functional.

The next fundamental theorem shows that every normed space admits non-
trivial functionals (except when X = {0}).

Theorem 2.15 (Hahn-Banach theorem). Let X be normed space and p :

X → R a seminorm. Let Y ⊆ X a subspace and ϕ0 a linear functional on Y
(that is, ϕ0 : Y → K linear) with

ϕ0 (y) ≤ p(y), y ∈ Y.

Then ϕ0 has an extension to a functional ϕ on X which satisfies

|ϕ(x)| ≤ p(x), x ∈ X. (2.5)

Proof. For Y = X there is nothing to show. Now assume Y 6= X. We dis-

tinguish between the real and the complex case. First assume that X is a real
vector space.
We divide the proof in two steps.
Step 1. Let z0 ∈ X \ Y and Z := span{z0 , Y }. We will show that ϕ0 can be
extended to some ψ ∈ Z 0 such that (2.5) holds for all z ∈ Z.
Obviously, every linear extension of ψ must be of the form

ψc (y + λz0 ) = ϕ0 (y) + λc, λ ∈ R, y ∈ Y

for some c ∈ R. We have to find c such that |ψc (z)| ≤ p(z), z ∈ Z, that is,

|ψc (y + λz0 )| ≤ p(y + λz0 ), y ∈ Y, λ ∈ R. (2.6)

By assumption on ϕ0

ϕ0 (x) − ϕ0 (y) = ϕ0 (x − y) ≤ p(x − y) ≤ p(x + z0 ) + p(y + z0 ), y, x ∈ Y,

implying

−ϕ0 (y) − p(y + z0 ) ≤ −ϕ0 (x) + p(x + z0 ), y, x ∈ Y,

so that

a := sup{−ϕ0 (x) − p(x + z0 ) : x ∈ Y } ≤ inf{−ϕ0 (x) + p(x + z0 ) : x ∈ Y } := b.

Chapter 2. Bounded maps; the dual space 33

Now let c ∈ [a, b] arbitrary. We show that then ψc is an extension of ϕ0 as

desired. Let z = y + λz0 ∈ Z with y ∈ Y and λ ∈ R. Obviously ψc is continuous
in 0, hence ψc ∈ Z 0 .
We have to show (2.6). For λ = 0 equation (2.6) clearly holds. For λ 6= 0:
 
λ > 0 : λc ≤ λb ≤ λ − ϕ0 ( λ1 y) + p( λ1 y + z0 ) = −ϕ0 (y) + p(y + λz0 ),
 
λ < 0 : λc ≤ λa ≤ λ − ϕ0 ( λ1 y) − p( λ1 y + z0 ) = −ϕ0 (y) + p(y + λz0 ).

In both cases we obtain ψc (z) = ψc (y + λz0 ) = ϕ0 (y) + λc ≤ p(y + λz0 ) = p(z).

Application to −z yields −ψc (z) = ψc (−z) ≤ p(−z) = p(z). In summary, we
have |ψ(z)| ≤ p(z), z ∈ Z.
Step 2. Let Φ be the set of all proper extensions of ϕ0 such that |ϕ(x)| ≤ p(x)
for all x ∈ D(ϕ) (the domain of ϕ). By Step 1, Φ is not empty and partially
ordered by
ϕ1 < ϕ2 ⇐⇒ ϕ2 is an extension of ϕ1 .
Every totally ordered subset Φ0 has the upper bound
[
D(f ) = D(ψ), f (x) = ψ(x) for x ∈ D(ψ).
ψ∈Φ0

By Zorn’s lemma, Φ contains a maximal element ϕ. This ϕ is defined on X

because otherwise, by Step 1, it would not be maximal.
Now we assume that X is a complex vector space. Consider X as a vector space
over R and define the functional
V0 : Y → R, V0 (y) = Re(ϕ(y)).
It is R-linear because for all x, y ∈ Y and α ∈ R
V0 (αx + y) = Re(ϕ0 (αx + y)) = Re(αϕ0 (x) + ϕ0 (y)) = α Re(ϕ0 (x)) + Re(ϕ0 (y))
= αV0 (x) + V0 (y).
In addition, V0 is bounded by the sublinear functional p
|V0 (y)| = | Re(ϕ0 (y))| ≤ |ϕ0 (y)| ≤ p(y), y ∈ Y.
By what we have already shown, there exists an R-linear extension V ∈ L(X, R)
of V0 with |V (x)| ≤ p(x), x ∈ X. Now define
ϕ : X → C, ϕ(x) = V (x) − iV (ix).
ϕ has the following properties:
(i) ϕ is an extension of ϕ0 . To see this, let y ∈ Y .
ϕ(y) = V0 (y) − iV0 (iy) = Re(ϕ0 (y)) − i Re(ϕ0 (iy)) = Re(ϕ0 (y)) − i Re(iϕ0 (y))
= Re(ϕ0 (y)) + i Im(ϕ0 (y)) = ϕ0 (y).
34 2.2. The dual space and the Hahn-Banach theorem

(ii) ϕ is C-linear. To show this, let x, y ∈ X and ζ = a + ib with a, b ∈ R.

ϕ(x + y) = V (x + y) − iV (i(x + y)) = V (x) + V (y) − iV (ix) − iV (iy)

= ϕ(x) + ϕ(y),

ϕ(ζx) = ϕ(ax) + ϕ(ibx) = V (ax) − iV (iax) + V (ibx) − iV (i2 bx)

= a[V (x) − iV (ix)] + b[V (ix) + iV (x)]
= (a + ib)[V (x) − iV (ix)] = ζϕ(x).

(iii) ϕ is bounded by p. To prove this, let x ∈ X and α ∈ R such that

|ϕ(x)| = eiα ϕ(x) = Re ϕ(eiα x) = V (eiα x) ≤ p (eiα x) = p(x).

In conclusion, ϕ is a C-linear continuous extension of ϕ0 which is bounded by p

as desired.

Remark. If in the Hahn-Banach theorem we consider only real normed spaces

and replace the seminorm p by a sublinear functional such that ϕ0 (y) ≤ q(y)
for all y ∈ Y , then ϕ0 can be extented to a functional ϕ : X → K such that
−q(x) ≤ ϕ(x) ≤ q(x) for all x ∈ X, see [Rud91, Theorem 3.2].

The Hahn-Banach theorem has some important corollaries.

Corollary 2.16. Let X be a normed space, Y ⊆ X a subspace and ϕ0 ∈ Y 0 .

Then there exists an extension ϕ ∈ X 0 of ϕ0 such that kϕk = kϕ0 k.

Proof. The map p : X → R, p(x) = kϕ0 k kxk is a sublinear functional on X and

|ϕ0 (y)| ≤ kϕ0 (y)k kyk = p(y) for all y ∈ Y . By the Hahn-Banach theorem, ϕ0
can be extended to a ϕ ∈ X 0 with |ϕ(x)| ≤ p(x) = kϕ0 k kxk, so that kϕk ≤ kϕ0 k.
On other hand kϕk ≥ kϕ0 k holds because ϕ0 is a restriction of ϕ.
The next corollary shows that X 0 does not consist only of the trivial functional
and that it separates points in X.

Corollary 2.17. Let X be a normed space, x ∈ X, x 6= 0. Then there exists a

ϕ ∈ X 0 such that ϕ(x) = kxk. In particular for all x, y ∈ X:
(i) x = 0 ⇐⇒ ∀ϕ ∈ X 0 ϕ(x) = 0,
(ii) x 6= y =⇒ ∃ ϕ ∈ X 0 ϕ(x) 6= ϕ(y).

Proof. Let Y := span{x} and ϕ0 ∈ Y 0 defined by ϕ0 (λx) = λkxk. Then ϕ0 (x) =

kxk and kϕ0 k = 1. By Corollary 2.16 there exists an extension ϕ ∈ X 0 of ϕ0
with the desired properties. Statement (i) is clear; (ii) follows when (i) is applied
to x − y.

Corollary 2.18. Let X, Y be a normed spaces.

Chapter 2. Bounded maps; the dual space 35

(i) kxk = sup{ϕ(x) : ϕ ∈ X 0 , kϕk = 1}, x ∈ X.

(ii) For T : X → Y linear

kT k = sup{ϕ(T x) : x ∈ X, kxk = 1, ϕ ∈ Y 0 , kϕk = 1}.

Proof. (i) For all ϕ ∈ X 0 with kϕk = 1: kxk = kϕk kxk ≥ |ϕ(x)|, hence
kxk ≥ sup{ϕ(x) : ϕ ∈ X 0 , kϕk = 1}. To show that in fact we have equality, we
recall that by Corollary 2.17 there exists a ϕ ∈ X 0 with kϕk = 1 and ϕ(x) = kxk.
Hence the formula in (i) is proved. Note the the supremum is in fact a maximum.

(ii) Let M := sup{ϕ(T x) : x ∈ X, kxk = 1, ϕ ∈ Y 0 , kϕk = 1}. We have to show

M = kT k. Obviously, M = ∞ if and only if kT k = ∞. Now assume kT k < ∞.
Let ε > 0. Then there exists an x ∈ X with kxk = 1 and kT xk ≥ kT k − ε.
Choose a ϕ ∈ X 0 such that kϕk = 1 and ϕ(T x) = kT xk. Then M ≥ ϕ(T x) =
kT k − ε. Since ε is arbitrary, it follows that M ≥ kT k. The revers inequality
follows from

ϕ(T x) ≤ kϕk kT xk ≤ kϕk kT k kxk = kT k, x ∈ X, kxk = 1, ϕ ∈ X 0 , kϕk = 1.

Corollary 2.19. Let X be a normed space, Y ⊆ X a closed subspace. For

every x0 ∈ X \ Y exists ϕ ∈ X 0 such that ϕ|Y = 0 and ϕ(x0 ) = 1.

Proof. Let π : X → X/Y be the canonical projection. Then π(y) = 0, y ∈ Y ,

and π(x0 ) 6= 0. Since X is a normed space by Example 1.13, there exists a
ψ ∈ (X/Y )0 such that ϕ(π(x0 )) 6= 0 and ϕ(π(x0 )) = 1. Obviously ϕ = ψ◦π ∈ X 0
and has the desired properties.

Corollary 2.20. Let X be a normed space, Y ⊆ X a subspace. Then the

following are equivalent:

(i) Y = X,
ϕ ∈ X 0.


(ii) ϕ|Y = 0 =⇒ ϕ = 0 ,

Theorem 2.21. Let X be a normed space.

X 0 separable =⇒ X separable.

Proof. By Proposition 1.9 the unit sphere SX 0 := {x0 ∈ X 0 : kx0 k = 1} is

separable. Choose dense subset {x0n : n ∈ N} of SX 0 . and xn ∈ SX := {x ∈ X :
kxk = 1} with kx0n (xn )k > 21 . Let U = span{xn : n ∈ N}. We will show U = X.
Assume this is not true. By Corollary 2.19 there exists an x0 ∈ SX 0 such that
x0 6= 0 and x0 |U = 0. Let n ∈ N such that kx0n − x0 k < 14 . This leads to the
contradiction
1 1
≤ |x0n (xn )| ≤ |x0n (xn ) − x0 (xn )| + |x0 (xn )| ≤ kx0n − x0 k + |x0 (xn )| < .
2 4
36 2.3. Examples of dual spaces

2.3 Examples of dual spaces

Theorem 2.22. (i) Let 1 ≤ p < ∞ and q such that
1 1
+ =1
p q
1
with the convention ∞ = 0. q is called the Hölder conjugate of p.
The following map is an isometric isomorphism:
∞
X
T : `q → (`p )0 , (T x)y = xn yn for x = (xn ) ∈ `q , y = (yn ) ∈ `p .
n=0

(ii) The following map is an isometric isomorphism:

∞
X
T : `1 → (c0 )0 , (T x)y = xn yn for x = (xn ) ∈ `1 , y = (yn ) ∈ c0 .
n=0

Proof. (i) Let 1 < p < ∞. T is well-defined by Hölder’s inequality and

∞
X 
|(T x)y| =  xn yn  ≤ kxkq kykp .
n=0

Linearity and injectivity of T is clear. The inequality above gives

kT xk ≤ kxkq , x ∈ `q . (2.7)

It remains to show surjectivity of T and that kT xk ≥ kxk, x ∈ `q . To this end,

let y 0 ∈ (`p )0 and set xn := y 0 (en ), n ∈ N, where en is the nth unit vector in `p .
We will show that x := (xn )n∈N ∈ `q and that T x = y 0 . For y 0 = 0 this is clear.
Now assume that y 0 6= 0. For n ∈ N define
|xn |q
(
tn := xn , xn 6= 0,
0, xn = 0.

Using pq − p = q we find
N
X N
X
|tn |p = |xn |p(q−1) , N ∈ N.
n=1 n=1

Hence, for all N ∈ N,

N
X N
X N
X N
X N
X
|xn |q = tn y 0 (en ) = y 0 tn en ≤ ky 0 k


xn tn = tn en p
n=1 n=1 n=1 n=1 n=1
N N
X 1 X 1
= ky 0 k |tn |p ≤ ky 0 k |xn |q

p

p
.
n=1 n=1
Chapter 2. Bounded maps; the dual space 37

For N large enough, the last factor in the line above is not zero, so, using
1 − p1 = 1q , we obtain
N
X
q1
|xn |q ≤ ky 0 k
n=1

implying that x ∈ `q . Since (T x) en = xn en = y 0 en , n ∈ N, and {en : n ∈ N}

a total subset of `p , it follows that T x = y 0 . In particular, with the inequality
above, kxkq ≤ ky 0 k = kT xk. Together with (2.7) it follows that kT xk = kxk,
that is, T is an isometry.
The proof for p = 1 is similar.
(ii) Well-definedness and injectivity of T are clear. Moreover kT xk ≤ kxk1 for
every x ∈ `1 because
∞ ∞
X  X
 xn yn ≤ kyk∞
 |xn | = kyk∞ kxk1 , y ∈ c0 , x ∈ `1 .
n=0 n=0

To show that T is surjective, let y 0 ∈ (c0 )0 and let xn := yn (en ) where en is

the nth unit vector in c0 . For n ∈ N choose αn ∈ R such that |y 0 (en )| =
exp(iαn )y 0 (en ). It follows that
∞
X ∞
X ∞
X ∞
X
|y 0 (en )| = exp(iαn ) y 0 (en ) = y 0


|xn | = exp(iαn ) en
n=0 n=0 n=0 n=0
∞
X
0
exp(iαn ) en ∞ = ky 0 k.

≤ ky k
n=0

Hence x ∈ `1 and kxk1 ≤ ky 0 k. As before, since {en : n ∈ N} is a total subset

of c0 , it follows that T x = y 0 and the proof is complete. (Note however, that
{en : n ∈ N} is not dense in `∞ .)
The theorem above shows that
(`p )0 ∼
= `q , 1 ≤ p < ∞,
0 ∼
(c0 ) = `1 .

Remark. Note that (`∞ )0  `1 . To see this, assume that (`∞ )0 ∼

= `1 . Since `1 is
separable, Theorem 2.21 would imply that also `∞ is separable, in contradiction
to Example 1.26.

Other important examples are given without proof in the following theorems.

Theorem 2.23. Let (Ω, Σ, µ) be a σ-finite measure space. Let 1 ≤ p < ∞ and
q such that p1 + 1q = 1. Then
Z
T : Lq (Ω) → (Lp (Ω))0 , (T f )(g) = f g dµ, f ∈ Lq (Ω), g ∈ Lp (Ω),
Ω
38 2.4. The Banach space adjoint and the bidual

is an isometric isomorphism.

Theorem 2.24 (Riesz’s representation theorem). Let K be a compact

metric space and M (K) the set of all regular Borel measures with finite varia-
tion, that is kµk < ∞ with
nX o
kµk := sup |µ(V )| : Z partition of K in pairwise disjoint measurable sets .
V ∈Z

1 1
Let 1 ≤ p < ∞ and q such that p + q = 1. Then
Z
T : M (K) → (C(K))0 , (T µ)(g) = g dµ, µ ∈ M (K), g ∈ C(K),
Ω

is an isometric isomorphism.

For a proof, see [Rud87, Theorem 6.19].

The theorems above show that

(Lp )0 ∼
= Lq , 1 ≤ p < ∞,
0 ∼
(C(K)) = M (K).

2.4 The Banach space adjoint and the bidual

Definition 2.25. Let X, Y be normed spaces and T ∈ L(X, Y ). The Banach
space adjoint of T is

T 0 : Y 0 → X 0, (T 0 y 0 )x := y 0 (T x), y 0 ∈ Y 0 , x ∈ X.

Obviously, T 0 is linear and continuous as composition of continuous functions,

hence T 0 ∈ L(Y 0 , X 0 ) and the following diagram commutes

T
X Y

x0 =y 0 ◦T y0

Theorem 2.26. Let X, Y, Z be normed spaces.

(i) The map L(X, Y ) → L(Y 0 , X 0 ), T 7→ T 0 , is linear and isometric, that is,
kT 0 k = kT k. In general, it is not surjective.
(ii) (ST )0 = T 0 S 0 for S ∈ L(Y, Z) and T ∈ L(X, Y ).
Chapter 2. Bounded maps; the dual space 39

Proof. (i) Linearity of T 7→ T 0 is clear. Immediately by the definition of T 0 we

have that

kT 0 y 0 k = ky 0 ◦ T k ≤ ky 0 k kT k, y0 ∈ Y 0 ,

hence kT 0 k ≤ kT k. By Corollary 2.18 kT k is

kT k = sup{y 0 (T x) : x ∈ X, kxk = 1, y 0 ∈ Y 0 , ky 0 k = 1}.

For every ε > 0 there exist x ∈ X, kxk = 1, y 0 ∈ Y 0 such that kT k − ε <

y 0 (T x) = (T 0 y 0 )(x) ≤ kT 0 k ky 0 k kxk = kT 0 k , so kT k ≤ kT 0 k.
(ii) For all z 0 ∈ Z 0 and x ∈ X we have ((ST )0 z 0 )(x) = z 0 (ST (x)) = z 0 (S(T x)) =
(S 0 z 0 )(T x) = T 0 (S 0 z 0 )x = (T 0 S 0 )(z 0 )(x), hence (ST )0 = T 0 S 0 .

Example 2.27. Let 1 ≤ p < ∞. The adjoint of the left shift

L : `p → `p , L(x1 , x2 , x3 , . . . ) = (x2 , x3 , . . . )

is the right shift.

Proof. Let 1
p + 1
q = 1 and y = (yn )n∈N ∈ lq ∼
= (lp )0 . Then for all x = (xn )n∈N ∈
lp :
∞
X ∞
X ∞
X ∞
X
(L0 y)x = y(Lx) = yn (Lx)n = yn xn+1 = yn−1 xn = (Ry)n xn
n=1 n=1 n=2 n=2
∞
X
= (Ry)n xn = (Ry)x.
n=1

Definition 2.28. Let X be a normed space. X 00 := (X 0 )0 is the bidual of X.

For every x ∈ X the linear map

JX (x) : X 0 → K, JX (x)x0 := x0 x

is linear and bounded by kxk, hence JX (x) ∈ X 00 .

Theorem 2.29. The map

JX : X → X 00 , JX (x)x0 = x0 x, x0 ∈ X 0

is a linear isometry. In general, it is not surjective.

Proof. We have seen above that JX is well-defined, linear and kJX (x)k ≤ kxk,
x ∈ X. Now let x ∈ X and choose ϕx ∈ X 0 such that ϕx (x) = kxk (Corol-
lary 2.17). It follows that kJX (x)ϕx k = |ϕx (x)| = kxk, hence kJX (x)k ≥ 1.
The preceding theorem gives another easy proof that every normed space X can
be completed (see Theorem 1.7).
40 2.4. The Banach space adjoint and the bidual

Corollary 2.30. Every normed space is isometrically isomorphic to a dense

subspace of a Banach space.

Proof. By the theorem above, X is isometrically isomorphic to JX (X) ⊆ X 00 .

Since X 00 is complete (Theorem 2.6), the closure JX (X) is a Banach space.

Definition 2.31. A Banach space is called reflexive if JX is surjective.

Examples 2.32. (i) Every finite-dimensional normed space is reflexive.

(ii) `p is reflexive for 1 < p < ∞ by Theorem 2.22.
(iii) c0 and `1 are not reflexive.

Note that there are non-reflexive Banach spaces X such that X ∼

= X 00 (but JX
is not surjective).

Lemma 2.33. Let X, Y be normed spaces and T ∈ L(X, Y ). Then T 00 ◦ JX =

JY ◦ T , that is, the following diagram commutes:

T
X Y

JX JY

X 00 Y 00
T 00

Proof. For x ∈ X and y 0 ∈ Y 0

[T 00 (JX (x))](y 0 ) = (JX (x))(T 0 y 0 ) = T 0 y 0 x = y 0 (T x) = (JY (T x))y 0 = [(JY ◦ T )(x)]y 0 .

If X and Y are identified with subspaces of X 00 and Y 00 via the canonical maps
JX and JY , then T 00 is an extension of T . Note that with this identification
S ∈ L(Y 0 , X 0 ) is adjoint operator of some T ∈ L(X, Y ) if and only if S 0 (X) ⊆ Y .

0
Lemma 2.34. Let X be a normed space. Then JX ◦ JX 0 = idX 0 .

Proof. Note that JX 0 : X 0 → X 000 and JX

0
: X 000 → X 0 . For x ∈ X, x0 ∈ X 0
0
[(JX ◦ JX 0 )x0 ](x) = [JX 0 x0 ](JX (x)) = [JX x]x0 = x0 x.

Theorem 2.35. (i) Every closed subspace of a reflexive normed space is re-
flexive.
(ii) A Banach space X is reflexive if and only if X 0 is reflexive.
Chapter 2. Bounded maps; the dual space 41

Proof. (i) Let U be a closed subspace of a reflexive normed space X and let
u00 ∈ U 00 . We have to find a u ∈ U such that JX (u) = u00 . Let x000 : X 0 →
K, x000 (x0 ) = u00 (x0 |U ). Obviously, x000 is linear and bounded because

|x000 (x0 )| = |u00 (x0 |U )| ≤ ku00 kkx0 |U k ≤ ku00 kkx0 k,

hence x000 ∈ X 00 . Since X is reflexive there exists an x0 ∈ X such that JX (x0 ) =

x000 . Assume that x0 ∈ / U . Since U is closed, there exists a ϕ ∈ X 0 such that
ϕ|U = 0 and ϕ(x0 ) = 1 (Corollary 2.19). On the other hand ϕ(x0 ) = 0 by choice
of x0 because

x0 (x0 ) = x000 (x0 ) = JX (x0 )x0 = u00 (x0 |U ), x0 ∈ X 0 ,

Therefore x0 ∈ U . It remains to be shown that JU (x0 ) = u00 , that is

u00 (u0 ) = u0 (x0 ), u0 ∈ U 0 .

Let u0 ∈ U 0 and choose an arbitrary extension ϕ ∈ X 0 (Corollary 2.16). By

definition of x0 it follows that

u00 (u0 ) = u00 (ϕ|U ) = x000 (ϕ) = ϕ(x0 ) = u0 (x0 ).

(ii) Let X be reflexive. We have to show that JX 0 : X 0 → X 000 is surjective. Let

x000 000 0 0 000
0 ∈ X . The map x0 : X → K, x0 (x) = x0 (JX (x)) is linear and bounded,
hence x0 ∈ X . We will show that JX 0 (x0 ) = x000
0 0 0 00 00
0 . Let x ∈ X . Since X is
00
reflexive, there exists an x ∈ X such that JX (x) = x . Therefore

JX 0 (x00 )x00 = x00 (x00 ) = JX (x)(x00 ) = x00 x = x000 000 00

0 (JX (x)) = x0 (x ),

hence indeed JX 0 (x00 ) = x000

0 .
Now assume that X 0 is reflexive. By what was is already proved, X 00 is reflexive.
Since X is a closed subspace of X 00 via the canonical map JX , X is reflexive by
part (i) of the theorem.

Corollary 2.36. A reflexive normed space X is separable if and only if X 0 is

separable.

Proof. That separability of X 0 implies separability of X was shown in The-

orem 2.21. If X is separable and reflexive, then also X 00 is separable. By
Theorem 2.35 X 0 is reflexive, so we can again apply Theorem 2.21 to obtain
that X 0 is separable.

Definition 2.37. Let X be a normed space. A sequence (xn )n∈N converges

weakly to x0 ∈ X if and only if

lim x0 (xn ) = x0 (x0 ), x0 ∈ X 0 .

n→∞

w
Notation: xn −
→ x or w- lim xn = x.
n→∞
42 2.4. The Banach space adjoint and the bidual

If it should be emphasised that a sequence converges with respect to the norm

in the given Banach space, then the sequence is called norm convergent. Some-
times the notion strongly convergent is used. Note, however, that in spaces
of linear operators the term “strong convergence” has another meaning (see
Defintion 3.12).
The next remark shows that strong convergence is indeed stronger than weak
convergence.

Remarks 2.38. (i) If the weak limit of a sequence exists, then it is unique,
because, by the Hahn-Banach theorem, the dual space separates points (Corol-
lary 2.17).
(ii) Every convergent sequence is weakly convergent with the same limit.
(iii) A weakly convergent sequence is not necessarily convergent. Consider for
example the sequence of the unit vectors (en )n∈N in c0 . Let ϕ ∈ c00 ∼
= `1 . Then
lim ϕ(en ) = 0 but the sequence of the unit vectors does not converge in norm.
n∈N

Example 2.39. Let (xn )n∈N be a bounded sequence in C([0, 1]). Then the
following is equivalent:
(i) (xn )n∈N converges weakly to y ∈ C[(0, 1)].
(ii) (xn )n∈N converges pointwise to y ∈ C[(0, 1)].
Proof. “(i) =⇒ (ii)” It is easy to see that for every t0 ∈ [0, 1] the point
evaluation x 7→ x(t0 ) is a bounded linear functional. Hence for all t ∈ [0, 1] the
sequence (xn (t)n∈N converges to some y(t). By assumption, [0, 1] → K, t 7→ y(t)
belongs to C([0, 1]).
“(ii) =⇒ (i)” follows from Riesz’s representation theorem (Theorem 2.24) and
the Lebesgue convergence theorem (see A.19).

Theorem 2.40. Every bounded sequence in a reflexive normed space contains

a weakly convergent subsequence.

Proof. Let X be a reflexive normed space and x = (xn )n∈N ⊆ X be a bounded

sequence.
First we assume that X is separable. By theorem 2.36, also X 0 is separable.
Let {ϕn : n ∈ N} be a dense subset of X 0 . We will construct a subsequence
y = (yn )n∈N of x such that for every j ∈ N the sequence (ϕj (yn ))n∈N converges.
The sequence (ϕ1 (xn ))n∈N is bounded, so it contains a convergent subsequence

(ϕ1 (xn1 ,1 ), ϕ1 (xn1 ,2 ), ϕ1 (xn1 ,3 ), . . . )

Now the sequence (ϕ2 (xn1 ,j ))j∈N is bounded, so it contains a convergent subse-
quence

(ϕ2 (xn2 ,1 ), ϕ2 (xn2 ,2 ), ϕ2 (xn2 ,3 ), . . . )

Chapter 2. Bounded maps; the dual space 43

Continuing like this, we obtain a sequence of subsequences xnm = (xnm ,j )j∈N ,

m ∈ N such that (ϕm (xnm ,j ))j∈N converges. Now the “diagonal sequence” y
with ym := xnm ,m has the desired property.
Now we will show that y is weakly convergent. Let x0 ∈ X 0 and ε > 0. Choose
an k ∈ N such that kx0 − ϕk k < 4M ε
where M := sup{kxn k : n ∈ N} < ∞. Let
N ∈ N such that |ϕk (yn ) − ϕk (ym )| < 2ε , m, n ≥ N . It follows for m, n ≥ N :

|x0 (yn ) − x0 (ym )| ≤ |x0 (yn ) − ϕk (yn )| + |ϕk (yn ) − ϕk (ym )| + |ϕk (ym ) − x0 (ym )|
≤ 2M kx0 − ϕk k + |ϕk (yn ) − ϕk (ym )|
ε ε
< + = ε.
2 2
This implies that (x0 (yn ))n∈N is a Cauchy sequence in K, hence it converges. To
show that (yn )n∈N converges weakly, define the map

ψ : X 0 → K, ψ(x0 ) = lim x0 (yn ).

n→∞

By what is already shown, ψ is well-defined and linear. It is also bounded

because

|ψ(x0 )| =  lim x0 (yn ) = lim |x0 (yn )| ≤ lim kx0 k k(yn )k ≤ M kx0 k.
 
n→∞ n→∞ n→∞

Hence ψ ∈ X 00 . Since X is reflexive, there exists a y0 ∈ X such that x0 (y0 ) =

ψ(x0 ) = lim x0 (yn ). Hence (yn )n∈N converges weakly to y0 .
n→∞

Now assume that X is not separable. Let Y := span{xn : n ∈ N} where (xn )n∈N
is the bounded sequence in X chosen at the beginning of the proof. Y is sep-
arable (Theorem 1.25) and reflexive (Theorem 2.35). Hence, by the first step
of the proof, there exists a subsequence (yn )n∈N ⊆ Y of (xn )n∈N and a y0 such
w
that yn −→ y0 in Y . Let x0 ∈ X 0 . Then x0 |Y 0 ∈ Y 0 , hence lim x0 (yn ) =
n→∞
w
lim x0 |Y 0 (yn ) = x0 |Y 0 (y0 ) = x0 (y0 ). Therefore we also have yn −
→ y0 in X.
n→∞
Chapter 3. Linear operators in Banach spaces 45

Chapter 3

Linear operators in Banach

spaces

3.1 Baire’s theorem

Theorem 3.1 (Baire-Hausdorff). Let (X, d) be a complete
T∞ metric space and
(An )n∈N be a family of open dense subsets of X. Then n=1 An is dense in X.

Taking complements, it is easily seen that the theorem above implies

Theorem. Let (X, d) be a complete

S∞ metric space and (Bn )n∈N be a family of
closed subsets of X such that n=1 Bn contains an open subset. Then at least
one of the sets Bn contains a non-empty open subset.

Proof of Theorem 3.1. For r > 0 and x ∈ X let B(x, r) := {ξ ∈ X : kx−ξk < r}.
We have to show that any open ball in X has non-empty intersection with
T
n∈N An . Let ε > 0 and x0 ∈ X.
A1 is open and dense in X, hence A1 ∩ B(x0 , ε) is open and not empty. Hence
there exist ε1 ∈ (0, 2−1 ε) and x1 ∈ A1 such that B(x1 , ε1 ) ⊆ A1 ∩ B(x0 , ε),
hence

B(x1 , ε21 ) ⊆ B(x1 , ε1 ) ⊆ A1 ∩ B(x0 , ε).

A2 is open and dense in X, hence A2 ∩ B(x1 , ε21 ) is open and not empty. Hence
there exist ε2 ∈ (0, 2−2 ε) and x2 ∈ A2 such that B(x2 , ε2 ) ⊆ A2 ∩ B(x1 , ε21 ),
hence

B(x2 , ε22 ) ⊆ B(x2 , ε2 ) ⊆ A2 ∩ B(x1 , ε21 ) ⊆ A2 ∩ A1 ∩ B(x0 , ε1 ).

In this way we obtain sequences (εn )n∈N and (xn )n∈N with 0 < εn < 2−n ε and

B(xn , ε2n ) ⊆ B(xn , εn ) ⊆ An ∩ B(xn−1 , εn−1 ) ⊆ An−1 ∩ . . . A2 ∩ A1 ∩ B(x0 , ε1 ).

(3.1)
46 3.2. Uniform boundedness principle

Observe that xn ∈ B(xN , ε2N ) for N ∈ N and n ≥ N . This implies that (xn )n∈N
is a Cauchy sequence in X because, for fixed N ∈ N and all n, m > N we obtain
d(xm , xn ) ≤ d(xm , xN )+d(xn , xN ) < 2−N +1 . Since X is complete, y := lim xn
n→∞
exists and x0 ∈ B(xN , εN ) for every N ∈ N because for fixed N , we have that
xn ∈ B(xN , ε2N ) if n ≥ N . Hence (3.1) implies

y ∈ B(xN , ε2N ) ⊆ B(xN −1 , εN −1 ) ⊆ AN −1 ∩ . . . A2 ∩ A1 ∩ B(x0 , ε1 ), N ≥ 2,

T
so y ∈ n∈N An ∩ B(x0 , ε).

Definition 3.2. Let (X, d) be a metric space.

• A ⊆ X is called nowhere dense in X, if A does not contain an open set.

• A ⊆ X is of first category if it is the countable union of nowhere dense
sets.
• A ⊆ X is of second category if it is not of first category.

Note that A is nowhere dense if and only if X \ A is dense in X.

An equivalent formulation of

Theorem 3.3 (Baire’s category theorem). A complete metric space is of

second category in itself.

Examples 3.4. Q is of first category in R. R is of second category in R.

3.2 Uniform boundedness principle

Definition 3.5. Let (X, d) be a metric space. A family F = (fλ )λ∈Λ of maps
X → R is called uniformly bounded if there exists an M ∈ R such that

|fλ (x)| ≤ M, x ∈ X, λ ∈ Λ.

The next theorem shows that a family of pointwise bounded continuous func-
tions on a complete metric space is necessarily uniformly continuous on a certain
ball.

Theorem 3.6 (Uniform boundedness principle). Let X be a complete met-

ric space, Y a normed space and F ⊆ C(X, Y ) a family of continuous functions
which is pointwise bounded, i. e.,

∀x ∈ X ∃ Cx ≥ 0 ∀f ∈ F kf (x)k < Cx .

Then there exists an M ∈ R, x0 ∈ X and r > 0 such that

∀ x ∈ Br (x0 ) ∀f ∈ F kf (x)k < M. (3.2)

Chapter 3. Linear operators in Banach spaces 47

Proof. For n ∈ N let

\
An := {x ∈ X : kf (x)k ≤ n}.
f ∈F

Note that for every n ∈ N the set {x ∈ X : kf (x)k ≤ n} is closed because f and
k · k are continuous. Since all An are intersections of closed sets, they are closed.
Let x ∈ X. Since F is pointwise bounded, there exists an nx ∈ N such that
x ∈ Anx , hence X ⊆ ∪n∈N An . By Baire’s theorem exists an N ∈ N, x0 ∈ X,
r > 0 such that Br (x0 ) ⊆ AN , that is, (3.2) is satisfied with M = N .
The Banach-Steinhaus theorem is obtained in the special case of linear bounded
functions.

Theorem 3.7 (Banach-Steinhaus theorem). Let X be a Banach space, Y

a normed space and F ⊆ L(X, Y ) a family of continuous linear functions which
is pointwise bounded, i. e.,

∀x ∈ X ∃ Cx ≥ 0 ∀f ∈ F kf (x)k < Cx .

Then there exists an M ∈ R such that

kf k < M, f ∈ F.

Proof. By the uniform boundedness principle there exists an open ball Br (x0 ) ⊆
X and an M 0 ∈ R such that kf (x)k < M 0 for all x ∈ Br (x0 ) and f ∈ F. For
x ∈ X with kxk = 1 and f ∈ F we find
1 1
kf (x)k = kf (rx)k = kf (x0 ) − f (x0 − rx)k
r r
1 2M 0
≤ (kf (x0 )k + kf (x0 − rx)k) ≤ =: M,
r | {z } r
∈Br (x0 )

showing that F is uniformly bounded by M .

Corollary 3.8. Let X be a normed space and A ⊆ X. Then the following are
equivalent:
(i) A is bounded.
(ii) For every x0 ∈ X 0 the set {x0 (a) : a ∈ A} is bounded.

Proof. “(i) =⇒ (ii)” is clear.

“(ii) =⇒ (i)” The family (JX (a))a∈A ⊆ X 00 is pointwise bounded by assump-
tion. By the Banach-Steinhaus theorem there exists a M ∈ R such that

kak = kJX (a)k ≤ M, a ∈ A.

Hence A is bounded.
48 3.2. Uniform boundedness principle

Corollary 3.9. Every weakly convergent sequence in a normed space is bounded.

Proof. Let X be a normed space and (xn )n∈N be a weakly convergent sequence
in X. By hypothesis, for every x0 ∈ X 0 the set {x0 (xn ) : n ∈ N} is bounded.
Therefore, by Corollary 3.8, the set {xn : n ∈ N} is bounded.

The following theorem follows directly from Theorem 2.40 and Corollary 3.9.

Theorem 3.10. Let (X, k · k) be a normed space, (xn )n∈N and x0 ∈ X. Then
the following is equivalent:

(i) x0 = w- lim xn .
n→∞

(ii) (xn )n∈N is bounded and there exists a total subset M 0 ⊆ X 0 such that

lim f (xn ) = f (x0 ), f ∈ M 0.

n→∞

Corollary 3.11. Let X be Banach space and A0 ⊆ X 0 . Then the following is

equivalent:

(i) A0 is bounded.
(ii) For all x ∈ X the set {a0 (x) : a0 ∈ A0 } is bounded.

Proof. The implication “(i) =⇒ (ii)” is clear. The other direction follows di-
rectly from the Banach-Steinhaus theorem.

Note that for “(ii) =⇒ (i)” the assumption that X is a Banach space is necessary.
For example, let d = {x = (xn )n∈N : xn 6= 0 for at most finitely many n} ⊆ `∞ .
d is a non-complete normed space (see Example 1.15 (5)). For m ∈ N define the
linear function ϕm : d → K by ϕm (en ) = mδm,n where δm,n is the Kronecker
delta. Obviously ϕm ∈ d0 and kϕm k = m, hence the family (ϕm ) is not bounded
in d0 , but for every fixed x ∈ d the set {ϕm (x) : m ∈ M } is.

Definition 3.12. Let X, Y be normed spaces, (Tn )n∈N ∈ L(X, Y ) a sequence

of bounded linear operators and T ∈ L(X, Y ).

(i) (Tn )n∈N converges to T , denoted by lim Tn = T , if and only if

n→∞

lim kTn − T k = 0.
n→∞

s
(ii) (Tn )n∈N converges strongly to T , denoted by s- lim Tn = T or Tn −
→ T , if
n→∞
and only if

lim kTn x − T xk = 0, x ∈ X.
n→∞
Chapter 3. Linear operators in Banach spaces 49

w
(iii) (Tn )n∈N converges weakly to T , denoted by w- lim Tn = T or Tn −
→ T , if
n→∞
and only if

lim |ϕ(Tn x) − ϕ(T x)| = 0, x ∈ X, ϕ ∈ Y 0 .

n→∞

Remark. (i) The limits are unique if they exist.

(ii) Convergence in norm implies strong convergence and the limits are equal.
Strong convergence implies weak convergence and the limits are equal.
The reverse implications are not true:

• Let X = `2 (N), Tn : X → X, Tn x = (x1 , . . . , xn , 0, . . . ) for x = (xm )m∈N .

Then T converges strongly to id but kTn − id k = 1 for all n ∈ N, so that
(Tn )n∈N does not converge to id in norm.
• Let X = `2 (N), Tn : X → X, Tn x = (0, . . . , 0, x1 , x2 , . . . ) (n leading
zeros) for x = (xm )m∈N . Then T converges weakly to 0 but kTn xk = 1 for
all n ∈ N, so that (Tn )n∈N does not converge strongly to 0.

Proposition 3.13. Let X be a Banach space, Y be a normed space and (Tn )n∈N ⊆
L(X, Y ) such that for all x ∈ X the limit T x := lim Tn x exists. Then T ∈
n∈N
L(X, Y ).

Proof. It is clear that T is well-defined and linear. By the uniform boundedness

principle, there exists an C ∈ R such that kTn k < C for all n ∈ N. Now let
x ∈ X with kxk = 1. Then kT xk = lim kTn xk ≤ supkTn k kxk ≤ C which
n→∞ n∈N
implies that T ∈ L(X, Y ).

We finish this section with a result on strong convergence of positive operators

on a space of continuous functions. An operator T on a function space is called
positivity preserving if T f ≥ 0 for every f ≥ 0 in the domain of T .

Theorem 3.14 (Korovkin). Let X = C[0, 2π] the space of the continuous
functions on [0, 2π] and let xj ∈ X with x0 (t) = 1, x1 (t) = cos(t), x2 (t) = sin(t)
for t ∈ [0, 2π]. Let (Tn )n∈N ⊆ L(X) be a sequence of positivity preserving
operators such that Tn xj → xj for n → ∞ and j = 0, 1, 2. Then (Tn )n∈N
converges strongly to id, that is, Tn x → x for all x ∈ X.

Proof. We define the auxiliary functions

t−s
yt (s) = sin2 , t, s ∈ [0, 2π].
2

Note that yt (s) = 12 (1−cos(s) cos(t)−sin(s) sin(t)), hence yt ∈ span{x0 , x1 , x2 },

in particular Tn yt → yt for n → ∞.
50 3.2. Uniform boundedness principle

Now fix x ∈ X and ε > 0. Since x is uniformly continuous there exists a δ > 0
such that for all s, t ∈ [0, 2π]
t−s
yt (s) = sin2 <δ =⇒ |x(t) − x(s)| < ε.
2
2kxk∞
Setting α = δ we obtain that
|x(t) − x(s)| ≤ ε + αyt (s), s, t ∈ [0, 2π],
because either s, t are such that yt (s) < δ, then |x(t) − x(s)| < δ by definition
of δ; or yt (s) ≥ δ, then |x(t) − x(s)| ≤ 2kxk∞ = αδ ≤ αyt (s). Hence we have
that
−ε − αyt (s) ≤ x(t) − x(s) ≤ ε + αyt (s), s, t ∈ [0, 2π]
=⇒ −εx0 − αyt ≤ x(t)x0 − x ≤ εx0 + αyt , t ∈ [0, 2π]
and since Tn is positive and yt is a positive function
−εTn x0 − αTn yt ≤ x(t)Tn x0 − Tn x ≤ εTn x0 + αTn yt , t ∈ [0, 2π].

Since Tn x0 → x0 and Tn yt → 21 (1 − cos(t)x1 − sin(t)x2 ) for n → ∞, we can find

N ∈ N large enough such that εTn x0 + αTn yt < εx0 + αyt + ε for all n ≥ N ,
hence
|x(t)Tn x0 − Tn x| ≤ εx0 + αyt + ε, t ∈ [0, 2π], n ≥ N.
Hence xTn x0 − Tn x0 converges to 0 in norm in X because by the inequality
above
|x(t)(Tn x0 )(t) − (Tn x)(t)| ≤ ε + αyt (t) + ε = 2ε, t ∈ [0, 2π], n ≥ N.
That Tn x → x follows now from
kx − xTn x0 k∞ + kxTn x0 − Tn k∞ ≤ kxk kx0 − Tn x0 k∞ + kxTn x0 − Tn k∞ .

Fourier Series
Definition 3.15. Let x : R → R a 2π-periodic integrable function. The Fourier
series of x is
∞
a0 X
S(x, t) = + (ak cos(kt) + bk sin(kt)),
2
k=1

where
1 π
Z
ak := x(s) cos(ks) ds, k ∈ N0 ,
π −π
1 π
Z
bk := x(s) sin(ks) ds, k ∈ N.
π −π
Chapter 3. Linear operators in Banach spaces 51

Note that the Fourier series is a formal series only. In the following we will
prove theorems on convergence of the Fourier series.
First we will use methods from Analysis 1 to show that for a continuously differ-
entiable periodic function its Fourier series converges uniformly to the function.
Next we will use the uniform boundedness principle to show that there exist
continuous functions whose Fourier series does not converge pointwise every-
where. Finally, the Korovkin theorem implies that the arithmetic means of the
partial sums of the Fourier series of a periodic function converges uniformly to
the function.
For a given 2π-periodic function and n ∈ N we define the nth partial sum
n
a0 X
sn (x, t) = + (ak cos(kt) + bk sin(kt)). (3.3)
2
k=1

Lemma 3.16.
( sin((n+ 1 )s)
π
s 6= 0,
Z
1 2
2 sin( 2s ) ,
sn (x, t) = x(s + t)Dn (s) ds with Dn (s) =
π −π n + 12 s = 0.
(3.4)
Dn is called Dirichlet kernel. Dn is continuous and
1 π
Z
Dn (s) ds = 1. (3.5)
π −π

Proof. Using the trigonometric identity cos(a) cos(b) + sin(a) sin(b) = cos(a − b)
and that x is 2π-periodic we obtain
n
a0 X
sn (x, t) = + (ak cos(kt) + bk sin(kt))
2
k=1
n
1 π
Z 1 X 
= x(s) + (cos(ks) cos(kt) + sin(ks) sin(kt)) ds
π −π 2
k=1
Z π n
1  1 X 
= x(s) + cos(k(s − t)) ds
π −π 2
k=1
n
1 π
Z 1 X 
= x(s + t) + cos(ks) ds.
π −π 2
k=1

Now we calculate for s 6= 0

n n n 2n
1 X 1 1 X ıs −iks 1 X ıks e−ins X ıks
+ cos(ks) = + (e + e )= e = e
2 2 2 2 2
k=1 k=1 k=−n k=0
1 1
e−ins eı2ns −1 1 eı(n+ 2 )s − e−ı(n+ 2 )s sin((n + 12 )s)
= = = = Dn (s).
2 eıs −1 2 eıs/2 − e−ıs/2 2 sin 2s
52 3.2. Uniform boundedness principle

1 1
Pn
Note that lim Dn (s) = n + 2 = 2 + k=1 cos(0). For the proof of (3.5) let
s→0
x = 1 a constant function on R. Then, by (3.3),
1 π
Z
Dn (s) ds = sn (x, t) = x(t) = 1.
π −π

Theorem 3.17. Let x : R → R be a 2π-periodic continuously differentiable

function. Then the Fourier series of x converges uniformly to x.

Proof. Let x : R → R a 2π-periodic continuously differentiable function. Let

ε > 0 and h ∈ (0, π) such that h < πkxε0 k∞ . Using (3.4) and (3.5) it follows that
1Z π 
|x(s) − sn (x, t)| =  (x(s + t) − x(t))Dn (s) ds 
 
π −π
1   −h  Z h Z π
Z 
≤ . . . ds  + | . . . | ds +  . . . ds  .
  
π

−π −h h
| {z } | {z } | {z }
=:An (t) =:Bn (t) =:Cn (t)

We have to show that An (t), Bn (t) and Cn (t) tend to 0 for n → ∞ uniformly
in t. Using the mean value theorem and that π2 σ ≤ sin(σ) for σ ∈ [0, π/2] we
obtain
Z h Z h
|x(s + t) − x(t)| 1 kx0 k |s|
Bn (t) = s | sin((n + 2 )s)| ds ≤ s ds
−h 2 sin | 2 | | {z } −h 2 sin | 2 |
≤1
π0ε
≤ 2hkx k∞ < .
2 2
Define the auxiliary function
x(s + t) − x(t)
ft (s) = , s ∈ [h, π], t ∈ [0, π].
2 sin( 2s )
2kxk∞
The functions ft are continuously differentiable and kft k∞ ≤ 2 sin(h/2) =: M1 ,
0
kx k∞
kft0 k∞
≤ =: M2 . Note that the bounds do not depend on t. Integrating
2 sin(h/2)
by parts, we find
Z π 
Cn (t) =  ft (s) sin((n + 21 )s) ds 
 
h
 cos((n + 1 )s) π Z π cos((n + 1 )s) 
2 2 0
=− f (s) + f (s) ds
  
1 t  1 t 
n+ 2 h h n+ 2
1 M
≤ 1 (2M1 + (π − h)M2 ) =: .
n+ 2 n + 21

Note that M 0 does not depend on t. When we choose N such that n+ M ε

1 < 2 we
2
obtain finally |x(s) − sn (x, t)| < ε for all t ∈ R, that is, kx − sn (x, · )k∞ < ε.
Chapter 3. Linear operators in Banach spaces 53

Theorem 3.18. There exists a 2π-periodic continuous function x whose Fourier

series does not converge everywhere pointwise to x.

Proof. We identify the 2π-periodic functions on R with


X := x ∈ C([−π, π]) : x(−π) = x(π) .
Clearly (X, k · k∞ ) is a Banach space.
Note that for fixed t ∈ [−π, π] and n ∈ N
sn ( · , t) : X → K
is linear and bounded, hence an element in X 0 .
Assume that for every x ∈ X its Fourier series converges pointwise to x. Then
for every x ∈ X and t ∈ [−π, π] the sequence (sn (x, t))n∈N is bounded (because
it converges to x(t)). By the uniform boundedness principle there exists Ct such
that ksn ( · , t)k ≤ Ct for all n ∈ N. In particular, we have
ksn ( · , 0)k ≤ C0 , n ∈ N.
It is easy to see that
Z π Z π
1   1
ksn (x, 0)k = x(s)Dn (s) ds ≤ kxk∞ |Dn (s)| ds

π π

−π −π
Rπ
hence ksn ( · , 0)k ≤ −π |Dn (s)| ds. On the other hand, the function y(s) =
sign(Dn (s)) can be approximated by continuous functions ym with kym k = 1
such that
1 π 1 π 1 π
Z Z Z
ksn (ym , 0)k = x(s)Dn (s) ds → sign(Dn (s))Dn (s) ds = |Dn (s)| ds
π −π π −π π −π
so that finally we obtain
Z π
1
ksn ( · , 0)k = |Dn (s)| ds < C0 , n ∈ N.
π −π

However ksn ( · , 0)k → ∞ for n → ∞ because

| sin((n + 12 )s)| | sin((n + 21 )s)|
Z π Z π Z π
|Dn (s)| ds = 2 s ds ≥ 2 ds
−π 0 2 sin 2 0 s
Z π(n+ 12 ) n−1
X Z (k+1)π | sin σ|
| sin σ|
= 2 dσ ≥ 2 dσ
σ σ
0 k=0 kπ
n−1 Z (k+1)π n−1
X 1 X 1
≥ 2 | sin σ| dσ = 4π .
π(k + 1) kπ π(k + 1)
k=0 k=0
n
X 1
= 4 .
k
k=1

Hence the theorem is proved.

54 3.2. Uniform boundedness principle

Finally we show that the arithmetic mean of the partial sums of the Fourier
series of a continuous function converge.

Theorem 3.19 (Fejér). As before let


X := x ∈ C([−π, π]) : x(−π) = x(π)
and let Tn ∈ L(X) defined by
n−1
1X
Tn x = sn (x, · ).
n
k=0

Then (Tn )n∈N converges strongly to id (i. e. Tn x → x for n → ∞, x ∈ X).

Proof. Note that the Tn are well-defined and that for all x ∈ X and t ∈ [−π, π]
n−1 Z n−1
1X π
Z π
1 x(s + t)) X 1
Tn x(t) = x(s + t) Dk (s) ds = s sin((k + )s) ds.
n −π nπ −π 2 sin 2 2
k=0 k=0

We simplify the sum in the integrand:

n−1 n−1  s n−1   s eins −1 
1
X X X
sin((k + 21 )s) = Im ei(k+ 2 )s = Im ei 2 eiks = Im ei 2 is
e −1
k=0 k=0 k=0
ins ins/2 ins/2
e −1 e (e − eins/2 )
= Im = Im
eis/2 − e−is/2 eis/2 − e−is/2
2i(cos(ns/2) + i sin(ns/2)) sin(ns/2) sin2 (ns/2)
= Im = .
2i sin(s/2) sin(s/2)
If we define the Fejér kernel
( 2
1 sin (ns/2)
2n sin(s/2) , s 6= 0,
Fn (s) := 1
2n s = 0,

we can write Tn x as
Z
1
Tn x(t) = −π π Fn (s)x(s + t) ds.
π
Note that all Fn are positive functions, hence the Tn are positive operators. To
show the theorem, it suffices to show that Tn xj → xj for x0 (t) = 1, x1 (t) =
cos(t), x2 (t) = sin(t) (Korovkin theorem). Using (3.3) it follows that sk (x0 , · ) =
x0 for all k ∈ N0 and that
s0 (x1 , · ) = s0 (x2 , · ) = 0,
sk (x2 , · ) = x1 , sk (x1 , · ) = x2 , k ∈ N.
n−1
Since Tn x0 = x0 , Tn xj = n xj for j = 1, 2 and n ∈ N the theorem is proved.
Chapter 3. Linear operators in Banach spaces 55

3.3 The open mapping theorem

Definition 3.20. A map f between metric spaces X and Y is called open if
the image of an open set in X is an open set in Y .

Note that an open map does not necessarily map closed sets to closed sets.
For example, the projection π : R × R → R, π((s, t)) = s, is open. The set
A := {(s, t) ∈ R × R : s ≥ 0, st ≥ 2} is closed in R × R but π(A) = (0, ∞) is
open in R.

Lemma 3.21. Let X, Y be Banach spaces and T ∈ L(X, Y ) such that

BY (0, r) ⊆ T (BX (0, 1)).

for some r > 0. Then for every ε ∈ (0, 1)

BY (0, (1 − ε)r) ⊆ T (BX (0, 1)).

Here BX (x0 , r) := {x ∈ X : kx − x0 k < r} and BY (y0 , r) := {y ∈ Y : ky − y0 k <

r} are open balls in X and Y respectively.

The lemma says that if T (BX (0, 1)) is dense in BY (0, r), then, for any 0 < ρ < r,
the ball BY (0, ρ) is contained in T (BX (0, 1)).
Proof. Note that the assertion is equivalent to

BY (0, r) ⊆ (1 − ε)−1 T (BX (0, 1)) = T (BX (0, (1 − ε)−1 )).

Fix ε > 0 and y0 ∈ BY (0, r). We have to show that there exists an x0 ∈ X with
kx0 k < (1 − ε)−1 and y0 = T (x0 ). By assumption, BY (0, r) ⊆ T (BX (0, 1)).
Hence there exists an x1 ∈ BX (0, 1) such that ky0 − T x1 k < εr. By scaling, we
know that T (BX (0, ε)) is dense in BY (0, εr). Since y0 − T x1 ∈ BY (0, εr), there
exists an x2 ∈ BX (0, ε) such that ky0 − T x1 − T x2 k < ε2 r. Since T (BX (0, ε2 ))
is dense in BY (0, ε2 r), there exists an x3 ∈ BX (0, ε2 ) such that ky0 − T x1 −
T x2 − T x3 k < ε3 r.
Continuing in this way, we obtain a sequence (xn )n∈N such that
n
X
kxn k < εn−1 , ky0 − T xk k < rεn , n ∈ N. (3.6)
k=1
P∞ P∞
It
Pfollows that x0 := k=1 xk exists and lies in B(0, (1−ε)−1 ) because k=1 kxk k <
∞
k=1 rε
k−1
= r(1 − ε)−1 . Since T is continuous, we know that
∞
X ∞

X
T (x0 ) = T xk = T xk .
k=1 k=1
Pn
By (3.6) it follows that k=1 T xk converges to y0 for n → ∞. Hence T x0 = y0
and the statement is proved.
56 3.3. The open mapping theorem

In the proof of the open mapping theorem we use the following fact.

Remark. Let T : X → Y be a linear map between normed spaces X and Y

and assume that TX (B(0, 1)) is dense in BY (y, δ) for some y ∈ Y and δ > 0.
Then TX (B(0, 1)) is dense in BY (0, δ).

Proof. Obviously it suffices to show that T (BX (0, 2)) is dense in BY (0, 2δ).
Since T is linear, it follows immediately that TX (B(0, 1)) is dense in BY (−y, δ).
Let z ∈ BY (0, 2δ) and ε > 0. Note that y − z/2 ∈ BY (y, δ) and −y − z/2 ∈
BY (−y, δ). Choose x1 , x2 ∈ BX (0, 1) such that kT x1 − (y − z/2)k < ε/2 and
kT x2 − (−y − z)k < ε/2. Since x1 + x2 ∈ BX (0, 2) and

kT (x1 + x2 ) − zk ≤ kT x1 − (y − z/2)k + kT x2 − (−y − z/2)k < ε,

it follows that z ∈ T (BX (0, 2)) because ε can be chosen arbitrarily small.

Theorem 3.22 (Open mapping theorem). Let X, Y be Banach spaces and

T ∈ L(X, Y ). Then T is open if and only if it is surjective.

Proof. If T is open, then it is obviously surjective.

Now assume that T is surjective. We use the notation of the preceding lemma.
By assumption
∞
[
Y = T (BX (0, k)).
k=1

Since Y is complete, by Baire’s category theorem there must exist an n ∈ N and

y ∈ Y and ε > 0 such BY (y, ε) ⊆ T (BX (0, n)), in other words, T (BX (0, 1)) is
dense in BY (y/n, ε/n). By the remark above T (BX (0, 1)) is dense in BY (0, ε/n),
so by Lemma 3.21 BY (0, δ) ⊆ T (BX (0, 1)) for all δ < ε/n.
Now let U ⊆ X be an open set and u ∈ U . Then there exists an open ball
BX (0, ε) such that u + BX (0, ε) ⊆ U . By what was shown above, there exists
an δ > 0 such that T u + BY (0, δ) ⊆ T u + T (BX (0, ε)) = T (u + BX (0, ε)) ⊆
T (U ).

The open mapping theorem has the following important corollaries.

Corollary 3.23 (Inverse mapping theorem). Let X, Y be Banach spaces

and T ∈ L(X, Y ) a bijection. Then T −1 exists and is continuous.

Proof. By the open mapping theorem T is open, so its inverse T −1 is continuous.

Corollary 3.24. Let X, Y be Banach spaces and T ∈ L(X, Y ) injective. Then

T −1 : rg(T ) → X is continuous if and only if rg(T ) is closed.
Chapter 3. Linear operators in Banach spaces 57

Proof. If rg(T ) is closed in Y then it is a Banach space. So by the previous

lemma, T : X → rg(T ) has a continuous inverse. On the other hand, if T −1 :
rg(T ) → X is continuous, then T is an isomorphism between X and rg(T ), so
rg(T ) is complete, hence closed in Y .

Corollary 3.25. Let X be a K-vector space and k · k1 and k · k2 norms on X

such that X is complete with respect to both norms. Assume that there exists
an α > 0 such that kxk2 ≤ αkxk1 for all x ∈ X. Then the two norms are
equivalent.

Proof. Let T : (X, k · k1 ) → (X, k · k2 ), T x = x. T is surjective and bounded

by α, so it is continuous. By the open mapping theorem, its inverse is con-
tinuous, hence bounded. The statement follows now from kxk1 = kT −1 xk1 ≤
kT −1 k kxk2 , x ∈ X.

3.4 The closed graph theorem

Let X, Y be normed spaces. Then X × Y is a normed space with either of the
norms

k · k : X × Y → R, k(x, y)k = kxk + kyk,

p
k · k : X × Y → R, k(x, y)k = kxk2 + kyk2 .

Note that the two norms defined above are equivalent.

Definition 3.26. Let X, Y be normed spaces, D a subspace of X and T : D →

Y linear. T is called closed if its graph

G(T ) := {(x, T x) : x ∈ D} ⊆ X × Y

is closed in X × Y . T is closable if G(T ) is the graph of an operator T . The

operator T is called the closure of T .

D is called the domain of T , also denoted by dom T . Sometimes the notations

T : X ⊇ D → Y or T (X → Y ) are used.

Obviously, the graph G(T ) is a subspace of X × Y .

Lemma 3.27. Let X, Y normed space and D ⊆ X a subspace. Then T : X ⊇

D → Y is closed if and only if for every sequence (xn )n∈N ⊆ D the following is
true:

(xn )n∈N and (T xn )n∈N converge

(3.7)
=⇒ x0 := lim xn ∈ D and lim T xn = T x0 .
n→∞ n→∞
58 3.4. The closed graph theorem

Proof. Assume that T is closed and let (xn )n∈N such that (xn )n∈N and (T xn )n∈N
converge. Then ((xn , T xn ))n∈N ⊆ G(T ) converges in X × Y . Since G(T ) is
closed, lim (xn , T xn ) = (x0 , y0 ) ∈ G(T ). By definition of G(T ) this implies
n→∞
lim xn = x0 ∈ D(T ) and T x0 = y0 = lim T xn .
n→∞ n→∞
Now assume that (3.7) holds and let ((xn , T xn ))n∈N ⊆ G(T ) be a sequence
that converges in X × Y . Then both (xn )n∈N and (T xn )n∈N converge, hence
x0 := lim xn ∈ D and lim T xn = T x0 which shows that lim (xn , T xn ) =
n→∞ n→∞ n→∞
(x0 , T x0 ) ∈ G(T ), hence G(T ) is closed.

Lemma 3.28. Let X, Y normed space and D ⊆ X a subspace. Then T : D → Y

is closable if and only if for every sequence (xn )n∈N ⊆ D the following is true:

lim xn = 0 and (T xn )n∈N converges =⇒ lim T xn = 0. (3.8)

n→∞ n→∞

The closure T of T is given by

D(T ) = {x ∈ X : ∃ (xn )n∈N ⊆ D with lim xn = x and (T xn )n∈N converges },

n→∞
T x = lim (T xn ) for (xn )n∈N ⊆ D with lim xn = x.
n→∞ n→∞
(3.9)

Proof. Assume that T is closable. Then G(T ) is the graph of a linear function.
Hence for a sequence (xn )n∈N ⊆ D with lim xn = 0 and lim T xn = y for
n→∞ n→∞
some y ∈ Y it follows that (0, y) ∈ G(T ) = G(T ). Hence y = T 0 = 0 because T
is linear.
Now assume that (3.8) holds and define T as in (3.9). T is well-defined because
for sequences (xn )n∈N and (x̃n )n∈N in D with lim xn = lim x̃n = x such that
n→∞ n→∞
(T xn )n∈N and (T x̃n )n∈N in D converge, it follows that (xn − x̃n )n∈N converges
to 0. Since T (xn − x̃n ) = T (xn − x̃n ) converges, it follows by assumption that
lim T xn − lim T x̃n = lim T (xn − x̃n ) = 0. Linearity of T is clear. By
n→∞ n→∞ n→∞
definition, G(T ) is the closure of G(T ), so T is the closure of T .

Remarks 3.29. Let X, Y be normed spaces.

(i) Every T ∈ L(X, Y ) is closed.

(ii) If T is closed and injective, then T −1 is closed.

Proof. Closedness of {(x, T x) : x ∈ X} ⊆ X × Y implies closeness of

{(T −1 y, y) : y ∈ rg(T )} ⊆ X × Y .

(iii) If T : D ⊇ X → Y is linear and continuous, then T is closable and

D(T ) = D(T ).
Chapter 3. Linear operators in Banach spaces 59

Examples 3.30. (i) A continuous operator that is not closed.

Let X be normed space, S ∈ L(X) and D a dense subset of X with
X \D = 6 ∅. (For example, d is dense in c0 .) Then T := S|D is continuous
because it is the restriction of a continuous function, but is not closed. To
see this, fix an x0 ∈ X \ D and choose a sequence (xn )n∈N ⊆ D which
converges to x0 . Then (T xn )n∈N converges (to Sx0 ). If T were closed,
this would imply that x0 ∈ D, contradicting the choice of x0 .
(ii) A closed operator that is not continuous.
Let X = C([−1, 1]), D = C 1 ([−1, 1]) ⊆ C([−1, 1]) and T : X ⊇ D →
X, T x = x0 . Then T is closed and not continuous.

Proof. Let (xn )n∈N ⊆ D such that (xn )n∈N and (T xn )n∈N converge. From
a well-known theorem in Analysis 1 it follows that x0 := lim xn is differ-
n→∞
entiable and T x0 = x00 = ( lim xn )0 = lim x0n = lim T xn .
n→∞ n→∞ n→∞
That T is not continuous was already shown in Example 2.7 (iv) (choose
xn (t) = n1 exp(−n(t + 1))).

(iii) Let X = L2 (−1, 1), D = C 1 ([0, 1]) ⊆ L2 ([0, 1]) and T : X ⊇ D →

X, T x = x0 . Then T is not closed.
1
Proof. Let xn : [−1, 1] → R, xn (t) = (t2 + n−2 ) 2 . Then (xn )n∈N ⊆ D
and xn → g for n → ∞ where g(t) = |t|, t ∈ [−1, 1]. The sequence of the
derivatives converges

1,
 t > 0,
t
x0n (t) = 1 → h(t) = −1, t < 0,
(t2 + n−1 ) 2 
0, t = 0.


Obviously h ∈ L2 (−1, 1). If T were closed, it would follow that g ∈

C 1 ([−1, 1]), a contradiction.

Definition 3.31. Let X, Y be Banach spaces, D ⊆ X a subspace and T : X ⊇

D → Y a linear operator. Then
k · kT : D → R, kxkT = kxk + kT xk
is called the graph norm of T .

It is easy to see that k · kT is a norm

p on D. Moreover, the norm defined above
is equivalent to the norm kxk0T = kxk2 + kT xk2 on D. Most of the time, the
graph norm defined in Definition 3.31 is easier to use in calculations. However,
the norm with the square root is sometimes more useful when operators in
Hilbert spaces are considered.

Lemma 3.32. Let X, Y be Banach spaces, D ⊆ X a subspace and T : X ⊇

D → Y a closed linear operator. Then
60 3.4. The closed graph theorem

(i) (D, k · kT ) is a Banach space.

(ii) Te : (D, k · kT ) → Y, Tex = T x, is continuous.

Proof. (i) To show completeness of (D, k · kT ) let (xn )n∈N ⊆ D be a Cauchy

sequence with respect to k · kT . Then, by definition of the graph norm, (xn )n∈N
is a Cauchy sequence in X and (T xn )n∈N is a Cauchy sequence in Y . Since
X and Y are complete, the sequences converge. Hence, by the closeness of T ,
k·kT
k · k- lim xn =: x0 ∈ D and xn −−−→ x0 .
n→∞
(ii) The statement follows from kTexkY ≤ kxkX + kT xkY = kxkT , x ∈ D.

Lemma 3.33. Let X, Y be Banach spaces, D ⊆ X a subspace and T : X ⊇

D → Y a closed surjective operator. Then T is open. If, in addition, T is
injective, then T −1 is continuous.

Proof. By Lemma 3.32 and the open mapping theorem (Theorem 3.22) the
operator iTe : (D, k · kT ) → Y, Tex = T x, is open. Let U ⊆ D open with respect
to the norm in X. Then U is also open with respect to the graph norm because
obviously i : (D, k · kT ) → (D, k · k), ix = x, is bounded, hence continuous.
Hence T (U ) = Te(U ) is open in Y .
Now assume in addition that T is injective. Then Te−1 : Y → (D, k · kT ) is
continuous by the inverse mapping theorem. Since i is continuous, also T −1 =
(Te ◦ i−1 )−1 = i ◦ Te−1 is continuous.

Lemma 3.34. Let X, Y be Banach spaces, D ⊆ X a subspace and T : X ⊇ D →

Y a closed injective linear operator such that T −1 : rg(T ) → X is continuous.
Then rg(T ) is closed.

Proof. Let (yn )n∈N be a Cauchy sequence in rg(T ) with y0 := lim yn . and
n→∞
xn := T −1 yn , n ∈ N. Then (xn )n∈N is a Cauchy sequence in D because kxn −
xm k = kT −1 yn − T −1 ym k ≤ kT −1 k kyn − ym k. Hence (xn )n∈N converges in X
and its limit x0 belongs to D and y0 = lim yn = T x0 ∈ rg(T ) because T is
n→∞
closed.

Theorem 3.35 (Closed graph theorem). Let X, Y be Banach spaces and

T : X → Y be a closed linear operator. Then T is bounded.

Proof. Note that the projections

π1 : G(T ) → X, π1 (x, T x) = x,
π2 : G(T ) → Y, π2 (x, T x) = T x
are continuous and that π1 is bijective. By assumption the graph G(T ) is closed
in X × Y , hence a Banach space, so π1 is open by the open mapping theorem
(Theorem 3.22). Hence T = π2 ◦ π1−1 is continuous.
Chapter 3. Linear operators in Banach spaces 61

Lemma 3.36. Let X, Y be Banach spaces, D ⊆ X a subspace and T : D → Y

linear. Then the following are equivalent:

(i) T is closed and D(T ) is closed.

(ii) T is closed and T is continuous.
(iii) D(T ) is closed and T is continuous.

Proof. (i) =⇒ (ii) follows from the closed graph theorem because by assumption
D is Banach space.
(ii) =⇒ (iii) and (iii) =⇒ (i) are clear.

Example 3.37. An everywhere defined linear operator that is not closed.

Let X be an infinite dimensional Banach space and (xλ )λ∈Λ an algebraic basis
of X. Without restriction we can assume kxλ k = 1, λ ∈ Λ. Choose N → Λ, n 7→
λn be an injection. Then the operator
X X
T : X → X, T (x) = n cλn xλn for x = cλn xλn ∈ X,
n∈N λ∈Λ

is well-defined. Assume that T is closed. By the closed graph theorem T must

be bounded, but kT xλn k = knxλn k = n while kxλn k = 1, n ∈ N contradicting
the boundedness of T .

3.5 Projections in Banach spaces

Definition 3.38. Let X be a vector space. P : X → X is called a projection
(on rg(P )) if P 2 = P .

Note that if P is a projection, then also id −P is a projection because (id −P )2 =

id −2P + P 2 = id −P .

Lemma 3.39. Let X be a normed space and P ∈ L(X) a projection. Then the
following holds:

(i) Either P = 0 or kP k ≥ 1.
(ii) ker(P ) and rg(P ) are closed.
(iii) X is isomorphic to ker P ⊕ rg(P ).

Proof. (i) Note that kP k = kP 2 k ≤ kP k2 , hence 0 ≤ kP k − kP k2 = kP k(1 −

kP k).
(ii) Since P is continuous, ker(P ) = P −1 ({0}) is closed. To see that rg(P ) is
closed, it suffices to show that rg(P ) = ker(id −P ). Indeed, x ∈ ker(id −P )
implies x = P x ∈ rg(P ) and y ∈ rg(P ) implies (P − id)y = P y − y = y − y = 0,
hence y ∈ ker(id −P ).
62 3.5. Projections in Banach spaces

(iii) Obviously x 7→ ((id −P )x, P x) ∈ ker(P ) ⊕ rg(P ) is well defined, linear,

bijective and continuous because id −P and P are continuous. By the inverse
mapping theorem then also the inverse operator is continuous which shows that
X and ker(P ) ⊕ rg(P ) are isomorphic.

Theorem 3.40. Let X be a normed space, U ⊆ X a finite dimensional sub-

space. Then there exists a linear continuous projection P of X to U with
kP k ≤ dim U .

Proof. From linear algebra we know that there exist bases (u1 , . . . , un ) of U and
(ϕ1 , . . . , ϕn ) of U 0 such that kuk k = kϕk k = 1 and ϕj (uk ) = δjk , j, k = 1, . . . , n.
By the Hahn-Banach theorem the ϕk can be extended to linear functionals ψk
on X with kϕk k = kψk k. We define
n
X
P : X → X, Px = ϕk (x)uk .
k=1

Obviously P is a linear bounded projection on U and

n
X n
X
kP xk ≤ kϕk k kxk kuk k = kxk = nkxk.
k=1 k=1

Theorem 3.41. Let X be Banach space, U, V ⊆ X closed subspaces such that

X and U ⊕ V are algebraically isomorphic. Then the following holds:

(i) X is isomorphic to V ⊕ U with k(u, v)k = kuk + kvk.

(ii) There exists a continuous linear projection of X on U .
(iii) V is isomorphic to X/U .

Proof. (i) Since U and V are Banach spaces, their sum U ⊕V is a Banach space.
The map U ⊕ V → X, (u, v) 7→ u + v is linear, continuous and bijective. Hence
by the inverse mapping theorem, also the inverse is continuous.
(ii) P : X → U, u + v 7→ u is the desired projection.
(iii) The map V 7→ X/V, v 7→ [v] is linear, bijective and continuous. Since U is
closed, X/U is a Banach space. By the inverse mapping theorem it follows that
V and X/U are isomorphic.

Definition 3.42. let X be a Banach space. A closed subspace U of X is called

complemented if there exists a continuous linear projection on U .

Remark 3.43. Note that not every closed subspace of a Banach space is com-
plemented in the sense of the theorem above. For example, c0 is not comple-
mented as subspace of `∞ .
Chapter 3. Linear operators in Banach spaces 63

3.6 Weak convergence

Definition 3.44. Let X be a set and U = (Uλ )λ∈Λ a family of subsets of sets in
X. The smallest topology on X such that all Uλ are open is called the topology
generated by U, denoted by τ (U).

Obviously τ (U) exists and is the intersection of topologies containing all Uλ .

Lemma 3.45. Let X be a set, U = (Uλ )λ∈Λ a family of subsets of X. Then

the topology generated by U consists of all sets of the form
n
[ \
Uγ,k , (3.10)
γ∈Γ k=1

that is, of arbitrary unions of finite intersections of sets in the family U.

Proof. Let τ (U) be the topology generated by U and σ(U) the system of sets
described in (3.10). It is not hard to see that σ(U) is a topology containing U,
hence containing τ (U). On the other hand, all sets of the form (3.10) are open
in τ (U), so σ(U) ⊆ τ (U).

Definition 3.46. Let X be a set, Λ be an index set and for every λ ∈ Λ let
(Yλ , τλ ) be a topological space. Consider a family F = (fλ : X → Yλ ) of
functions. The smallest topology on X such that all fλ are continuous, is called
the initial topology on X, denoted by σ(X, F).

Note that τ (F) = τ {fλ−1 (Uλ ) : λ ∈ Λ, Uλ ∈ τλ }.

Definition 3.47. Let X be a normed space. The topology σ(X, X 0 ) is called

the weak topology on X. The topology σ(X 0 , X) is called the weak ∗ topology on
X 0 when X is identified with a subset of X 00 by the canonical map JX .

Note that σ(X 0 , X) ⊆ σ(X 0 , X 00 ) ⊆ σk·k .

Lemma 3.48. Let X be a normed space. A sequence (xn )n∈N ⊆ X is weakly

convergent to some x0 ∈ X (in the sense of Definition 2.37) if and only if it
converges in the weak topology σ(X, X 0 ).

Proof. Assume that (xn )n∈N is weakly convergent with x0 := w- lim xn and let
n→∞
U be a σ(X, X 0 )-open set containing x0 . Then there exist ϕ1 , . . . , ϕn such that
n
\
x0 ∈ ϕ−1
j (Vj ) ⊆ U
k=1

with Vj open subsets in R containing ϕj (x0 ). Since lim ϕ(xn ) = ϕ(x0 ) for all
n→∞
64 3.6. Weak convergence

ϕ ∈ X 0 , we can choose an T N ∈ N such that ϕj (xn ) ∈ Uj for all n ≥ N and all

n
j = 1, . . . , n. Hence xn ∈ k=1 {ϕ−1j (Vj )} ⊆ U for all n ≥ N .
Now assume that (xn )n∈N ⊆ X converges to x0 in the weak topology. Since
by definition of σ(X, X 0 ) all functionals ϕ ∈ X 0 are continuous, it follows that
(ϕ(xn ))n∈N converges to ϕ(x0 ) for every ϕ ∈ X 0 .

Lemma 3.49. Let X be a normed space, (xn )n∈N ⊆ X and (ϕn )n∈N ⊆ X 0 .
(i) x0 = w- lim xn =⇒ kx0 k ≤ lim inf xn .
n→∞ n→∞

(ii) ϕ0 = w ∗ - lim ϕn =⇒ kϕ0 k ≤ lim inf ϕn .

n→∞ n→∞

Proof. (i) For x0 = 0 the assertion is clear. By the Hahn-Banach theorem there
exists an ϕ ∈ X 0 such that ϕ(x0 ) = kx0 k and kϕk = 1. Hence
kx0 k = k lim ϕ(xn )k ≤ lim inf kϕk kxn k = lim inf kxn k.
n→∞ n→∞ n→∞

(ii) Let ε > 0. Then there exists an x ∈ X with kxk = 1 such that kϕ0 k − ε <
kϕ0 (x)k. The statement follows as above:
kϕ0 k − ε < kϕ0 (x)k = lim kϕn (x)k ≤ lim inf kϕn k kxk = lim inf kϕn k.
n→∞ n→∞ n→∞

Definition 3.50. Let X be a topological space. A function f : X → R is called

upper semicontinuous if lim sup f (xn ) ≤ f (x). It is called lower semicontinuous
xn →x
if lim inf f (xn ) ≥ f (x).
xn →x

Hence the lemma above states that k · k is lower semicontinuous in the weak
topology.

Definition 3.51. For λ ∈ Λ let (Xλ , τλ ) be topological spaces. Define

Y n [ o
X := Xλ := f : Λ → Xλ : f (λ) ∈ Xλ , λ ∈ Λ .
λ∈Λ λ∈Λ

The product topology on X is the weakest topology such that for every λ ∈ Λ
the projection
πλ : X → Xλ , πj (f ) = f (j),
is continuous.

Lemma 3.52. Let X as above with the product topology. Let O ⊆ P(X) be the
family of all sets U ⊆ X such that for every u ∈ U there exist λj ∈ Λ, Uj ⊆ Xλj
open, j = 1, . . . , n, such that
n
\
u ∈ {s ∈ X : s(λj ) ∈ Uj , j = 1, . . . , n} = πλ−1
j
(Uj ) ⊆ U.
j=1 | {z }
open in O
Chapter 3. Linear operators in Banach spaces 65

Then O is the product topology on X.

Proof. This is a special case of Lemma 3.48.

Theorem 3.53 (Banach-Alaoglu). Let X be a normed space. Then the closed

unit ball K10 := {ϕ ∈ X 0 : kϕk ≤ 1} is weak ∗-compact.

Proof. For x ∈ X define the set Ax := {z ∈ K : |z| ≤ kxk} and let A :=

Q
x∈X Ax together with the product topology. By Tychonoff’s theorem A is
compact. Note that elements a ∈ A are maps X → K with |a(x)| ≤ kxk, x ∈ X.
Hence K10 ⊆ A because |ϕ(x)| ≤ kϕk kxk ≤ kxk for every ϕ ∈ K10 . The product
topology on A is the weakest topology on A such that for every x ∈ X the map
πx : A → K, a 7→ a(x) is continuous. Hence the topology on K10 induced by A
is exactly the weak ∗-topology on K10 . So it suffices to show that K10 is closed in
A with the product topology.
Let ϕ ∈ K10 and let x, y ∈ X and ε > 0. Then

U := {a ∈ A : |a(x + y) − ϕ(x + y)| < ε, |a(x) − ϕ(x)| < ε, |a(y) − ϕ(y)| < ε}

is an open neighbourhood of ϕ. Hence there exists a g ∈ U ∩ K10 . Since g is

linear, it follows that

|ϕ(x + y) − ϕ(x) − ϕ(y)| = |ϕ(x + y) − ϕ(x) − ϕ(y) − g(x + y) + g(x) + g(y)|

≤ |ϕ(x + y) − g(x + y)| + |ϕ(x) − g(x)| + |ϕ(y) − g(y)| < 3ε.

Since ε was arbitrary, this implies ϕ(x + y) = ϕ(x) + ϕ(y). Similarly it can be
shown that ϕ(λx) = λϕ(x) for λ ∈ K and x ∈ X. It follows that ϕ is linear.
Since ϕ ∈ A, it follows that kϕk ≤ 1, hence ϕ ∈ K10 .
Chapter 4. Hilbert spaces 67

Chapter 4

Hilbert spaces

4.1 Hilbert spaces

Definition 4.1. Let X be a K-vector space. A map

h· , ·i : X × X → K

is a sesquilinear form on X if for all x, y, z ∈ X, λ ∈ K

(i) hλx + y , zi = λhx , zi + hy , zi,

(ii) hx , λy + zi = λhx , yi + hx , zi.

The inner product is called

• hermitian ⇐⇒ hx , yi = hy , xi, x, z ∈ X,
• positive semidefinite ⇐⇒ hx , xi ≥ 0, x ∈ X,
• positive (definite) ⇐⇒ hx , xi > 0, x ∈ X \ {0}.

Definition 4.2. A positive definite hermitian sesquilinear form on a K-vector

X is called an inner product on X and (X, h· , ·i) is called an inner product space
(or pre-Hilbert space).

Note that hx , xi ∈ R, x ∈ X, for a hermitian sesquilinear form X because

hx , xi = hx , xi.

Lemma 4.3 (Cauchy-Schwarz inequality). Let X be a K-vector space with

inner product h· , ·i. Then for all x, y ∈ X

|hx , yi|2 ≤ |hx , xi| |hy , yi|, (4.1)

with equality if and only if x and y are linearly dependent.

68 4.1. Hilbert spaces

Proof. For x = 0 or y = 0 there is nothing to show. Now assume that y 6= 0.

For all λ ∈ K

0 ≤ hx + λy , x + λyi = hx , xi + λhy , xi + λhx , yi + |λ|2 hy , yi.

In particular, when we choose λ = − hx ,yi

hy ,yi we obtain

|hy , xi|2 |hx , yi|2 |hx , yi|2

0 ≤ hx + λy , x + λyi = hx , xi − − +
hy , yi hy , yi hy , yi
2
|hx , yi|
= hx , xi −
hy , yi

which proves (4.1). If there exist α, β ∈ K such that αx+βy = 0, then obviously
equality holds in (4.1). On the other hand, if equality holds, then hx + λy , x +
λyi = 0 with λ chosen as above, so x and y are linearly dependent.

Note that (4.1) is true also in a space X with a semidefinite hermitian sesquilin-
ear form but equality in (4.1) does not imply that x and y are linearly dependent.

Lemma 4.4. An inner product space (X, h· , ·i) becomes a normed space by
1
setting kxk := hx , xi 2 , x ∈ X.

Proof. The only property of a norm that does not follow immediately from the
definition of k · k is the triangle inequality. To prove the triangle inequality,
choose x, y ∈ X. Using the Cauchy-Schwarz inequality, we find

kx + yk2 = kxk2 + 2 Rehx , yi + kyk2 ≤ kxk2 + 2|hx , yi| + kyk2

≤ kxk2 + 2kxk kyk + kyk2 = (kxk + kyk)2 .

In the following, we will always consider inner product spaces endowed with the
topology induced by the norm.

Definition 4.5. A complete inner product space is called a Hilbert space.

Lemma 4.6. Note that the scalar product on a inner product space X is a
continuous map X × X → K when X × X is equipped with the norm k(x, y)k =
kxkX + kykX .

Proof. The statement follows from

|hx1 , x2 i − hy1 , y2 i| = |hx1 , x2 − y2 i − hy1 − x1 , y2 i|

≤ kx1 k kx2 − y2 k − ky1 − x1 k ky2 k.

The polarisation formula allows to express the inner product of two elements of
X in terms of their norms.
Chapter 4. Hilbert spaces 69

Theorem 4.7 (Polarisation formula). Let X be an inner product space over

K and x, y ∈ X. Then

1 
hx , yi = kx + yk2 − kx − yk2 , if K = R,
4
1 
hx , yi = kx + yk2 − kx − yk2 + ikx + iyk2 − ikx − iyk2 , if K = C.
4

Proof. Straightforward calculation.

A necessary and sufficient criterion for a normed space to be an inner product

space is the following.

Theorem 4.8 (Parallelogram identity). Let X be normed space. Then the

norm on X is generated by an inner product if and only if for all x, y ∈ X the
parallelogram identity is satisfied:

kx + yk2 + kx − yk2 = 2kxk2 + 2kyk2 .

In this case, the inner product is given by the polarisation formula.

Proof. Assume that the norm is generated by the inner product h· , ·i and let
1
kxk = hx , xi 2 . Then for all x, y ∈ X parallelogram identity holds:

kx + yk2 + kx − yk2 = kxk2 + kyk2 + 2 Rehx , yi + kxk2 + kyk2 − 2 Rehx , yi

= 2kxk2 + 2kyk2 .

Now assume that the norm on X is such that the parallelogram identity holds
and for x, y ∈ X define hx , yi by the polarisation formula. We prove that h· , ·i
is an inner product on X in the case K = C. The case K = R can be proved
analogously.

• Positivity.

4hx , xi = kx + xk2 − kx − xk2 + i kx + ixk2 − i kx − ixk2

= 4kxk2 + i kx + ixk2 − i kix + xk2 = 4kxk2 ≥ 0.

• Hermiticity.

4hx , yi = kx + yk2 − kx − yk2 + i kx + iyk2 − i kx − iyk2

= ky + xk2 − ky − xk2 + i k − ix + yk2 − i kix + yk2 = 4hy , xi.
70 4.1. Hilbert spaces

• Additivity.

4(hx , yi + hx , zi)
= kx + yk2 − kx − yk2 + i kx + iyk2 − i kx − iyk2
+ kx + zk2 − kx − zk2 + i kx + izk2 − i kx − izk2
y+z y − z 2
y+z y − z 2
= x + + − x − −

2 2 2 2

y+z y − z 2
y+z y − z 2
+ x + − − x − +

2 2 2 2

y+z y − z 2
y+z y − z 2
+ i x + i +i − i x − i −i

2 2 2 2

y+z y − z 2
y+z y − z 2
+ i x + i −i − i x − i +i

2 2 2 2

y + z 2 y − z 2 y + z 2 y − z 2

= 2 x + + 2 − 2 x − − 2

2 2 2 2

y − z 2
y + z 2
y + z 2
y − z 2
+ 2i x + i + 2i − 2i x − i − 2i

2 2 2 2

y + z 2 y + z 2 y + z 2 2
y + z

= 2 x + − 2 x − + 2i x + i − 2i x − i

2 2 2 2

y+z
= 2 · 4hx , i.
2
If we choose z = 0 we find hx , yi = 2hx , y2 i, hence

y + z
hx , yi + hx , zi = 2 x , = hx , y + zi.
2

• Homogeneity. From the additivity we obtain hλx , yi = λhx , yi for all λ ∈

Q. Note that hix , yi = ihx , yi, hence homogeneity is proved for λ ∈ Q+iQ.
Hence for fixed x, y ∈ C the two continuous functions C → C, λ 7→ λhx , yi
and C → C, λ 7→ hλx , yi must be equal because they are equal on the
dense subset Q + iQ of C.

Theorem 4.9. The completion of an inner product space is an inner product

space.

Proof. By continuity of the norm, the parallelogram identity holds on the com-
pletion X of an inner product space X. So X is an inner product space.

Examples 4.10. (i) Rn and Cn with the Euclidean inner product

n
X
hx , yi = xk yk , x = (xk )nk=1 , y = (yk )nk=1 ,
k=1

are inner product spaces.

Chapter 4. Hilbert spaces 71

(ii) `2 (N) with

∞
X
hx , yi = xk yk , x = (xk )k∈N , y = (yk )k∈N ,
k=1

is an inner product space.

(iii) Let R([0, 1]) be the vector space of the Riemann integrable functions on
the interval [0, 1]. Then
Z 1
hf , gi = f (t)g(t) dt, f, g ∈ R([0, 1]),
0

defines a sesquilinear form on R([0, 1]) which is not positive definite, since,
for example, χ{0} 6= 0, but hχ{0} , χ{0} i = 0.
The restriction of h· , ·i to the space of the continuous functions C([0, 1]) is
an inner product which is not complete (its closure is the space L2 ([0, 1])).

4.2 Orthogonality
Definition 4.11. Let X be an inner product space.

(i) Elements x, y ∈ X are called orthogonal, denoted by x ⊥ y, if and only if

hx , yi = 0
(ii) Subsets A, B ⊆ X are called orthogonal, denoted by A ⊥ B, if and only if
ha , bi = 0 for all a ∈ A, b ∈ B.
(iii) The orthogonal complement of a set M ⊆ X is

M ⊥ := {x ∈ X : x ⊥ m, m ∈ M }.

Remarks 4.12. (i) Pythagoras’ theorem holds: kx + yk2 = kxk2 + kyk2 if

x ⊥ y.
(ii) For every set M ⊆ X its orthogonal complement M ⊥ is a closed subspace
of X.
(iii) A ⊆ (A⊥ )⊥ for every subset A ⊆ X.
(iv) A⊥ = (span A)⊥ for every subset A ⊆ X.

Theorem 4.13 (Projection theorem). Let H be a Hilbert space, M ⊆ H a

nonempty closed and convex subset and x0 ∈ H. Then there exists exactly one
y0 ∈ M such that kx0 − y0 k = dist(x0 , M ).

Proof. Recall that dist(x0 , M ) := inf{kx0 − yk : y ∈ M }. If x0 ∈ M then the

assertion is clear (choose y0 = x0 ).
Now assume that x0 ∈
/ M . Without restriction we may assume x0 = 0.
72 4.2. Orthogonality

Existence of y0 . Let d := dist(x0 , M ) = inf{kyk : y ∈ M }. Then there exists

a sequence (yn )n∈N ⊆ M such that lim kyn k = d. We will show that (yn )n∈N
n→∞
is a Cauchy sequence. Note that k yn +y
2 k ≥ d2 because yn +y
m 2
2
m
∈ M by the
convexity of M . Hence the parallelogram identity (Theorem 4.8) yields
y − y 2 y − y 2 y + y 2
n m n m n m
≤ + − d2
2 2 2

1
= (kyn k2 + kym k2 ) − d2 −→ 0, n, m → ∞.
2
Since X is a Banach space, (yn )n∈N converges to some y0 ∈ X, and since M is
closed, y0 ∈ M .
Uniqueness of y0 . Assume that there are y0 , ỹ0 ∈ M such that ky0 k = kỹ0 k =
d = dist(x0 , M ). The parallelogram identity yields
y + ỹ 2 y + ỹ 2 y − ỹ 2 1
0 0 0 0 0 0
d2 ≤ ≤ + = (ky0 k2 + kỹ0 k2 ) = d2 .
2 2 2 2
It follows that ky0 − ỹ0 k = 0, so y0 = ỹ0 .

Lemma 4.14. Let M be a closed and convex subset of a Hilbert space H and
fix x0 ∈ H. For y0 ∈ M the following are equivalent:
(i) kx0 − y0 k = dist(x0 , M ),
(ii) Rehx0 − y0 , y − y0 i ≤ 0, y ∈ M.

Proof. (i) =⇒ (ii) For t ∈ [0, 1] and y ∈ M let yt := y0 + t(y − y0 ). Then

yt ∈ M by the convexity of M and by assumption on y0

kx0 − y0 k2 ≤ kx0 − yt k2 = kx0 − y0 − t(y − y0 )k2

= kx0 − y0 k2 − 2t Rehx0 − y0 , y − y0 i + t2 ky − y0 k2 .

So for all t ∈ (0, 1]

2 Rehx0 − y0 , y − y0 i ≤ tky − y0 k2

which implies Rehx0 − y0 , y − y0 i ≤ 0.

(ii) =⇒ (i) Let y ∈ M . By assumption

kx0 − yk2 = k(x0 − y0 ) + (y0 − y)k2

= kx0 − y0 k2 + ky0 − yk2 + 2 Rehx0 − y0 , y0 − yi ≥ kx0 − y0 k2 .

Lemma 4.15. Let U be a closed subspace of a Hilbert space H and fix x0 ∈ H.

For y0 ∈ U the following are equivalent:
(i) kx0 − y0 k = dist(x0 , U ),
(ii) x0 − y0 ⊥ U .
Chapter 4. Hilbert spaces 73

Proof. (i) =⇒ (ii) Let y ∈ U . If y = 0, then obviously hx0 − y0 , yi = 0. If

kyk = 1, let λ = kyk−1 hx0 − y0 , yi. By assumption
kx0 − y0 k2 ≤ kx0 − y0 − λyk2
= kx0 − y0 k2 − λhx0 − y0 , yi − λhy , x0 − y0 i + |λ|2 kyk2
= kx0 − y0 k2 + (1 − 2kyk−2 )|hx0 − y0 , yi|2
= kx0 − y0 k2 − |hx0 − y0 , yi|2
so hx0 − y0 , yi = 0. By linearity of U then x0 − y0 ⊥ y for all y ∈ U .
(ii) =⇒ (i) Let y ∈ U . By assumption
kx0 − yk2 = k(x0 − y0 ) + (y0 − y)k2 = kx0 − y0 k2 + ky0 − yk2 ≥ kx0 − y0 k2 .
Recall that a linear operator P : X → X on a Banach space X is called a
projection if and only if P 2 = P (see Definition 3.38).

Theorem 4.16. Let H be a Hilbert space, U ⊆ H a closed subspace with U 6=

{0}. Then there exists a projection PU ∈ L(H) on U such that kPU k = 1 and
ker(PU ) = U ⊥ . Also id −PU is continuous projection with k id −PU k = 0 if
U = H and k id −PU k = 1 if U 6= H. If U ⊕ U ⊥ is equipped with the norm
1
k(u, v)k = (kuk2 + kvk2 ) 2 , then H = U ⊕ U ⊥ .

Definition 4.17. PU as in the theorem is called the orthogonal projection on

U.

Proof of Theorem 4.16. Fix x0 ∈ H and let PU (x0 ) := y0 the unique element
y0 ∈ U such that kx0 − y0 k = dist(x0 , U ). Then rg(PU ) = U and PU2 = PU ,
hence PU is a projection on U .
By Lemma 4.15, PU (x0 ) is the unique element in U such that x0 −PU (x0 ) ∈ U ⊥ .
Rehx0 − PU (x0 ) , y − PU (x0 )i ≤ 0, y ∈ U.

We will show that PU is linear. Let x1 , x2 ∈ H and λ ∈ K. Since U ⊥ is a

subspace, we obtain
λx1 − x2 − (λPU (x1 ) − PU (x2 )) = λ(x1 − PU (x1 )) − (x2 − PU (x2 )) ∈ U ⊥ .
Hence, by definition of PU ,
PU (λx1 − x2 ) = λPU (x1 ) − PU (x2 ).

We already know that rg(PU ) = U . ker(PU ) = U ⊥ because

PU (x) = 0 ⇐⇒ x0 ∈ U ⊥ .

Therefore id −PU is a projection with rg(id −PU ) = U ⊥ and ker(id −U ) = U .

By Pythagoras’ theorem we obtain
kx0 k2 = kPU (x0 ) + (id −PU )(x0 )k2 = kPU (x0 )k2 + k(id −PU )(x0 )k2 .
74 4.2. Orthogonality

In particular, H = U ⊕ U ⊥ with norm as in the statement, and kPU k ≤ 1 and

k id −PU k ≤ 1. Lemma 3.39 implies kPU k = 1, k id −PU k = 1 if U 6= H and
k id −PU k = 0 if U = H.

Lemma 4.18. Let U be a subspace of a Hilbert space H. Then U = U ⊥⊥ .

Proof. By the projection theorem (Theorem 4.16), for every closed subspace V

PV = id −PV ⊥ = id −(id −PV ⊥⊥ ) = PV ⊥⊥ ,

hence V = V ⊥⊥ . Application to V = U shows the statement.

Definition 4.19. Let X, Y be vector spaces. A map X → Y is called antilinear

or conjugate linear if f (λx + y) = λf (x) + f (y) for all λ ∈ K and x, y ∈ X.

Theorem 4.20 (Fréchet-Riesz representation theorem). Let H be a Hilbert

space. Then the map

Φ : H → H 0, y 7→ h· , yi

is an isometric antilinear bijection.

Proof. Obviously Φ(0) = 0 ∈ H 0 . The Cauchy-Schwarz inequality yields

kΦ(y)(x)k = |hx , yi| ≤ kxk kyk, x, y ∈ H,

hence kΦ(y)k ≤ kyk If y 6= 0, then set x = kyk−1 y. Note that kxk = 1 and
kΦ(y)xk = kyk, implying that kΦ(y)k = kyk. So we have shown that Φ is
well-defined and an isometry. In particular, Φ is injective.
To show that Φ is surjective, fix an ϕ ∈ H 0 . If ϕ = 0, then ϕ = Φ(0). Otherwise
we can assume that kϕk = 1. Since ker{ϕ} is closed, there exists a decomposition
H = ker ϕ ⊕ (ker ϕ)⊥ . Note that rg(ϕ) = K, hence dim(ker ϕ)⊥ = 1. Choose
y0 ∈ (ker ϕ)⊥ with ϕ(y0 ) = 1. Then (ker ϕ)⊥ = span{y0 }. For x = u + λy0 ∈
ker ϕ ⊕ (ker ϕ)⊥ ,

hx , ky0 k−2 y0 i = λ = λϕ(y) + ϕ(u) = ϕ(x),

−1
hence
ϕ = h· , ky0 k y0 i. Since Φ is an isometry, it follows that 1 = kϕk =
ky0 k
ky0 k2 = ky10 k , so ky0 k = 1.

Corollary 4.21. (i) Every Hilbert space is reflexive.

(ii) The dual H 0 of a Hilbert space H is an inner product space by

hΦ(x) , Φ(y)iH 0 = hy , xiH

with Φ : H → H 0 as in Theorem 4.20.

Chapter 4. Hilbert spaces 75

Proof. (ii) is clear. Let Ψ : H 0 → H 00 as in Theorem 4.20. Then it is easy to

check that Ψ ◦ Φ = JH , so JH is surjective, implying that H is reflexive.

Corollary 4.22. Let H be a Hilbert space.

(i) A sequence (xn )n∈N ⊆ H converges weakly to x0 ∈ H if and only if

hxn − x0 , yi → 0, y ∈ H.

(ii) Every bounded sequence (xn )n∈N ⊆ H contains a weakly convergent sub-
sequence.

Proof. (i) follows from the Riesz-Fréchet theorem, and (ii) follows with Theo-
rem 2.40.

4.3 Orthonormal systems

Definition 4.23. Let H be a Hilbert space. A family S = (xλ )λ∈Λ of vectors
in H is called an orthonormal system if hxλ , xλ0 i = δλλ0 . A orthonormal system
S is an orthonormal basis (or a complete orthonormal system) if and only if for
every orthonormal system T

S ⊆ T =⇒ S = T.

Examples 4.24. (i) The unit vectors (en )n∈N in `2 (N) are a orthonormal
system.

(ii) Let H = L2 (−π, π). An orthonormal system in H is

n 1 o n 1 o n 1 o
S= √ ∪ √ sin(n · ) : n ∈ N ∪ √ cos(n · ) : n ∈ N .
2π π π

Lemma 4.25 (Gram-Schmidt). Let H be a Hilbert space and (xn )n∈N a

family of linearly independent vectors. Then there exists a orthonormal system
S = (sn )n∈N such that span S = span{xn : n ∈ N}.

Proof. Let s1 := kx1 k−1 x1 . Next set y2 := x2 − hx1 , s1 is1 . Note that y2 6= 0
because x2 and x1 are linearly independent. Let s2 := ky2 k−1 y2 . Then s1 ⊥ s2
and ks1 k = ks2 k = 1. Now for k ≥ 1 let
n
X
yn+1 := xn+1 − hxk , sk isk , sn+1 := kyn+1 k−1 yn+1 .
k=1

Since x1 , . . . , xn+1 are linearly independent, sn+1 is well-defined. By con-

struction, sn+1 ⊥ sj for j = 1, . . . n. Note that for every n ∈ N, sn ∈
span{x1 , . . . , xn } and xn ∈ span S, hence span S = {xn : n ∈ N}.
76 4.3. Orthonormal systems

Example. Let H = L2 ((0, 1)) and xn ∈ H defined by xn (t) = tn .qApplication

of the Gram-Schmidt orthogonalisation yields polynomials sn (t) = n + 21 Pn (t)
n
1 d
where Pn (t) = 2n n! dtn (t
2
− 1)n is the nth Legendre polynomial.

Theorem 4.26 (Bessel inequality). Let H be a Hilbert space, {sn : n ∈ N}

a orthonormal system in H. Then
∞
X
|hx , sn i|2 ≤ kxk2 , x ∈ H.
n=1

PN
Proof. For N ∈ N let xN := x− n=1 hx , sn isn . Since xN ⊥ sn for n = 1, . . . , N ,
Pythagoras’ theorem yields
XN 2 N
X N
X
kxk2 = kxN k2 + hx , sn isn = kxN k2 + |hx , sn i|2 ≥ |hx , sn i|2 .

n=1 n=1 n=1

Lemma 4.27. Let H be a Hilbert space, S = (sλ )λ∈Λ a orthonormal system in

H. Then for every x ∈ H the set

Sx := {λ ∈ Λ : hx , sλ i =
6 0}

is at most countable.

Proof. By the Bessel inequality, for every n ∈ N the set

n 1o
Sx,n := λ ∈ Λ : |hx , sλ i| ≥
n
S∞
is finite. Hence Sx = n=1 Sx,n is at most countable.
P
Definition 4.28. Let X be a normed space, (xλ )λ∈Λ ⊆ X. Then λ∈Λ xλ
P∞ to x ∈ H if and only if Λ0 := {λ ∈ Λ : xλ 6= 0} is at
converges unconditionally
most countable and n=1 xλn = x for every enumeration Λ0 = {λn : n ∈ N}.

Recall that in finite dimensional Banach spaces unconditional convergence is

equivalent to absolute convergence. In every infinite dimensional Banach space,
however, there exists a unconditionally convergent series that does not converge
absolutely (Dvoretzky-Rogers theorem).

Corollary 4.29 (Bessel inequality). Let H be a Hilbert space and S ⊆ H a

orthonormal system. Then
X
|hx , si| ≤ kxk2 , x ∈ H.
s∈S
Chapter 4. Hilbert spaces 77

Proof. For fixed x ∈ H, the set Sx = {s ∈ S : hx , si 6= 0} is at most countable

(Lemma 4.27), so the claim follows from the Bessel inequality for countable
orthonormal systems.

Theorem 4.30. Let H be a Hilbert space and S ⊆ H a orthonormal system.

Then
X
P : H → H, Px = hx , sis
s∈S

is an orthogonal projection on span S and the series is unconditionally conver-

gent.

Proof. First we proof that the series in the definition of P is unconditionally

convergent (this proves then well-definedness of P ). Fix x ∈ H. For fixed
x ∈ H, the set Sx = {s ∈ S : hx , si 6= 0} is at most countable
Pn (Lemma
4.27).
Let Sx = {sn : n ∈ N} be an enumeration of Sx . Then k=1 hx , sk isk n∈N is
a Cauchy sequence because
XM 2 M
X
hx , sk isk = |hx , sk i|2 −→ 0, M, K → ∞


k=N k=N
P∞
by Bessel’s inequality. Since H is complete, yP:= k=1 hx , sk isk exists. Let
∞
π : N → N be a permutation. Then also yπ := k=1 hx , sπ (k)isπ(k) exists. We
have to show that y = yπ . For all z ∈ H
∞
X ∞
X
hy , zi = hy , sn ihsn , zi = hy , sπ ihsπ , zi = hyπ , zi.
n=1 n=1
P∞
We have used that n=1 hy , sn ihsn , zi is absolute convergent and can therefore
be rearranged, because, by Hölder’s inequality and Bessel’s inequality
∞
X ∞
2  X ∞
 X 
|hy , sn ihsn , zi| ≤ |hy , sn i|2 |hsn , zi|2 ≤ kyk2 kzk2 < ∞.
n=1 n=1 n=1

Since y − yπ ⊥ z, z ∈ H, it follows that y = yπ . Therefore the series in the

definition of P is unconditionally convergent and P is well-defined.
It is clear that P is a linear and kP k ≤ 1 follows from Corollary 4.29. Let
⊥
x ∈ H. We have to show that x − P x ∈ span S (Theorem 4.16). This is clear
because
D X E D X E
x− hx , si , s0 = x − hx , si , s0 = 0, s0 ∈ S.
s∈S s∈Sx

Theorem 4.31. Let H be a Hilbert space and S ⊆ H a orthonormal system.

Then the following is equivalent.
78 4.3. Orthonormal systems

(i) S is a complete orthonormal system.

(ii) x ⊥ S =⇒ x = 0, x ∈ H.
(iii) H = span S.
X
(iv) x = hx , sis, x ∈ H.
s∈S
X
(v) hx , yi = hx , sihs , yi, x, y ∈ H.
s∈S
X
(vi) Parseval’s equality holds: kxk2 = |hx , si|2 , x ∈ H.
s∈S

Proof. (i) =⇒ (ii) If there exists an x ∈ H such that x ∈ S ⊥ \ {0}, then

S 0 := S ∪ {kxk−1 x} is a orthonormal system with S ( S 0 , contradicting the
maximality of S.
(ii) =⇒ (iii) follows from Lemma 4.18.
P
(iii) =⇒ (iv) By theorem 4.30, x 7→ s∈S hx , sis is the orthogonal projection
on span S = H.
(iv) =⇒ (v) straightforward.
(v) =⇒ (vi) Choose x = y.
(vi) =⇒ (i) Assume there exists an orthonormal system S 0 ) S. Then for
every s0 ∈ S 0 \ S we get the contradiction
X
1 = ks0 k2 = |hs0 , si|2 = 0.
s∈S

Now we show that the orthonormal systems in Example 4.24 are complete.

Examples 4.32. (i) The set of the unit vectors {en : n ∈ N} in `2 (N) are a
complete orthonormal system in `2 (N) because {en : n ∈ N} = `2 (N).

(ii) Let Γ be a set and define

n
`2 (Γ) := f : Γ → K : f (γ) 6= 0 for at most countably many γ ∈ Γ and
X o
|f (γ)|2 < ∞ .
γ∈Γ

P
Then hf , gi = γ∈Γ f (γ)g(γ) is a well-defined inner product (note that
only countably many terms are 6= 0 and the sum is absolutely convergent
by Hölder’s inequality). As in the case Γ = N it can be shown that `2 (Γ)
is a Hilbert space and (fλ )λ∈Γ where fλ (γ) = δλγ (Kronecker delta) is a
complete orthonormal system in `2 (Γ).
Chapter 4. Hilbert spaces 79

(iii) Let H = L2 (0, 1) and

n 1 o n 1 o n 1 o
S= √ ∪ √ sin(n · ) : n ∈ N ∪ √ cos(n · ) : n ∈ N .
2π π π

Note that span S is the set of all trigonometric polynomials. Without

restriction we can assume that K = R. By the theorem of Fejér, the
trigonometric polynomials are dense in C2π := {f ∈ C([−π, π]) : f (−π) =
f (π)} with respect to k · k∞ , hence also with respect to k · k2 . Since C2π
is k · k2 -dense in L2 ([−π, π]), S is a total subset of L2 ([−π, π]).

Lemma 4.33. Let H be an infinite dimensional Hilbert space. Then the fol-
lowing is equivalent.

(i) H is separable.
(ii) Every complete orthonormal system in H is countable.
(iii) There exists an countable complete orthonormal system in H.

Proof. (i) =⇒ (ii) Assume S ⊆ H is an uncountable complete orthonormal

1
system in H. Let ε ∈ (0, 2−p2 ) and s 6= s0 ∈ S. 0
√ Then Bε (s) ∩ Bε (s ) = ∅ because
0 2 0 2
by Pythagoras ks − s k = ksk + ks k = 2. Let A be a dense subset of H.
For every s ∈ S there exists an as ∈ A such that as ∈ Bε (s). In particular,
as 6= as0 if s 6= s0 , so A cannot be countable, thus H is not separable.
(ii) =⇒ (iii) The existence of a complete orthonormal system in H follows
from Zorn’s lemma. By assumption, it must be complete.
(iii) =⇒ (i) Let S be a countable orthonormal system in H. Then span S = H
by Theorem 4.31 and H is separable by Theorem 1.25.

Lemma 4.34. Let H be Hilbert space and S and T be complete orthonormal

system in H. Then |S| = |T |.

Proof. The statement is proved in linear algebra if |S| < ∞. Now assume that
S is not finite. For x ∈ S the set Tx := {y ∈ TS : hx , yi =
6 0} is at most countable
by Lemma 4.27. By Theorem 4.31 (ii) T ⊆ x∈S Tx , hence |T | ≤ |S||N| = |S|.
Analogously, |S| ≤ |T ||N| = |T |. By the Schröder-Bernstein theorem then
|S| = |T |.

Theorem 4.35. Let H be a Hilbert space and S an orthonormal basis of H.

Then H ∼
= `2 (S) (see Example 4.32 (ii)).

Proof. Define T : H → `2 (S) by T x(s) = hx , si, x ∈ H, s ∈ S. T is well-defined

by Bessel’s inequality. Then T : H → `2 (S) is linear and isometric byPParseval’s
equality. To show that T is surjective, let y ∈ `2 (S) and define x := s∈S y(s)s.
Then x ∈ H (Theorem 4.30) and T x = y.

Note that by construction hT x , T yi = hx , yi, x, y ∈ H.

80 4.4. Linear operators in Hilbert spaces

Corollary 4.36. If H is a separable Hilbert space, then H ∼

= `2 (N).

Corollary 4.37 (Fischer-Riesz theorem). L2 [0, 1] ∼

= `2 (N).

4.4 Linear operators in Hilbert spaces

Definition 4.38. Let H1 , H2 be Hilbert spaces and Φj : Hj → Hj0 the canon-
ical isomorphism in the Fréchet-Riesz representation theorem (Theorem 4.20).
Let T ∈ L(H1 , H2 ). Its (Hilbert space) adjoint operator is T ∗ := Φ−1 0
1 T Φ2 ∈
0
L(H2 , H1 ) where T is the Banach space adjoint of T (see Definition 2.25).

Hence T ∗ is characterised by

hT x , yi = hx , T ∗ yi, x ∈ H1 , y ∈ H2 .

Theorem 4.39. Let H1 , H2 , H3 be Hilbert spaces, S, T ∈ L(H1 , H2 ), R ∈

L(H2 , H3 ) and λ ∈ K.

(i) (λS + T )∗ = λS ∗ + T ∗ .
(ii) (RT )∗ = T ∗ R∗ .
(iii) T ∗ ∈ L(H2 , H1 ) and kT ∗ k = kT k.
(iv) T ∗∗ = T .
(v) kT T ∗ k = kT ∗ T k = kT k2 .
(vi) ker T = (rg(T ∗ ))⊥ , ker T ∗ = (rg(T ))⊥ .
(vii) If T is invertible, then (T −1 )∗ = (T ∗ )−1 .

Proof. (i)–(iv) are clear. For the proof of (v) note that for kxk = 1

kT xk2 = hT x , T xi = hx , T ∗ T xi ≤ kxk kT ∗ T xk ≤ kT ∗ T k ≤ kT ∗ k kT k = kT k2 .

Taking the supremum over all x ∈ H with kxk = 1 shows the desired equalities.

(vi) ker T = (rg T ∗ )⊥ because for x ∈ H

Tx = 0 ⇐⇒ ∀y ∈ H2 hT x , yi = 0 ⇐⇒ ∀y ∈ H2 hx , T ∗ yi = 0
∗
⇐⇒ x ⊥ rg(T ).

Then also ker T ∗ = (rg(T ∗∗ ))⊥ = (rg(T ))⊥ .

Definition 4.40. Let H1 , H2 be Hilbert spaces, T ∈ L(H1 , H2 ).

(i) T is called unitary if T is invertible and T T ∗ = idH2 and T ∗ T = idH1 .

(ii) T is called normal if H1 = H2 and T T ∗ = T ∗ T .
(iii) T is called selfadjoint if H1 = H2 and T = T ∗ .
Chapter 4. Hilbert spaces 81

Remarks. (i) T selfadjoint =⇒ T normal.

(ii) T ∈ L(H1 , H2 ) =⇒ T T ∗ and T ∗ T are selfadjoint.

Next we show that a length preserving linear map between Hilbert spaces also
preserves angles.

Lemma 4.41. Let H1 , H2 be Hilbert spaces and T ∈ L(H1 , H2 ).

(i) T is an isometry ⇐⇒ hT x , T yi = hx , yi, x, y ∈ H1 .
(ii) T is unitary ⇐⇒ T is a surjective isometry.

Proof. (i) The direction “⇐” is clear; “⇒” follows from the polarisation formula
(Theorem 4.7).
(ii) “⇒” Since T is unitary, if follows that rg(T ) ⊇ rg(T T ∗ ) = rg(idH2 ) = H2 ,
so T is surjective. T is an isometry because for all x, y ∈ H1

hT x , T yi = hT ∗ T x , yi = hx , yi,

“⇐” Assume that T as a surjective isometry. Since

hx , y − T ∗ T yi = hx , yi − hT x , T yi = 0, x, y ∈ H1 ,

it follows that T ∗ T y = y, so T ∗ T = idH1 . In particular T ∗ is surjective. Now

we will show that T ∗ is an isometry. Let ξ, η ∈ H2 . Then there exist x, y ∈ H1
such that T x = ξ and T y = η. It follows that

hT ∗ ξ , T ∗ ηi = hT ∗ T x , T ∗ T yi = hx , yi = hT x , T yi = hξ , ηi.

By the same argument as for T we conclude that idH2 = T ∗∗ T ∗ = T T ∗ .

Examples 4.42. (i) Let H1 , H2 be Hilbert spaces with dim H1 = dim H2 =

n < ∞. After choice of bases, a linear operator T : H1 → H2 has a
representation (aij )nij=1 ∈ Mn (C). The matrix corresponding to T ∗ is
then (aji )nij=1 .
(ii) Let H = L2 [0, 1]. For k ∈ L∞ ([0, 1] × [0, 1]) define
Z 1
Tk : L2 [0, 1] → L2 [0, 1], (Tk f )(t) = k(s, t)f (s) ds.
0

Then Tk ∈ L2 [0, 1] and

Z 1
Tk∗ : L2 [0, 1] → L2 [0, 1], (Tk f )(t) = k(s, t)f (s) ds,
0

that is Tk∗ = Tk .
82 4.4. Linear operators in Hilbert spaces

Theorem 4.43 (Hellinger-Toeplitz). Let H be a Hilbert space, T : H → H

a linear operator such that

hT x , yi = hx , T yi, x, y ∈ H.

Then T is bounded, hence selfadjoint.

Proof. It suffices to show that T is closed because D(T ) = H is closed. Let

(xn )n∈N ⊆ H with xn → 0 and T xn → y. Observe that

kyk2 = lim hT xn , yi = lim hxn , T yi = h lim xn , T yi = h0 , T yi = 0,

n→∞ n→∞ n→∞

so y = 0. This implies that T is closable, hence closed since D(T ) = H.

Theorem 4.44. Let H be a complex Hilbert space. For T ∈ L(H) the following
is equivalent.
(i) hT x , xi ∈ R, x ∈ H.
(ii) T is selfadjoint.

Proof. (ii) =⇒ (i) follows from

hT x , xi = hx , T xi = hT x , xi, x ∈ H.

(i) =⇒ (ii) Let x, y ∈ H and λ ∈ C.

A := hT (λx + y) , λx + yi = |λ|2 hT x , xi + hT y , yi + λhT x , yi + λhT y , xi,

B := hT (λx + y) , λx + yi = |λ|2 hT x , xi + hT y , yi + λhy , T xi + λhx , T yi.

By assumption, A = B, so in the special cases λ = 1 and λ = i we obtain

hT x , yi + hT y , xi = hy , T xi + hx , T yi,
hT x , yi − hT y , xi = −hy , T xi + hx , T yi,

so finally hT x , yi = hx , T yi.

Theorem 4.45. Let H be a Hilbert space, T ∈ L(H) selfadjoint. Then

kT k = sup |hT x , xi|.

kxk≤1

Proof. Let M := supkxk≤1 |hT x , xi|. Obviously M ≤ kT k because for kxk ≤ 1

|hT x , xi| ≤ kT kkxk2 ≤ kT k.

To show the reverse inequality fix x, y ∈ H. Observe that

hT (x + y) , x + yi − hT (x − y) , x − yi = 2hT x , yi + 2hT y , xi
= 2hT x , yi + 2hy , T xi = 4 RehT x , yi.
Chapter 4. Hilbert spaces 83

Hence, by the parallelogram identity (Theorem 4.8), for kxk ≤ 1, kyk ≤ 1,

1

RehT x , yi ≤ |hT (x + y) , x + yi| + |hT (x − y) , x − yi|
4
1
M
≤ M kx + yk2 + M kx − yk2 = (kxk2 + kyk2 ) ≤ M.
4 2
Now choose λ ∈ C, |λ| = 1 such that λhT x , yi = |hT x , yi|, so

|hT x , yi| = hT (λx) , yi = | RehT (λx) , yi| ≤ M, kxk ≤ 1, kyk ≤ 1.

In particular, kh· , T xik ≤ M , so kT xk ≤ 1 for kxk ≤ 1. This shows kT k ≤

M.

Corollary 4.46. Let H be a Hilbert space and T ∈ L(H) selfadjoint. If

hT x , xi = 0, x ∈ H, then T = 0.

Note that the condition hT x , xi = 0 automatically implies that T is selfadjoint

in the case of a complex Hilbert space. In a real Hilbert spaces H the assumption
that T is selfadjointis necessary for the statement in the corollary. For example,

0 1

let T = : R2 → R2 the rotation about 90◦ . Then T 6= 0 but
−1 0
hT x , xi = 0 for all x ∈ R2 .

Lemma 4.47. Let H be a Hilbert space, T ∈ L(H) a normal operator. Then

kT xk = kT ∗ xk, x ∈ H,

in particular, ker T = ker T ∗ .

Proof. 0 = hT ∗ T x − T T ∗ x , xi = kT xk2 − kT ∗ xk2 .

Definition 4.48. Let H be a Hilbert space. A bounded selfadjoint operator

T ∈ L(H) is called non-negative, denoted by T ≥ 0, if hT x , xi ≥ 0 for all x ∈ H.
It is called positive, denoted by T > 0, if hT x , xi > 0 for all x ∈ H \ {0}. We
write T ≤ S if and only if S − T ≥ 0. A sequence (Tn )n∈N ∈ L(H) is increasing
if and only if Tn ≤ Tn+1 , n ∈ N. A sequence (Tn )n∈N ∈ L(H) is decreasing if
and only if (−Tn )n∈N ∈ L(H) is increasing.

Theorem 4.49. Let H be a Hilbert space. Every monotonic bounded sequence

of selfadjoint linear operators on H converges strongly.

Proof. Let (Tn )n∈N be a bounded monotonic sequence of selfadjoint operators.

Without restriction we assume that it is increasing. Let

snm : H × H → K, snm (x, y) = h(Tn − Tm )x , yi

84 4.5. Projections in Hilbert spaces

is a positive semidefinite sesquilinear form on H if n ≥ m. Let M be a bound

of (Tn )n∈N . Note that then kTn − Tm k ≤ 2M . Then, using Cauchy-Schwarz
inequality, we find for n ≥ m and x ∈ H with kxk = 1

k(Tn − Tm )xk2 = h(Tn − Tm )x , (Tn − Tm )xi = snm (x, (Tn − Tm )x)

1 1
≤ snm (x, x) 2 snm ((Tn − Tm )x, (Tn − Tm )x) 2
1 1
= h(Tn − Tm )x , xi 2 h(Tn − Tm )x , (Tn − Tm )2 xi 2
1 1
≤ h(Tn − Tm )x , xi 2 kTn − Tm k 2 kTn − Tm k
3 1
≤ (2M ) 2 h(Tn − Tm )x , xi 2 .

By assumption (hTn x , xi)n∈N is a monotonically increasing bounded sequence

in R, hence convergent. It follows that (Tn x)n∈N is a Cauchy sequence, hence T
converges strongly to some T ∈ L(H) (Proposition 3.13). That T is selfadjoint
follows from

hT x , yi = lim hTn x , yi = lim hx , Tn yi = hx , T yi, x, y ∈ H.

n→∞ n→∞

4.5 Projections in Hilbert spaces

Proposition 4.50. Let H be a Hilbert space, P ∈ L(H) a projection. If P 6= 0
then the following is equivalent.
(i) P is an orthogonal projection.
(ii) kP k = 1.
(iii) P is selfadjoint.
(iv) P is normal.
(v) hP x , xi ≥ 0, x ∈ H.

Proof. (i) =⇒ (ii) follows from Theorem 4.16.

(ii) =⇒ (i) Let x ∈ ker P and y ∈ rg(P ). Then for all λ ∈ K

kλyk2 = kP (x + λy)k2 ≤ kx + λyk2 = kxk2 + |λ|2 kyk2 + 2 Re(λhx , yi).

In particular, 0 ≤ kxk2 + 2λ Rehx , yi for all λ ∈ R, and 0 ≤ kxk2 + 2iλ Imhx , yi

for all λ ∈ iR, hence Rehx , yi = Imhx , yi = 0.
(i) =⇒ (iii) Observe that hP x , yi = hx , P yi for all x, y ∈ H because

hP x , yi = hP x , y − P y + P yi = hP x , P yi,
hx , P yi = hx − P x + P x , P yi = hP x , P yi.

(iii) =⇒ (iv) is clear.

(iv) =⇒ (i) By Lemma 4.47, ker P = ker P ∗ = (rg P )⊥ .
(i) =⇒ (v) For all x ∈ H: hP x , xi = hP x , x − P x + P xi = hP x , P xi ≥ 0.
Chapter 4. Hilbert spaces 85

(v) =⇒ (i) Let x ∈ ker P , y ∈ rg P . Since for all λ ∈ R

0 ≤ hP (x + λy) , x + λyi = hλy , x + λyi = λ2 kyk2 + λhy , xi,

it follows that hx , yi = 0.

Lemma 4.51. Let H Hilbert space H. A linear operator P : H → H is an

orthogonal projection if and only if P 2 = P and hx , P yi = hy , P xi for all
x, y ∈ H.

Proof. Assume that P is an orthogonal projection. Then P 2 = P and by Propo-

sition 4.50 P is selfadjoint.
If P 2 = P and hx , P yi = hy , P xi for all x, y ∈ H, then P is a projection. By
the theorem of Hellinger-Toeplitz (Theorem 4.43) P is selfadjoint, hence P is
an orthogonal projection by Proposition 4.50.

Lemma 4.52. Let H be a Hilbert space, U1 , U2 ⊆ H closed subspaces and P1 ,

P2 the corresponding orthogonal projections. Then the following is equivalent:
(i) P1 P2 = P2 P1 = 0.
(ii) U1 ⊥ U2 .
(iii) P1 + P2 is an orthogonal projection.
If one of the equivalent conditions above hold, then rg(P1 + P2 ) = U1 ⊕ U2 .

Proof. (i) =⇒ (ii) By assumption, U2 = rg P2 ⊆ ker P1 = (rg P1 )⊥ = U1⊥ ,

hence U1 ⊥ U2 .
(ii) =⇒ (i) By assumption, rg P2 = U2 ⊆ U1⊥ = ker P1 , hence P1 P2 = 0. Since
(ii) is symmetric in U1 and U2 , it follows also that P2 P1 = 0.
(i),(ii) =⇒ (iii) Observe that P1 P2 = P2 P1 = 0, so P1 + P2 is a projection
because

(P1 + P2 )2 = P12 + P1 P2 + P2 P1 + P22 = P1 + P2 .

Since the sum of two selfadjoint operators is selfadjoint, P1 + P2 is selfadjoint,

hence, by Proposition 4.50 an orthogonal projection.
(iii) =⇒ (i) Since P1 + P2 is an orthogonal projection, it follows that

P1 P2 + P2 P1 = (P1 + P2 )2 − (P1 + P2 ) = 0.

In particular 0 = (P1 P2 +P2 P1 )P2 x = (id +P2 )P1 P2 x. Note that for y ∈ H \{0}
the vectors (id −P2 )y and P2 y are linearly independent, hence (id +P2 )y =
(id −P2 )y + 2P2 y is zero if and only if (id −P2 )y = 0 and P2 y = 0, hence y = 0.
Therefore rg P1 P2 ⊆ ker(id +P2 ) = {0}.

Lemma 4.53. Let H be a Hilbert space and P1 and P2 orthogonal projections

on subspaces U1 and U2 .
86 4.5. Projections in Hilbert spaces

(i) P1 P2 is an orthogonal projection if and only if P1 P2 = P2 P1 . In this case,

P1 P2 is an projection on U1 ∩ U2 .

(ii) P1 − P2 is an orthogonal projection if and only if P1 P2 = P2 P1 = P2 .

Proof. (i) If P1 P2 is an orthonormal projection, then, by Proposition 4.50,

P1 P2 is selfadjoint, that is P1 P2 = (P1 P2 )∗ = P2∗ P1∗ = P2 P1 . On the other
hand, if P1 and P2 commute, then it is easy to verify that (P1 P2 )2 = P1 P2
and (P1 P2 )∗ = P1 P2 , hence P1 P2 is an orthogonal projection. In this case,
rg(P1 P2 ) = rg(P2 P1 ), so rg(P1 P2 ) ⊆ U1 ∩ U2 . On the other hand, P1 P2 x = x
for every x ∈ U1 ∩ U2 , so also rg(P1 P2 ) ⊇ U1 ∩ U2 holds.
(ii) Using Lemma 4.52 we obtain

P1 − P2 orthonormal projection ⇐⇒ 1 − (P1 − P2 ) orthonormal projection

⇐⇒ (1 − P1 ) + P2 orthonormal projection
⇐⇒ P2 (1 − P1 ) = (1 − P1 )P2 = 0
⇐⇒ P2 P1 = P1 P2 = P2 .

Lemma 4.54. Let H be a Hilbert space and P1 , P2 orthogonal projections on

H0 , H1 ⊆ H. Then the following is equivalent.

(i) H0 ⊆ H1 ,
(ii) kP0 xk ≤ kP1 xk, x ∈ H.
(iii) hP0 x , xi ≤ hP1 x , xi, x ∈ H.
(iv) P0 P1 = P0 .

Proof. (ii) ⇐⇒ (iii) Let x ∈ H and P an orthogonal projection. Then hP x , xi =

hP 2 x , xi = hP x , P xi = kP xk2 .
(i) ⇐⇒ (iv)

P0 P1 = P0 ⇐⇒ P0 (id −P1 ) = 0 ⇐⇒ rg(id −P1 ) ⊆ ker P0

⊥ ⊥
⇐⇒ (rg P1 ) ⊆ (rg P0 ) ⇐⇒ H1⊥ ⊆ H0⊥
⇐⇒ H0 ⊆ H1 .

(iv) =⇒ (ii) For all x ∈ H: kP0 xk = kP0 P1 xk ≤ kP0 kkP1 xk ≤ kP1 xk.
(iii) =⇒ (i) Let x ∈ H1⊥ = ker P1 . Then 0 = hP1 x , xi ≥ hP0 x , xi = kP0 k2 ,
hence H1⊥ ⊆ ker P0 = H0⊥ .

Lemma 4.55. Let H be a Hilbert space and (Pn )n∈N a sequence of orthogonal
projections with hPm x , xi ≤ hPn x , xi for all x ∈ X and m < n. Then (Pn )n∈N
converges strongly to an orthogonal projection.
Chapter 4. Hilbert spaces 87

Proof. By Theorem 4.49 we already know that s- lim Pn =: P exists and is a

selfadjoint operator. It remains to be shown that P is a projection, that is, that
P 2 = P . For x ∈ H and n ∈ N

P 2 x = (P − Pn + Pn )(P − Pn + Pn )x = (P − Pn )P x + Pn (P − Pn )x + Pn2 x.

Note that (P − Pn )P x → 0, n → ∞, and also Pn (P − Pn )x because kPn k = 1,

n ∈ N. Since Pn2 x = Pn x → P x, it follows that P 2 = P .

4.6 The adjoint of an unbounded operator

In sections 2.4 and section 4.4 we have defined the adjoint of bounded linear
operators between Banach or Hilbert spaces. Now we define the adjoint of an un-
bounded linear operator. Recall that T (X → Y ) denotes a possibly unbounded
linear operators defined on a subspace D(T ) ⊆ X.

Definition 4.56. Let X, Y be Banach spaces and D(T ) ⊆ X a dense subspace.

For a linear map T : X ⊇ D(T ) → Y we define

D(T 0 ) := {ϕ ∈ Y 0 : x 7→ ϕ(T x) is a bounded linear functional on D(T )},

Since D(T ) is dense in X, the map D(T ) → K, x 7→ ϕ(T x) has a unique

continuous extension T 0 ϕ ∈ X 0 for ϕ ∈ D(T 0 ). Hence the Banach space adjoint
T0

T 0 : Y 0 ⊇ D(T 0 ) → X 0 , (T 0 ϕ)(x) = ϕ(T x), x ∈ D(T ), ϕ ∈ D(T 0 ).

is well-defined.

Theorem 4.57. Let X, Y be Banach spaces, D(T ) ⊆ X a dense subspace and

T : X ⊇ D(T ) → Y be a linear operator. Then T 0 is closed.

Proof. Let G(T 0 ) = {(y 0 , T 0 y 0 ) : ϕ ∈ D(T 0 )} ⊆ Y 0 × X 0 be the graph of T 0 .

Note that (y 0 , x0 ) ∈ G(T 0 ) if and only if x0 x = y 0 (T x) for all x ∈ D(T ). Now let
((yn0 , x0n ))n∈N ⊆ G(T 0 ) a convergent sequence with lim (yn0 , x0n ) = (y00 , x00 ). For
n→∞
all x ∈ D(T ) it follows that

x00 x = lim x0n x = lim yn0 (T x) = lim y00 (T x),

n→∞ n→∞ n→∞

thus (y00 , x00 ) ∈ G(T 0 ) which implies that T 0 is closed.

Definition 4.58. Let X, Y be Banach spaces. For linear operators S, T from

X to Y we write S ⊆ T if T is an extension of S, that is, if D(S) ⊆ D(T ) and
T |D(S) = S.

Theorem 4.59. Let X, Y, Z be Banach spaces.

88 4.6. The adjoint of an unbounded operator

(i) Let (S, D(S)) and (T, D(T )) be densely defined linear operators X → Y .
If S ⊆ T then T 0 ⊆ S 0 .
(ii) Assume S(X → Y ) and T (Y → Z) are densely defined such that also T S
is densely defined. Then S 0 T 0 ⊆ (T S)0 .
(iii) Assume S(X → Y ) and T (X → Y ) are densely defined such that also
T + S is densely defined. Then (S 0 + T 0 ) ⊆ (S + T )0 .

Proof. (i) is clear from the definition of the adjoint operator.

(ii) Let z 0 ∈ D(S 0 T 0 ). Then T 0 z 0 ∈ D(S 0 ) and the map
D(S) → K, x 7→ (T 0 z 0 )(Sx)
is continuous. Then also its restriction
D(T S) → K, x 7→ (T 0 z 0 )(Sx) = z 0 (T Sx)
is continuous. Note that by assumption D(T S) is dense in X, hence z 0 ∈
D((T S)0 ) and (T S)0 z 0 = S 0 T 0 z 0 .
(iii) Let y 0 ∈ D(T 0 + S 0 ) = D(T 0 ) ∩ D(S 0 ). Then the map
D(T + S) → K, x 7→ y 0 (T x) + y 0 (Sx) = y 0 ((T + S)x)
is continuous. Since by assumption D(T + S) is dense in X, y 0 ∈ D((T + S)0 )
and (T + S)0 y 0 = (T 0 + S 0 )y 0 .
If S and T are bounded, then “=” holds in (ii) and (iii) (Theorem 2.26). Note
that for unbounded linear operators T 0 +S 0 = (T +S)0 is not necessarily true. For
example, if T (X → Y ) is a densely defined unbounded linear operator such that
also T 0 is densely defined with D(T 0 ) 6= Y 0 . Then D(T 0 − T 0 ) 6= Y 0 = D(T − T )0 .

Corollary 4.60. Let X be a Banach space, T a densely defined linear operator

in X with bounded inverse T −1 ∈ L(X). Then T 0 is invertible and
(T 0 )−1 = (T −1 )0 .

Proof. By Theorem 4.59 (ii) it follows that (T −1 )0 T 0 ⊆ (T T −1 )0 = id0X = idX ,

hence (T −1 )0 T 0 = idD(T 0 ) .
Again by Theorem 4.59 (ii) we find T 0 (T −1 )0 ⊆ (T −1 T )0 = id0D(T ) = idX , so it
suffices to show D(T 0 (T −1 )0 ) = D(T 0 ). Let ϕ ∈ D(T 0 ) and η = (T −1 )0 ϕ. For
0
every x ∈ D(T ) it follows that η(T x) = ((T −1 )0 ϕ)(T x) = ϕ(T −1 T x) = ϕ(x),
which implies η ∈ D(T 0 ), hence D(T 0 (T −1 )0 ) = D(T 0 ).
More general is Theorem 4.65 due to Phillips.

Definition 4.61. Let X be a Banach space. For subspaces A ⊆ X and B ⊆ X 0

we define the annihilators
A◦ := {ϕ ∈ X 0 : ϕ(x) = 0, x ∈ A} ⊆ X 0 ,
◦
B := {x ∈ X : ϕ(x) = 0, ϕ ∈ B} ⊆ X.
Chapter 4. Hilbert spaces 89

Remark 4.62. The sets A◦ and ◦ B are closed subspaces and ◦ (A◦ ) = A. If X
is reflexive, then also (◦ B)◦ = B.

Proof. Obviously, A◦ and ◦ B are subspaces. Let (x0n )n∈N ⊆ A◦ be a convergent

sequence. Then x00 := lim x0n ∈ A◦ because x00 x = lim x0n x = 0 for all x ∈ A.
n→∞ n→∞
Let (xn )n∈N ⊆ ◦ B be a convergent sequence. Then x0 := lim xn ∈ ◦ B because
n→∞
ϕx0 = lim ϕxn = 0 for all ϕ ∈ B.
n→∞
Now we show that ◦ (A◦ ) = A. Since obviously A ⊆ ◦ (A◦ ), also A ⊆ ◦ (A◦ ).
Assume that there exists an a ∈ ◦ (A◦ ) \ A. By a corollary to the Hahn-Banach
theorem (Corollary 2.19) there exists a ϕ ∈ X 0 such that ϕ|A = 0 and ϕ(a) 6= 0.
Therefore ϕ ∈ A◦ , so by definition of ◦ (A◦ ), also ϕ(a) = 0.
(◦ B)◦ = B follows if we identify X with X 00 using the canonical map JX .

Lemma 4.63. Let X, Y be Banach space, Y 6= {0} and T (X → Y ) a densely

defined closed linear operator and y0 ∈ Y \ {0}. Then there exists a ϕ ∈ D(T 0 )
such that ϕ(y0 ) 6= 0, in particular, D(T 0 ) 6= {0}.

Proof. By assumption, the graph G(T ) of T is closed and (0, y0 ) 6= G(T ). Hence,
by a corollary to the Hahn-Banach theorem (Corollary 2.19) there exists ψ ∈
(X × Y )0 such that ψ|G(T ) = 0 and ψ((0, y0 )) 6= 0. Let ϕ : Y → K, ϕ(y) =
ψ((0, y)). Obviously ϕ ∈ Y 0 and ϕ(y0 ) 6= 0. Moreover, ϕ ∈ D(T 0 ) because for
all x ∈ D(T )

ϕ(T x) = ψ((0, T x)) = ψ((x, T x) − (x, 0)) = ψ((x, T x)) − ψ((x, 0))
= −ψ((x, 0)).

Theorem 4.64. Let X and Y be Banach spaces. For a densely defined closed
linear operator T (X → Y ) the following holds:
◦
(i) rg(T )◦ = rg(T ) = ker T 0 .
(ii) rg T = ◦ (ker T 0 ).
(iii) rg T = Y ⇐⇒ T 0 is injective.
(iv) ◦ (rg T 0 ) ∩ D(T ) = ker T.
(v) rg T 0 ⊆ (ker T )◦ .

Proof. (i) The first equality is clear. The second equality follows from

ϕ ∈ rg(T )◦ ⇐⇒ ∀ y ∈ rg(T ) ϕ(y) = 0

⇐⇒ ∀ x ∈ D(T ) ϕ(T x) = 0
⇐⇒ ϕ ∈ D(T 0 ), T 0 ϕ = 0
⇐⇒ ϕ ∈ ker(T 0 ).

(ii) rg T = ◦ ((rg T )◦ ) = ◦ (ker T 0 ) by (i) and Remark 4.62.

90 4.6. The adjoint of an unbounded operator

(iii) By (ii), rg T = Y if and only if ◦ (ker T 0 ) = Y . This is the case if and only
if ϕ(y) = 0 for all ϕ ∈ ker T 0 and y ∈ Y , that is, if and only if ker T 0 = {0}.
(iv) Let x ∈ ker(T ) and x0 ∈ rg T 0 . Choose y 0 ∈ D(T 0 ) with T 0 y 0 = x0 . Then
x0 x = (T 0 y 0 )x = y 0 (T x) = y 0 (0) = 0, hence x ∈ ◦ (rg T 0 ).
Now let x ∈ ◦ (rg T 0 ) ∩ D(T ). Then y 0 (T x) = (T 0 y 0 )x = 0 for all y 0 ∈ Y 0 . Since
T is closed, it follows by Lemma 4.63 that T x = 0, hence x ∈ ker T .
(v) Let x0 ∈ rg(T 0 ) and x ∈ ker T . Choose y 0 ∈ D(T 0 ) such that T 0 y 0 = x0 .
Then x0 x = (T 0 y 0 )x = y 0 (T x) = y 0 (0) = 0. It follows that rg(T 0 ) ⊆ (ker T )◦ ,
and since (ker T )◦ is closed, the statement is proved.

Theorem 4.65 (Phillips). Let X, Y be a Banach spaces, T (X → Y ) a densely

defined injective linear operator with rg(T ) = Y . Then

(T 0 )−1 = (T −1 )0 (4.2)

and T −1 is bounded if and only if T is closed and (T 0 )−1 is bounded on X 0 .

(T −1 denotes the inverse of T : D(T ) → rg(T ), similar for (T 0 )−1 .)

Proof.

Theorem 4.66 (Closed range theorem). Let X, Y be reflexive Banach spaces

and T : X ⊇ D(T ) → Y a closed densely defined linear operator. The following
is equivalent:

(i) rg(T ) is closed.

(ii) rg(T 0 ) is closed.
(iii) T : X ⊇ D(T ) → rg(T ) is open.
(iv) T 0 : Y 0 ⊇ D(T 0 ) → rg(T 0 ) is open.
(v) rg(T ) = ◦ (ker T 0 ).
(vi) rg(T 0 ) = (ker T )◦ .

Proof. (i) ⇐⇒ (iii) Since T is closed, (D(T ), k · kT ) is a Banach space and

Te : (D(T ), k · kT ) → rg T, Tex = T x

is continuous (Lemma 3.32). Observe that also i : (D(T ), k · kT ) → X, x 7→ x is

continuous and that T = Te ◦ i−1 : X ⊇ D(T ) → Y . Note that rg T is a Banach
space.
If rg T is closed, then Te : (D, k · kT ) → rg T is open by the open mapping
theorem (Theorem 3.22), then also T = Te ◦ i−1 : X ⊇ D(T ) → rg T is open as
composition of open maps. If T : D(T ) → rg T is open, then it is surjective,
hence rg T is closed.
Note that T 0 is closed (Theorem 4.57), hence (ii) ⇐⇒ (iv) is proved analogously.
Chapter 4. Hilbert spaces 91

(i) ⇐⇒ (v) follows from theorem 4.64 (ii).

(ii) ⇐⇒ (vi) follows from theorem 4.64 (ii)

rg(T 0 ) = ◦ (ker T 00 ) = (ker T )◦ .

(iii) ⇐⇒ (iv) Recall that T is open if and only if there exists an , r > 0 such
that the image of the open ball in X with centre 0 and radius r contains the open
unit ball in Y . That is, there exists a r > 0 such that T (BX (0, r)) ⊇ BY (0, 1).
Assume that T is open and let r as above.
To show that T 0 is open, we have to show that for every x00 ∈ rg(T 0 ) with
kx00 k < 1, there exists a y00 ∈ D(T 0 ) with T 0 y00 = x00 and ky00 k < r. Define
a linear functional ϕ on rg(T ) as follows: for y ∈ rg T with kyk < 1 choose
x ∈ D(T ) such that kxk < r and T x = y. Set ϕ(y) = x00 x and extend ϕ linearly
to rg T . Note that |ϕ(y)| = |x00 x| ≤ kx00 kkxk ≤ rkyk, ϕ is bounded, so by
the theorem of Hahn-Banach it can be extended to a functional y00 ∈ Y 0 with
ky00 k ≤ r. Note that

D(T ) → K, 7→ y00 (T x) = ϕ(T x) = x00 x

is continuous, so y00 ∈ D(T ).

(iv) ⇐⇒ (iii) Follows analogously if we note that T 00 = T by the reflexivity of
X and Y .

Definition 4.67. Let H1 , H2 be Hilbert spaces and D(T ) ⊆ H1 a dense sub-

space. For a linear map T : H1 ⊇ D(T ) → H2 its Hilbert space adjoint T ∗ is
defined by

D(T ∗ ) := {y ∈ H2 : x 7→ hT x , yi is a bounded on D(T )},

T ∗ : H2 ⊇ D(T ∗ ) → H1 , T ∗ y = y∗ ,

where y ∗ ∈ H1 such that hT x , yi = hx , y ∗ i for all x ∈ D(T ).

Note that for y ∈ D(T ∗ ) the map x 7→ hT x , yi is continuous and densely defined
and can therefore be extended uniquely to an element ϕy ∈ H10 . By the Riesz
representation theorem (Theorem 4.20) there exists exactly one y ∗ ∈ H1 as
desired.

Definition 4.68. Let H1 , H2 be Hilbert spaces and D(T ) ⊆ H1 , D(S) ⊆ H2

subspaces. The linear maps T : H1 ⊇ D(T ) → H2 and S : H2 ⊇ D(S) → H1
are called formally adjoint if

hT x , yiH2 = hx , SyiH1 , x ∈ D(T ), y ∈ D(S).

Note that the formal adjoint of a non-densely defined linear operator is not
unique; in particular, the operator trivial operator with D = {0} is formally
adjoint to every linear operator.
If T is densely defined, then its adjoint T ∗ is its maximal formally adjoint
operator.
92 4.6. The adjoint of an unbounded operator

Lemma 4.69. Let H1 and H2 be Hilbert spaces and define

U : H1 × H2 → H2 × H1 , (x, y) 7→ (y, −x).

If T (H1 → H2 ) is a densely defined linear operator, then

G(T ∗ ) = U (G(T )⊥ ) = [U (G(T ))]⊥ . (4.3)

Proof. Observe that U is unitary, hence U (G(T )⊥ ) = [U (G(T ))]⊥ . The first
equality in (4.3) follows from

(y0 , x0 ) ∈ G(T ∗ ) ⇐⇒ hT x , y0 iY = hx , x0 iX , x ∈ D(T )

⇐⇒ hT x , y0 i − hx , x0 i = 0, x ∈ D(T )
⇐⇒ h(T x, −x) , (y0 , x0 )iH2 ×H1 = 0, x ∈ D(T )
⇐⇒ hU (x, T x) , (y0 , x0 )iH2 ×H1 = 0, x ∈ D(T )
⊥
⇐⇒ (y0 , x0 ) ∈ [U (G(T ))] .

Theorem 4.70. Let H1 and H2 be Hilbert spaces. For a densely defined linear
operator T (X → Y ) the following holds:
(i) T ∗ is closed.
(ii) If T is closable, then T ∗ is densely defined and T ∗∗ = T .

Proof. (i) follows immediately from (4.3).

(ii) Let y0 ∈ D(T ∗ )⊥ . Then hy0 , yi = 0 for all y ∈ D(T ). This implies

0 = h(0, y0 ) , (−z, y)iH1 ×H2 = h(0, y0 ) , U (y, z)iH1 ×H2 , (y, z) ∈ G(T ∗ ).

Hence by Lemma 4.69,

(0, y0 ) ∈ [U −1 (G(T ∗ ))]⊥ = G(T )⊥⊥ = G(T ) = G(T ).

It follows that y0 = T 0 = 0, so D(T ∗ ) = Y . Let

V : H2 × H1 → H1 × H2 , V (y, x) = (x, −y).

Obviously V U = − idH1 ×H2 and application of Lemma 4.69 to T ∗ yields

G(T ∗∗ ) = [V (G(T ∗ ))]⊥ = [V U (G(T )⊥ )]⊥ = [−(G(T )⊥ )]⊥ = G(T )⊥⊥ = G(T )
= G(T ).

hence T ∗∗ = T .

Theorem 4.71. Let H1 , H2 , H3 be Hilbert spaces.

(i) Let T (H1 → H2 ) and S(H1 → H2 ) be densely defined linear operators. If
S ⊆ T then T ∗ ⊆ S ∗ .
Chapter 4. Hilbert spaces 93

(ii) Assume S(H1 → H2 ) and T (H2 → H3 ) are densely defined with T S = H1 .

Then S ∗ T ∗ ⊆ (T S)∗ .

(iii) Assume S(H1 → H2 ) and T (H1 → H2 ) are densely defined with T + S =

H1 . Then (T ∗ + S ∗ ) ⊆ (S + T )∗ .

If S and T are bounded, then “=” holds in (ii) and (iii).

Proof. As is in the Banach space case.

Corollary 4.72. Let H be a Hilbert space, T a densely defined linear operator

in H with bounded inverse T −1 ∈ L(H). Then T ∗ is invertible and

(T ∗ )−1 = (T −1 )∗ =: T −∗ .

Proof. By Theorem 4.71 (ii) it follows that (T −1 )∗ T ∗ ⊆ (T T −1 )∗ = idH ∗ = idH ,

hence (T −1 )∗ T ∗ = idD(T ∗ ) .
Again by Theorem 4.71 (ii) we find T ∗ (T −1 )∗ ⊆ (T −1 T )∗ = id∗D(T ) = idH , so it
suffices to show D(T ∗ (T −1 )∗ ) = D(T ∗ ). Let y ∈ D(T ∗ ) and z = (T −1 )∗ y. For
every x ∈ D(T ) it follows that hT x , zi = hT x , (T −1 )∗ yi = hT −1 T x , yi = hx , yi,
so z ∈ D(T ∗ ) which implies D(T ∗ (T −1 )∗ ) = D(T ∗ ).

Theorem 4.73. Let H1 , H2 be Hilbert spaces, T (H1 → H2 ) a densely defined

closed linear operator. Then the following holds.
⊥
(i) rg(T )⊥ = rg(T ) = ker T ∗ .

(ii) rg(T ) = (ker T ∗ )⊥ .

(iii) rg(T ∗ )⊥ = ker T .

(iv) rg(T ∗ ) = (ker T )⊥ .

Proof. (i) Note that y ∈ rg(T )⊥ if and only if hT x , yi for all x ∈ D(T ). This
is equivalent to y ∈ D(T ∗ ) and T ∗ y = 0.
⊥⊥
(ii) By (i) rg(T ) = rg(T ) = (ker T ∗ )⊥ .
(iii) By Theorem 4.70 T ∗ is closed and densely defined and T ∗∗ = T . Appli-
cation of (i) to T ∗ shows rg(T ∗ )⊥ = ker T .
(iv) Application of (ii) to T ∗ shows rg(T ∗ ) = (ker T )⊥ .

Example 4.74. Let H = L2 [0, 1]. Let

D(T1 ) := W21 (0, 1) = {x ∈ L2 [0, 1] : x absolutely continuous, x0 ∈ L2 [0, 1]},

D(T2 ) := D(T1 ) ∩ {x ∈ L2 [0, 1] : x(0) = x(1)}
D(T3 ) := D(T1 ) ∩ {x ∈ L2 [0, 1] : x(0) = x(1) = 0}.
94 4.6. The adjoint of an unbounded operator

For k = 1, 2, 3 let

Tk : H ⊇ D(Tk ) → H, Tk x = ix0 .

Obviously, the Tk are well-defined and D(Tk ) is dense in H (Theorem A.27).

We will show: T1∗ = T3 , T3∗ = T1 , T2∗ = T2 , in particular all Tk are closed.
Proof. Let x, y ∈ D(T1 ). Then, using integration by parts,
Z 1 1 Z 1
hT1 x , yi = ix0 (t)y(t) dt = ix(t)y(t) − ix(t)y 0 (t) dt

0 0 0
= ix(1)y(1) − ix(0)y(0) + hx , T1 yi.

In particular we obtain

hT x , yi = hx , T yi, x ∈ D(T1 ), y ∈ D(T3 ),

hT x , yi = hx , T yi, x, y ∈ D(T2 ).

This shows that

D(T3 ) ⊆ D(T1∗ ), D(T2 ) ⊆ D(T2∗ ) and D(T1 ) ⊆ D(T3∗ )

and T1∗ |D(T3 ) = T3 , T3∗ |D(T3 ) = T1 and T2∗ |D(T3 ) = T2 .

To prove the inclusion D(T1∗ ) ⊆ D(T3 ) let g ∈ D(T1∗ ) and ϕ = T1∗ g. Define
Rt
Φ(t) = 0 ϕ(s) ds. Then Φ is absolutely continuous and Φ0 = ϕ. For x ∈ D(T1 )
Z 1 Z 1
ix0 (t)g(t) dt = hT1 x , gi = hx , ϕi = ix(t)ϕ(t) dt
0 0
1 Z 1
= x(t)Φ(t) − ix0 (t)Φ(t) dt

0 0
Z 1
= x(1)Φ(1) − ix0 (t)Φ(t) dt.
0

Note that Φ(1) = 0 as can be seen if x is chosen to be a constant function.

Hence
Z 1
ix0 (t)(g(t)iΦ(x)) dt = 0, x ∈ D(T1 ),
0

implying that g + iΦ ∈ rg(T1 )⊥ = {0}. It follows that g is absolutely continuous

and g(0) = iϕ(0) = 0, g(1) = iϕ(1) = 0, so g ∈ D(T3 ).
Analogously, T2∗ = T2 and T3∗ = T1 can be shown.

Definition 4.75. Let H be a Hilbert spaces, D(T ) ⊆ H a dense subspace and

T : H ⊇ D(T ) → H a linear map.
(i) T is called symmetric if T ⊆ T ∗ .
Chapter 4. Hilbert spaces 95

(ii) T is called selfadjoint if T = T ∗ .

(iii) T is called essentially selfadjoint if T = T ∗ .

The operator T2 in the example above is selfadjoint, the operator T3 is symmet-

ric.
Chapter 5. Spectrum of linear operators 97

Chapter 5

Spectrum of linear
operators

If not stated explicitely otherwise, all Hilbert and Banach spaces in this chapter
are assumed to be complex vector spaces.

5.1 The spectrum of a linear operator

Definition 5.1. Let X be a Banach space and T (X → X) a densely defined
linear operator.

ρ(T ) := {λ ∈ C : λ id −T is bijective} resolvent set of T,

σ(T ) := C \ ρ(T ) spectrum of T.

The spectrum of T is further divided in point spectrum σp (T ), continuous spec-

trum σc (T ) and residual spectrum σr (T ):

σp (T ) := {λ ∈ C : λ id −T is not injective},
σc (T ) := {λ ∈ C : λ id −T is injective, rg(T − λ id) 6= X, rg(T − λ id) = X},
σr (T ) := {λ ∈ C : λ id −T is injective, rg(T − λ id) 6= X}.

It follows immediately from the definition that

σ(T ) = σp (T ) ∪˙ σc (T ) ∪˙ σr (T ).

In the following, we often write λ − T instead of λ id −T .

Definition 5.2. (i) Elements λ ∈ σp (T ) are called eigenvalues of T .

(ii) For λ ∈ σp (T ) we define the geometric eigenspace of T in λ, Nλ (T ), and

98 5.2. The resolvent

the algebraic eigenspace of T in λ, Aλ (T ), by

Nλ (T ) := ker(T − λ),
Aλ (T ) := {x ∈ X : (T − λ)n x = 0 for some n ∈ N}.

(iii) For λ ∈ ρ(T ) the resolvent of T in λ is (λ id −T )−1 := R(λ, T ). The map

ρ(T ) → L(X), λ 7→ R(λ, T )

is the resolvent map.

Remark 5.3. If T is closed, then (T − λ)−1 is closed if it exists. Therefore, by

the closed graph theorem,

ρ(T ) = {λ ∈ C : T − λ is injective and (T − λ)−1 ∈ L(X)}.

Remark 5.4. Often the resolvent set of a linear operator is defined slightly
differently: Let T (X → X) is a densely defined linear operator. Then λ ∈ ρ(T )
if and only if λ − T is bijective and (λ − T )−1 ∈ L(X). With this definition
it follows that ρ(T ) = ∅ for every non-closed T (X → X) because one of the
following cases holds:
(i) λ − T is not bijective =⇒ λ ∈ / ρ(T );
(ii) λ − T is bijective, then (λ − T )−1 is defined everywhere and not closed, so
it cannot be bounded, which implies λ ∈ / ρ(T ).

Remark 5.5. If dim X < ∞, then σc (T ) = σr (T ) = ∅ and σp (T ) is the set of

all eigenvalues of T .

Theorem 5.6 (Spectral mapping theorem for polynomials). Let X be a

Banach space, T ∈ L(X) and P ∈ C[X] a polynomial. Then

σ(P (T )) = P (σ(T )).

Proof. Let λ ∈ C. Then there exists a polynomial Q such that P (X) − P (λ) =
(X − λ)Q(X). In particular, P (T ) − P (λ) = (T − λ)Q(T ) = Q(T )(T − λ).
Hence, if λ ∈ σ(T ), then (T − λ) is not bijective, so P (T ) − P (λ) is not bijective
which implies P (σ(T )) ⊆ σ(P (T )).
Now assume µ ∈ σ(P (T )). There exist a, λ1 , . . . , λn ∈ C such that P (X) − µ =
a(X − λ1 ) · · · (X − λn ). Since P (T ) − µ is not invertible, at least one of the
terms λj − T cannot be invertible, that is at least one λj must belong to the
spectrum of T and µ = P (λj ) ∈ P (σ(T )).

5.2 The resolvent

In this section we will study the resolvent map ρ(T ) → L(X), λ 7→ R(λ, T ) =
(λ − T )−1 . We will show that its domain is open and that it is analytic.
Chapter 5. Spectrum of linear operators 99

Lemma 5.7. Let X be a Banach space and T (X → X) a closed linear operator.

1
(i) kR(λ0 , T )k ≥ for all λ0 ∈ ρ(T ).
dist(λ0 , σ(T ))
(ii) For λ0 ∈ ρ(T ) and λ ∈ C with |λ − λ0 | < kR(λ0 , T )k−1
∞
X
R(λ, T ) = (λ0 − λ)n (R(λ0 , T ))n+1 .
n=0

Note that (ii) shows that locally around a λ0 ∈ ρ(T ) the resolvent has a power
series expansion with coefficients depending only on λ0 and T .
Proof of Lemma 5.7. Recall that for a bounded linear operator S ∈ L(X) with
kSk < 1 the operator (id −S)−1 ∈ L(X) and it is given explicitly by the Neu-
mann series (Theorem 2.10)
∞
X
(id −S)−1 = Sn.
n=0

Let λ0 ∈ ρ(T ). For λ ∈ C we find

λ − T = λ0 − T − (λ0 − λ) = id −(λ0 − λ)(λ0 − T )−1 (λ0 − T ).
 

If |λ0 − λ| < k(λ0 − T )−1 k−1 , then the term in brackets is invertible, hence so
is λ − T and we obtain
−1
(λ − T )−1 = (λ0 − T )−1 id −(λ0 − λ)(λ0 − T )−1


X∞ 
−1
= (λ0 − T ) (λ0 − λ)n (λ0 − T )−n
n=0
∞
X
= (λ0 − λ)n (λ0 − T )−(n+1)
n=0

which proves (ii). If µ ∈ C with |µ| < k(T − λ0 )−1 k−1 , then λ0 + µ ∈ ρ(T ),
hence dist(λ0 , σ(T )) ≥ k(T − λ0 )−1 k−1 , so also (i) is proved.
As a corollary we obtain the following theorem.

Theorem 5.8. Let X be a Banach space and T (X → X) a closed linear oper-

ator.
(i) σ(T ) is closed.
(ii) If T ∈ L(X), then σ(T ) is compact.

Proof. (i) C \ σ(T ) = ρ(T ) is open by Lemma 5.7.

(ii) Let λ ∈ C with |λ| > kT k. Then λ − T = λ(id −λ−1 T ) is invertible since
kλ−1 T k < 1 (Neumann series, Theorem 2.10), hence λ ∈ ρ(T ) It follows that
{λ ∈ C : |λ| > kT k} ⊇ ρ(T ). Since σ is closed and bounded, it is compact.
100 5.2. The resolvent

Next we prove the so-called resolvent identities.

Theorem 5.9. Let X be a Banach space and T (X → X), S(X → X) a linear

operators with D(S) = D(T ).
(i) 1st resolvent identity:

R(λ, T ) − R(µ, T ) = (µ − λ)R(λ, T )R(µ, T ), λ, µ ∈ ρ(T ).

In particular, the resolvents commute.

(ii) 2nd resolvent identity:

R(λ, T ) − R(λ, S) = R(λ, T )(T − S)R(λ, S), λ ∈ ρ(T ) ∩ ρ(S).

Proof. (i) follows from a straightforward calculation:

R(λ, T ) − R(µ, T ) = (λ − T )−1 − (µ − T )−1

= (λ − T )−1 µ − T − (λ − T ) (µ − T )−1
 

= (µ − λ)R(λ, T )R(µ, T ).

(ii) is shown similarly:

R(λ, T ) − R(λ, S) = (λ − T )−1 − (λ − S)−1

= (λ − T )−1 λ − S − (λ − T ) (λ − S)−1
 

= R(λ, T )(T − S)R(λ, S),

Next we study properties of the resolvent map ρ(T ) → L(X), λ 7→ R(λ, T ). By

Lemma 5.7 we already now that its domain is open and that it is analytic, that
is, locally it has a power series representation.

Definition 5.10. Let Ω ∈ C be an open set, X a Banach space and f : Ω → X.

(i) f is called holomorphic in z0 ∈ Ω if and only if the limit

f (z) − f (z0 )
lim
z→z0 z − z0
exists in the norm topology. f is called holomorphic if and only if it is
holomorphic in every z0 ∈ Ω.
(ii) f is called weakly holomorphic in z0 ∈ Ω if and only if the limit

f (z) − f (z0 )
lim
z→z0 z − z0
exists in the weak topology. f is called weakly holomorphic if and only if
it is weakly holomorphic in every z0 ∈ Ω. Hence, for every ϕ ∈ X 0 the
map Ω → C, z 7→ ϕ(f (z)) is holomorphic in the usual sense.
Chapter 5. Spectrum of linear operators 101

Lemma 5.11. Let X be a Banach space. A sequence (xn )n∈N ⊆ X is a Cauchy

sequence if and only if the sequence (ϕ(xn ))n∈N ⊆ X is uniformly Cauchy for
ϕ ∈ X 0 with kϕk ≤ 1 (that is, for every ε > 0 exists a N ∈ N such that
|ϕ(xn ) − ϕ(xm )| < ε for all m, n ≥ N and all ϕ ∈ X 0 with kϕk ≤ 1).

Proof. Assume that (xn )n∈N ⊆ X is a Cauchy sequence and let ε > 0. Then
there exists a N ∈ N such that kxn − xm k < ε for m, n ≥ N . It follows that
kϕ(xn ) − ϕ(xm )k ≤ kϕkkxn − xm k < ε for all m, n ≥ N and all ϕ ∈ X 0 with
kϕk ≤ 1.
Now let ε > 0 and assume that there exists an N ∈ N such that |ϕ(xn ) −
ϕ(xm )| < ε for all m, n ≥ N and all ϕ ∈ X 0 with kϕk ≤ 1. Recall that the map
JX : X → X 00 is an isometry. It follows for m, n ≥ N

kxn − xm k = kJX xn − JX xm k = sup{|(JX xn − JX xm )ϕ| : ϕ ∈ X 0 , kϕk ≤ 1}

= sup{|ϕ(xn ) − ϕ(xm )| : ϕ ∈ X 0 , kϕk ≤ 1} < ε.

Recall the following fundamental theorem of complex analysis.

Theorem 5.12 (Cauchy’s integral formula). Let Ω ∈ C open and let f :

Ω → C holomorphic. Let z0 ∈ Ω and r > 0 such that Kr (z0 ) := {z ∈ C :
|z − z0 | ≤ r} ⊆ Ω. Then
Z
1 f (z)
f (a) = dz, a ∈ Br (z0 ) (5.1)
2πi Γr (z0 ) z−a

where Γr (z0 ) is the positively oriented boundary of Kr (z0 ). More generally, for
n ∈ N0 ,
Z
(n) n! f (z)
f (a) = dz, a ∈ Br (z0 ). (5.2)
2πi Γr (z0 ) (z − a)n+1

Theorem 5.13 (Dunford). Let X be a Banach space and let Ω ∈ C open. A

map f : Ω → X is holomorphic if and only if it is weakly holomorphic.

Proof. Clearly, holomorphy of f implies weak holomorphy. Now assume that f

is weakly holomorphic. Let z0 ∈ Ω. Choose r > 0 such that Kr (z0 ) = {z ∈ C :
|z − z0 | ≤ r} ∈ Ω. and let Γr (z0 ) be the positively oriented boundary of Kr (z0 ).
For every ϕ ∈ X 0 Cauchy’s integral formula (5.1) yields
Z
1 ϕ(f (z))
ϕ(f (a)) = dz, a ∈ Br (z0 ).
2πi Γr (z0 ) z−a

For a ∈ Br (z0 ) and 0 < |h| < r − |z0 − a| it follows that a + h ∈ Kr (z0 ), hence
102 5.2. The resolvent

with Cauchy’s integral formula we obtain

1 
ϕ(f (a + h)) − ϕ(f (a)) − (ϕ ◦ f )0 (a)
h Z
1 1h 1 1 h i
= − − ϕ(f (z)) dz
2πi Γr (z0 ) h z − a − h z − a (z − a)2
Z
1 h 1 1 i
= − ϕ(f (z)) dz
2πi Γr (z0 ) (z − a)(z − a − h) (z − a)2
Z
h ϕ(f (z))
= dz.
2πi Γr (z0 ) (z − a)2 (z − a − h)

Since z 7→ ϕ(f (z)) is holomorphic in a neighbourhood of Γr (z0 ), it is in partic-

ular continuous. Hence there exists Cϕ such that

|ϕ(f (z))| < Cϕ , z ∈ Γr (z0 ).

By a corollary to the theorem of Banach-Steinhaus (Corollary 3.8), there exists

C > 0 such that

kf (z)k < C, z ∈ Γr (z0 ).

Hence we obtain
 1  d 
ϕ(f (a + h)) − ϕ(f (a)) − (ϕ ◦ f )(a)  ≤ hkϕkC 0 .
 
h dz


This implies that

1
 1 
lim ϕ f (a + h) − f (a) = lim ϕ(f (a + h)) − ϕ(f (a)) = (ϕ ◦ f )0 (a),
h→0 h h→0 h

uniformly for ϕ ∈ X 0 , kϕk ≤ 1. Therefore, by Lemma 5.11, lim h1 f (a + h) −

h→0
f (a) exists.

Theorem 5.14 (Dunford). Let X be a Banach space, Ω ⊆ C open and T :

Ω → L(X). Then the following is equivalent:
(i) T is holomorphic in the operator norm.
(ii) T is strongly holomorphic.
(iii) T is weakly holomorphic.

Proof. (i) =⇒ (ii) follows from the definition. (ii) ⇐⇒ (iii) follows form Theo-
rem 5.13. It remains to prove (iii) =⇒ (i). As in the proof of Theorem 5.13 we
obtain for x ∈ X and ϕ ∈ X 0
Z
1  d h ϕ(T (z)x)
ϕ(T (a+h)x−T (a)x) − |z=a (ϕT (z)x) = dz.
h dz 2πi Γr (z0 ) (z − a)2 (z − a − h)
Chapter 5. Spectrum of linear operators 103

Since z 7→ ϕ(T (z)x) is holomorphic in a neighbourhood of Γr (z0 ), it is continu-

ous, so there exists Cx,ϕ such that

|ϕ(T (z)x)| < Cx,ϕ , z ∈ Γr (z0 ).

By a corollary to the theorem of Banach-Steinhaus (Corollary 3.8), there exists

Cx > 0 such that

kT (z)xk < Cx , z ∈ Γr (z0 ),

and by the theorem of Banach-Steinhaus (Theorem 3.7), there exists C > 0

such that

kT (z)k < C, z ∈ Γr (z0 ).

This implies that

1   1 
lim ϕ(T (a + h)x − T (a)x) = ϕ lim (T (a + h)x − T (a)x)
h→0 h h→0 h

exists, uniformly for ϕ ∈ X 0 , kϕk ≤ 1. Therefore, by Lemma 5.11,

1
lim (T (a + h)x − T (a)x)
h→0 h
exists and convergence is uniform for x ∈ X with kxk = 1. Analogously as in
the proof of Lemma 5.11 it follows the existence of
1
lim (T (a + h) − T (a)).
h→0 h

Theorem 5.15. Let X be a Banach space, T (X → X) a densely defined closed

linear operator. Then the resolvent map

ρ(T ) → L(X), λ 7→ R(λ, T ) = (λ − T )−1

is holomorphic.

Proof. Let λ0 ∈ ρ(T ) and λ ∈ C with |λ − λ0 | < kR(λ0 , T )k. For fixed x ∈ X
and ϕ ∈ X 0 we have by Lemma 5.7
∞
 X  
ϕ(R(λ, T )x) = ϕ (λ − λ0 )n (R(λ0 , T ))n+1 x
n=0
∞
X
(λ − λ0 )n ϕ (R(λ0 , T ))n+1 x


=
n=0

where we used that the operator series converges and ϕ is continuous. Since
the last sum is absolutely convergent, it follows that λ → ϕ(R(λ, T )x) is ana-
lytic locally at λ0 , hence holomorphic. Since weak holomorphy is equivalent to
holomorphy in the operator norm (Theorem 5.14), the theorem is proved.
104 5.2. The resolvent

The preceding theorem allows us to apply theorems of complex analysis to the

resolvent map.

Theorem 5.16. Let X be a Banach space and T ∈ L(X). Then σ(T ) 6= ∅.

Proof. Assume σ(T ) = ∅. Observe that this implies X 6= {0} and T −1 ∈ L(X).
Let λ ∈ C with |λ| > kT k. Then λ ∈ ρ(T ) and using the Neumann series
∞ ∞
X X 1
kR(λ, T )k = λn T −(n+1) ≤ |λn |kT k−(n+1) = .

n=0 n=0
kT k − |λ|

In particular, kR(λ, T )k → 0 for |λ| → ∞. Hence for every x ∈ X and ϕ ∈ X 0

the map λ → ϕ(R(λ, T )x) is holomorphic and bounded in C, so constant by
the Liouville theorem. Since ϕ(R(λ, T )x) → 0 for |λ| → ∞, it follows that
ϕ(R(λ, T )x) = 0 for all λ ∈ C, x ∈ X and ϕ ∈ X 0 . By a corollary to the
Hahn-Banach theorem (Corollary 2.16) it follows that R(λ, T )x = 0 for all
x ∈ X and λ ∈ C, hence R(λ, T ) = 0, λ ∈ C. This contradicts the fact that
1 = kT T −1 k ≤ kT kkT −1 k = 0.
The following example shows that for unbounded linear operators the cases
σ(T ) = ∅ and σ(T ) = C are possible.

Examples 5.17. (i) Let X = C([0, 1]) and

T : X ⊇ C 1 ([0, 1]) → X, T x = x0 .
Then T is unbounded and closed and σ(T ) = σp (T ) = C.
(ii) Let X = {x ∈ C([0, 1]) : x(0) = 0}, D(T ) = {x ∈ X ∩ C 1 ([0, 1]) : x0 ∈ X}
and
T : X ⊇ D(T ) → X, T x = x0 .
Then T is unbounded and closed and σ(T ) = ∅.

Proof. (i) Obviously, T is unbounded and densely defined. If (xn )n∈N ⊆ D(T )
such that xn → x and T xn → y ∈ X, then, by a theorem of Analysis 1, x is
differentiable, hence in D(T ) and T x = x0 = y which implies that T is closed.
For every λ ∈ C the differential equation x0 −λx = 0 has the solution xλ (t) = eλt .
Note that xλ ∈ D(T ) and (T − λ)xλ = 0, so λ ∈ σp (T ).
(ii) Obviously, T is unbounded and densely defined. If (xn )n∈N ⊆ D(T ) such
that xn → x and T xn → y ∈ X, then, by a theorem of Analysis 1, x is
differentiable and x0 = y. Moreover, x(0) = lim xn (0) = 0, so in D(T ) and
n→∞
T x = x0 = y which implies that T is closed.
For every λ ∈ C and every y ∈ X the initial value problem x0 − λx = y, x(0)
has exactly one solution xλ given by
Z t
xλ (t) = eλt e−λs y(s) ds.
0
Chapter 5. Spectrum of linear operators 105

Obviously xλ ∈ C 1 [0, 1], xλ (0) = 0 and x0λ (0) = λxλ (0) + y(0) = 0. Hence T − λ
is bijective, in particular λ ∈ ρ(T ).
Note that in the last example the continuity of (T − λ) can be seen immediately:
n Z t  o
−1
e−λs y(s) ds : t ∈ [0, 1]
 λt
k(T − λ) yk∞ = kxλ k∞ = sup  e

0
Z 1
≤ kyk∞ max{1, eλ } e−λs ds.
0

Definition 5.18. Let X be a Banach space The spectral radius of T ∈ L(X) is

1
r(T ) := lim sup kT n k n .

Theorem 5.19. Let X be a Banach space, T ∈ L(X) and r(T ) its spectral
radius.
(i) r(T ) ≤ kT m k1/m ≤ kT k for all m ∈ N, in particular r(T ) = lim kT m k1/m .
m→∞
(ii) σ(T ) ⊆ {λ ∈ C : |λ| ≤ r(T )}.
(iii) If X is a complex Banach space, then there exists a λ ∈ σ(T ) such that
|λ| = r(T ), in particular
r(T ) = max{|λ| : λ ∈ σ(T )}.

(iv) If X is Hilbert space and T is normal, then r(T ) = kT k.

(v) If X is a complex Hilbert space and T is normal with r(T ) = 0, then
T = 0.

Proof. (i) Let m ∈ N arbitrary. For every n ∈ N there exist pn , qn ∈ N0 with

qn < m and n = pn m + qn . Let M := max{1, kT k, . . . , kT m−1 k}. Then
kT n k = kT pn m+qn k ≤ kT pn m k kT qn k ≤ M kT m kpn .
1 1 1 qm 1
This implies r(T ) = lim sup kT n k n ≤ lim sup M 2 kT m k m − nm = kT m k m .
n→∞ n→∞
P∞
(ii) By the formula of Hadamard, the radius of convergence of n=0 z n+1 kT n k
1
is (lim sup kT n k n )−1 = r(T )−1 . Hence for all λ ∈ C, |λ| > r(T ), the series
P∞ n→∞
−(n+1) n
n=0 λ T =: A converges in norm. By Theorem 2.10 (Neumann series),
A is the inverse of λ − T . Because T is closed, it follows that {λ ∈ C : |λ| >
r(T )} ⊆ ρ(T ), or equivalently {λ ∈ C : |λ| ≤ r(T )} ⊆ σ(T ).
(iii) Let r0 := max{|λ| : λ ∈ σ(T )}. It follows from (ii) that r0 ≤ r(T ). Now
choose any µ ∈ C with |µ| > r0 . We have to show that |µ| > r(T ). Observe
that by definition of R(T ) and by the formula of Hadamard
∞
X
−1
(λ − T ) = λ−(n+1) T n , |λ| > r(T ), (5.3)
n=0
106 5.3. The spectrum of the adjoint operator

where the series on the right hand side converges in norm. In particular, for
every ϕ ∈ L(X)0
∞
X
ϕ(λ − T )−1 = λ−(n+1) ϕ(T n ), |λ| > r(T ).
n=0

Hence λ 7→ ϕ(T − λ)−1 defines an analytic function for |λ| > r(T ). It follows
from complex analysis that then the equality in (5.3) holds for all λ in the largest
open ring
P∞ where λ 7→ ϕ(λ − T ) is analytic, that is for all λ > r(T ). In partic-
−(n+1)
ular, n=0 µ ϕ(T n ) converges for every ϕ ∈ L(X)0 , hence it is weakly
convergent, and therefore (µ−(n+1) ϕ(T n ))n∈N converges to 0. It follows that
(µ−(n+1) T n )n∈N is weakly convergent to 0, hence it is bounded (Corollary 3.9).
1 1 1
Let M ∈ R such that kµ−(n+1) kT n k < M , n ∈ N. Then kkT n k n < M n µ1+ n
1
for all n ∈ N, in particular r(T ) = lim kkT n k n ≤ µ.
n→∞
(iv) Recall that kT T ∗ k = kT k2 for a normal operator T (Theorem 4.39). Hence

kT 2 k2 = kT 2 (T ∗ )2 k = k(T T ∗ )2 k = k(T T ∗ )k2 = kT k4 ,

n n
hence kT 2 k = kT k2 . By induction, it can be shown that hence kT 2 k = kT k2
for all n ∈ N, implying that
1 n 1
r(T ) = lim kT n k n = lim kT 2 k 2n = lim kT k = kT k.
n→∞ n→∞ n→∞

(v) follows directly from (iv).

Note that in general r(T ) < kT k, for example r(T ) = 0 for every nilpotent
linear operator.

5.3 The spectrum of the adjoint operator

Lemma 5.20. (i) Let X be a Banach space and T (X → X) a densely defined
closed linear operator. Then σ(T 0 ) = σ(T ) and R(λ, T )0 = R(λ, T 0 ) for
λ ∈ ρ(T ).

(ii) Let H be a Hilbert space and T (H → H) a densely defined closed linear

operator. Then σ(T ∗ ) = σ(T ) = {λ ∈ C : λ ∈ σ(T )} and R(λ, T )∗ =
R(λ∗ , T ∗ ) for λ ∈ ρ(T ).

Proof. The assertions follow from Theorem 4.65.

Lemma 5.21. Let X be a Banach space and T (X → X) densely defined and

closed.

(i) λ ∈ σp (T ) =⇒ λ ∈ σp (T 0 ) ∪ σr (T 0 ).

(ii) λ ∈ σr (T ) =⇒ λ ∈ σp (T 0 ).
Chapter 5. Spectrum of linear operators 107

Proof. (i) If λ ∈ σp (T ), then ker(λ − T ) ) {0}, rg(λ − T 0 ) ⊆ ker(T )◦ 6= X. It

follows that λ ∈ σp (T 0 ) or λ ∈ σr (T 0 ).
(i) If λ ∈ σr (T ), then rg(λ − T ) 6= X. By Theorem 4.64 rg(λ − T ) = X if and
only if (λ − T )0 = λ − T 0 is not injective, hence λ ∈ σp (T 0 ).

Theorem 5.22. Let H be a complex Hilbert space, T (H → H) a symmetric

operator and λ ∈ C \ R.

(i) k(λ − T )xk ≥ | Im(λ)| kxk for all x ∈ D(T ).

In particular T − λ : D(T ) → rg(T − λ) is invertible with continuous
inverse and the point spectrum of T is real.
(ii) If T is closed, then rg(λ − T ) is closed.

Proof. (i) For all x ∈ D(T )

   
k(λ − T )xk kxk ≥ h(λ − T )x , xi = h(Re λ − T )x , xi + ihIm λx , xi
≥ | Im λ|kxk.

In particular, λ − T is injective, which implies that λ ∈/ σp (T ).

(i) If (λ − T ) is continuous and closed, to its domain rg(λ − T ) is closed.

Theorem 5.23. Let H be a complex Hilbert space and T (H → H) a symmetric

operator. Then the following is equivalent.

(i) T is selfadjoint.
(ii) rg(λ − T ) = H for all z ∈ C \ R.
(iii) rg(±i − T ) = H.
(iv) There exist z± ∈ C with Im z+ > 0 and Im z− < 0 such that rg(z± − T ) =
H.
(v) σ(T ) ⊆ R.
(vi) T is closed and ker(±i − T ∗ ) = H.

Proof. (i) =⇒ (ii) Let λ ∈ C\R. Then rg(λ−T ) 6= H is closed by Theorem 5.22
and λ∗ ∈/ σp (T ). It follows by Theorem 4.73 that

rg(λ − T ) = rg(λ − T )⊥⊥ = ker(λ∗ − T ∗ )⊥ = ker(λ∗ − T )⊥ = {0}⊥ = H.

(ii) =⇒ (i) By assumption, T is symmetric, hence T ⊆ T ∗ , so it suffices to

show that D(T ∗ ) ⊆ D(T ). Let λ ∈ C \ R. Then λ − T and λ − T are bijective.
For x ∈ D(T ∗ ) there exists a y ∈ D(T ) such that (λ − T ∗ )x = (λ − T )y. Since
T ⊆ T ∗ , it follows that T y = T ∗ , hence x − y ∈ ker(λ − T ∗ ) = {0} which implies
x = y ∈ D(T ).
(ii) =⇒ (iii) =⇒ (iv) is obvious.
108 5.3. The spectrum of the adjoint operator

(iv) =⇒ (v) Let z± ∈ C with Im z+ > 0 and Im z− < 0 such that rg(z± − T ) =
H. By Theorem 5.22, it follows that z± −T is injective and its inverse is bounded
by |=z± |. Hence, by Lemma 5.7, every λ ∈ C with |λ − z± | < |=z± | belongs to
ρ(T ). Given any λ ∈ C \ R, repeating the argument above finitely many times
shows that λ ∈ ρ(T ).
(v) =⇒ (ii) is obvious.
(vi) =⇒ (iii) Since T is closed, the range of ±i − T is closed by Theorem 5.22.
Therefore rg(±i − T ) = rg(±i − T )⊥⊥ = ker(∓i − T ∗ )⊥ = {0}⊥ = H.
(i) =⇒ (vi) Since T = T ∗ , it is closed and C \ R ⊆ ρ(T ), in particular ker(±i −
T ) = {0}.
Analogously, we find a characterisation of essentially selfadjoint operators.

Theorem 5.24. Let H be a complex Hilbert space and T (H → H) a symmetric

operator. Then the following is equivalent.

(i) T is essentially selfadjoint.

(ii) rg(λ − T ) = H for all z ∈ C \ R.
(iii) rg(±i − T ) = H.
(iv) There exist z± ∈ C with Im z+ > 0 and Im z− < 0 such that rg(z± − T ) =
H.
(v) σ(T ) ⊆ R.
(vi) ker(±i − T ∗ ) = H.

Definition 5.25. Let X be a Banach space and T (X → X) densely defined

and closed. λ ∈ C is called approximate eigenvalue if there exists a sequence
(xn )n∈N ⊆ X such that kxn k = 1 for all n ∈ N and lim (T − λ)xn = 0. The set
n→∞
of all approximate eigenvalues is denoted by σap (T ).

Proposition 5.26. (i) Every approximate eigenvalue belongs to σ(T ).

(ii) Every boundary point of σ(T ) ⊆ C is an approximate eigenvalue of T .
(iii) If X is a Hilbert space and if T is selfadjoint, then every λ ∈ σ(T ) is an
approximate eigenvalue of T .

Proof. (i) Let λ be an approximate eigenvalue of T . Choose a sequence (xn )n∈N ⊆

D(T ) such that kxn k = 1 for all n ∈ N and (λ − T )xn → 0. Assume that
λ ∈ ρ(T ). Then R(λ, T ) = (λ − T )−1 is bounded, therefore

lim xn = lim R(λ − T )(λ − T )xn = R(λ − T ) lim (λ − T )xn = 0,

n→∞ n→∞ n→∞

in contradiction to kxn k = 1 for all n ∈ N.

(ii) Let λ be a boundary point of σ(T ). Then there exists a sequence (λn )n∈N ⊆
ρ(T ) which converges to λ. For every n ∈ N choose xn ∈ X such that
Chapter 5. Spectrum of linear operators 109

kxn k = 1 and kR(λn , T )xn k ≥ 12 kR(λn , T )k. From Lemma 5.7 we know
that kR(λn , T )k ≥ dist(λn1,σ(T )) . Set yn := kR(λn , T )k−1 R(λn , T )xn . Then
yn ∈ D(T ) and kyn k = 1 for all n ∈ N. Moreover
k(λ − T )yn k ≤ k(λ − λn )yn k + k(λn − T )yn k
= |λ − λn | + kR(λn − T )xn k−1
≤ |λ − λn | + 2kR(λn − T )k−1 −→ 0, n → ∞.
Hence λ ∈ σap (T ).
(iii) By Theorem 5.23 the spectrum of a selfadjoint operator is real, so σ(T ) =
∂σ(T ) ⊆ σap (T ) ⊆ σ(T ).

Lemma 5.27. Let H be Hilbert space and T ∈ L(H) selfadjoint. Then σ(T ) ⊆
[m, M ] where m := inf{hT x , xi : kxk = 1} and M := sup{hT x , xi : kxk = 1}.
Moreover, m, M ∈ σ(T ).

Proof. Let λ ∈ R, λ < m. Then λ − T is injective because for all x ∈ X

k(λ − T )xkkxk ≥ h(λ − T )x , xi ≥ (λ − m)kxk2 . (5.4)
In particular, rg(λ−T ) = D((λ−T )−1 ) is closed because (λ−T )−1 : rg(λ−T ) →
H is closed and continuous by (5.4). Hence rg(λ − T ) = rg(λ − T ) = ker(λ −
T )⊥ = H. It follows that (−∞, m) ∈ ρ(T ). Analogously (M, ∞) ∈ ρ(T ) is
shown.
Now we show that m ∈ σ(T ). By Proposition 5.26 it suffices to show that m ∈
σap (T ). By definition of m there exists a sequence (xn )n∈N such that kxn k = 1
for all n ∈ N and hT xn , xn i & m. Since s(x, y) := h(T − m)x , yi defines a
positive semidefinite sesquilinear form, Cauchy-Schwarz inequality implies
1 1
k(T − m)xn k2 = |s(xn , (T − m)xn )| ≤ s(xn , xn ) 2 s((T − m)xn ) 2
1 1
= h(T − m)xn , xn i 2 h(T − m)2 xn , (T − m)xn i 2 .
Since the first term in the product tends to 0 for n → ∞ and the second term
3
is bounded by (kT k − m) 2 < ∞, it follows that k(T − m)xn k tends to 0 for
n → ∞. This shows that m ∈ σap (T ). The proof of M ∈ σ(T ) is analogous.

5.4 Compact operators

Recall that a metric space M is compact if and only if every open cover of M
contains a finite cover. M is called totally bounded if and only if every for every
ε > 0 there exists a covering of M with finitely many open balls of radius ε.
M is called precompact (or precompact) if and only if M is compact. It can be
shown that a totally bounded metric M is compact if and only if M is complete.
In particular, a subset of a complete metric space is totally bounded if and only
if its closure is compact. A subset of a metric space is called relatively compact
if and only if its closure is compact.
110 5.4. Compact operators

Definition 5.28. Let X, Y be normed spaces. An operator T ∈ L(X, Y ) is

called compact if for every bounded set A ⊆ X the set T (A) is relatively com-
pact. The set of all compact operators from X to Y is denoted by K(X, Y ).

Remark 5.29. Sometimes compact operators are called completely continuous.

Remarks 5.30. (i) Every compact linear operator is bounded.

(ii) T ∈ L(X, Y ) is compact if and only if for every bounded sequence (xn )n∈N
the sequence (T xn )n∈N contains a convergent subsequence.
(iii) T ∈ L(X, Y ) is compact if and only if T (BX (0, 1)) is relatively compact.
(iv) Let T ∈ L(X, Y ) with finite dimensional rg(T ). The T is compact.
(v) The identity map id ∈ L(X) is compact if and only if X is finite-dimensional.

Theorem 5.31. Let X, Y be Banach spaces. Then K(X, Y ) is a closed subspace

of L(X, Y ).

Proof. Obviously, 0 ∈ K(X, Y ) and Remark 5.30 (ii) implies that the linear
combination of compact operators is compact. Now let (Tn )n∈N ⊆ K(X, Y ) a
Cauchy sequence. Since L(X, Y ) is complete, there exists a T ∈ L(X, Y ) such
that Tn → T . We have to show T ∈ K(X, Y ). Take an arbitrary bounded
sequence (xn )n∈N ⊆ X and choose M ∈ R such that kxn k ≤ M , n ∈ N.
(1) (1)
Since T1 is compact, there exists a subsequence (xn ) such that (T1 xn )n∈N
(k)
converges. Continuing like this, for every k ≥ 2 we find a subsequence (xn ) of
(k−1) (k) (n)
(xn ) such that (Tk xn )n∈N converges. Let (yn )n∈N = (xn )n∈N the diagonal
sequence. Then, for every k ∈ N, the sequence (Tk yn )n∈N converges. Let ε > 0.
ε
Choose k ∈ N such that kT −Tk k < 3M and N ∈ N such that kTk xn −Tk xm k ≤ 3ε
for m, n ≥ N . Then, for all m, n ≥ N ,

kT yn − T ym k ≤ kT yn − Tk yn k + kTk yn − Tk ym k + kTk ym − T ym k
Mε ε Mε
≤ + + = ε.
3M 3 3M
Hence (T yn )n∈N is Cauchy sequence in the Banach space Y , hence convergent.

Lemma 5.32. Let X, Y, Z be Banach spaces, S ∈ L(X, Y ) and T ∈ L(Y, Z).

Then T S is compact if at least one of the operators S or T is compact.

Proof. Let (xn )n∈N be a bounded sequence in X. If S is compact, then there

exists a subsequence (xnk )k∈N such that (Sxnk )k∈N converges. By continuity of
T , also (T Sxnk )k∈N converges.
Now assume that T is compact. Since S is bounded, (Sxn )k∈N is bounded,
hence there exists a subsequence (xnk )k∈N such that (T Sxnk )k∈N converges.
Chapter 5. Spectrum of linear operators 111

Theorem 5.33 (Schauder). Let X, Y be Banach space and T ∈ L(X, Y ).

Then T is compact if and only if T 0 is compact.

For the proof we use the Ascoli-Arzelá theorem.

Theorem 5.34 (Arzelá-Ascoli). Let (M, d) be a compact metric space and

A ⊆ C(M ) a family of real or complex valued continuous functions on M such
that

(i) A is bounded,
(ii) A is closed,
(iii) A is equicontinuous, that is,

∀ε > 0 ∃δ > 0 ∀f ∈ A d(x, y) < δ =⇒ |f (x) − f (y)| < ε.

Then A is compact.

Proof. See, e. g., [Rud91] or [Yos95].

Proof of Theorem 5.33. First assume that T is compact. Let KX (0, 1) := {x ∈

X : kxk ≤ 1} be the closed unit ball in X. By assumption K := T (KX (0, 1))
is compact in Y and bounded by kT k. Now let (ϕn )n∈N ⊆ Y 0 be a bounded
sequence and C ∈ R such that kϕn k ≤ C, n ∈ N. We define the functions

fn : K → K, fn (y) := ϕn (y).

Then (fn )n∈N is bounded by C and equicontinuous because |f (y1 ) − f (y2 )| ≤

Cky1 − y2 k for all y1 , y2 ∈ K. By the Ascoli-Arzelá, (fn )n∈N is compact, so
there exists a convergent subsequence (fnk )k∈N . Then also (T 0 ϕnk )k∈N converges
because

kT 0 ϕnk − T 0 ϕnm k = sup{kϕnk (T x) − ϕnm (T x)k : x ∈ KX (0, 1)}

= sup{kϕnk (y) − ϕnm (y)k : y ∈ K} = kfnk − fnm k.

Now assume that T 0 is compact. Then T 00 ∈ L(X 00 , Y 00 ) is compact. By

Lemma 5.32 T 00 ◦ JX is compact. Recall that JY ◦ T = T ◦ JX (Lemma 2.33),
so JY ◦ T : X → Y 00 is compact. Since Y is closed in Y 00 , T : X → Y is
compact.

Example 5.35. Let k ∈ C([0, 1]2 ) and

Z 1
Tk : C([0, 1]) → C([0, 1]), (Tk x)(t) = k(s, t)x(s) ds.
0

Then Tk is compact.
112 5.4. Compact operators

Proof. Obviously Tk is well-defined and bounded. Let (xn )n∈N ⊆ C([0, 1]) a
bounded sequence with bound C. Hence (Tk xn )n∈N is bounded. To show that
it is equicontinuous fix ε > 0. Since k is uniformly continuous, there exists
a δ > 0 such that |k(s, t) − k(s0 , t0 )| < ε if k(s, t) − (s0 , t0 )k < δ. Now for
t1 , t2 ∈ [0, 1] with |t1 − t2 | < δ and n ∈ N we obtain
Z 1
|Tk xn (t1 ) − Tk xn (t2 )| ≤ |k(s, t1 ) − k(s, t2 )||xn (s)| ds < εkxn k∞ ≤ Cε.
0

By the Ascoli-Arzelá theorem it follows that (Tk xn )n∈N is relatively compact,

hence it contains a convergent subsequence.

Let X be vector space and T : X → X a linear operator. Then obviously

{0} ⊆ ker T ⊆ ker T 2 ⊆ ker T 3 ⊆ . . . ,
X ⊇ rg T ⊇ rg T 2 ⊇ rg T 3 ⊇ . . . .

Lemma 5.36. Let X a vector space and T : X → X a linear operator.

(i) Assume that ker T k+1 = ker T k for some k ∈ N0 . Then ker T n = ker T k
for all integer n ≥ k.
(ii) Assume that rg T k+1 = rg T k for some k ∈ N0 . Then rg T n = rg T k for
all integer n ≥ k.

Proof. We prove the lemma by induction. The case when n = k is clear by

assumption.
(i) Assume that n > k and ker T n = ker T k . Then
ker T n+1 = {x ∈ X : T n+1 x = 0} = {x ∈ X : T x ∈ ker T k } = ker T k+1 = ker T k .

(i) Assume that n > k and rg T n = rg T k . Then

rg T n+1 = T (rg T n ) = T (rg T k ) = rg T k+1 = rg T k .

Definition 5.37. Let X be a vector space and T : X → X a linear operator.

We define
(
min{k ∈ N0 : ker T k = ker T k+1 }, if the minimum exists,
ascent of T := α(T ) :=
∞ else
(
min{k ∈ N0 : rg T k = rg T k+1 }, if the minimum exists,
descent of T := δ(T ) :=
∞ else.

Lemma 5.38. Let X be a vector space and T : X → X a linear operator. If

both the ascent α(T ) and the descent δ(T ) are finite, then α(T ) = δ(T ) =: p
and X = rg(T p ) ⊕ ker(T p ).
Chapter 5. Spectrum of linear operators 113

Proof. Let p := α(T ) and q := δ(T ). We divide the proof in several steps.
Step 1. rg(T p ) ∩ ker(T n ) = {0} for every n ∈ N0 .
To see this, choose x ∈ rg(T p ) ∩ ker(T n ). Then there exists a y ∈ X such that
x = T p y, so 0 = T n x = T p+n y. Hence y ∈ ker T p+n = ker T p by Lemma 5.36 i.
It follows that x = T p y = 0.
Step 2. X = rg(T n ) + ker(T q ) for every n ∈ N0 .
For the proof fix x ∈ X. Then T q x ⊆ rg(T q ) = rg(T q+n ). Hence there exists
y ∈ X such that T q x = T q+n y. Then T q (x − T n y) = 0, and therefore x =
T n y + (x − T n y) ∈ rg(T n ) + ker(T q ).
Step 3. α(T ) ≤ δ(T ) = q.
Let x ∈ ker T q+1 . We have to show x ∈ ker T q . By step 2, with n = p,
there exist x1 ∈ rg(T p ) and x2 ∈ ker(T q ) such that x = x1 + x2 . Hence
x1 = x − x2 ⊆ ker(T q+1 ) ∩ rg(T p ) = {0} by step 1. Therefore x = x2 ∈ ker(T q ).

Step 4. δ(T ) ≤ α(T ) = p.

By step 1 and step 2, we have that X = rg(T p ) ⊕ ker(T q ). Since rg(T p+1 ) ∩
ker(T q ) ⊆ rg(T p ) ∩ ker(T q ) = {0}, we also have X = rg(T q+1 ) ⊕ ker(T q ),
implying rg R(T p+1 ) = rg(T p ), hence δ ≤ p.

Theorem 5.39. Let X be a Banach space, T ∈ L(X) a compact operator and

λ ∈ C \ {0}.
(i) ker(λ − T )n is finite dimensional for every n ∈ N0 .
(ii) If U ⊆ X is a closed subspace with U ∩ ker(λ − T )n = {0}, then (λ − T )(U )
is closed and λ − T : U → rg((λ − T )|U has a bounded inverse.
(iii) rg(λ − T )n is closed for every n ∈ N0 .
Pn
Proof. Note that (λ − T )n = λn − n=1 nk λn−k T k and the operator sum is

compact. Hence it suffices to show the assertions for n = 1.

(i) Observe that T |ker(λ−T ) = λ id |ker(λ−T ) . Hence λ id |ker(λ−T ) is compact.
By Remark 5.30 (v) this is case if and only if ker(λ − T ) is finite dimensional.
(ii) Since U ∩ ker(λ − T ) = {0}, the restriction (λ − T )|U is invertible. We

−1
will show that its inverse is bounded. Assume (λ − T )|U is not bounded.
Then there exists a sequence (xn )n∈N such that kxn k = 1 for all n ∈ N and
lim (λ − T )xn = 0. Since T is compact, there exists a convergent subsequence
n→∞
(T xnk )k∈N . Hence

λxnk = T xnk + (λ − T )xnk −→ lim T xnk =: y.

| {z } n→∞
→0

Note that y ∈ U because U is closed. Moreover, y ∈ ker(λ − T ) because

(λ − T )y = (λ − T ) lim xnk = lim (λ − T )xnk = 0.

n→∞ n→∞
114 5.4. Compact operators

Hence y ∈ ker(λ − T ) ∩ U = {0} in contradiction to kyk = lim kλxn k = λ 6= 0.

n→∞

−1
Hence (λ − T )|U : rg(λ − T )|U → U is bounded. Since it is also closed, its
domain rg(λ − T )|U must be closed.
(iii) By (i) we already know that dim ker(λ − T ) < ∞. Then by the following
lemma 5.40 there exists a closed subspace U ⊆ X such that X = ker(λ − T ) ⊕ U .
Hence rg(λ − T ) = rg((λ − T )|U ) is closed by (ii).

Lemma 5.40. Let X be a Banach space and M ⊆ X a finite dimensional

subspace. Then there exists a closed subspace U of X such that X = M ⊕ U .

Proof. Let x1 , . . . , xn a basis of M . Then there exist ϕ1 , . . . , ϕn ∈ M 0 such that

kϕk k = 1 and ϕk (xj )δkj for all j, k = 1, . . . , n. By the Hahn-Banach theorem
0
the ϕk can be extended Pn to functionals ψk ∈ X with kψ k k = 1, k = 1, . . . , n. Let
P : X → X, P x = j=1 ϕj (x)x. Obviously P = P 2 , hence P is a projection.
Note that M = P (X). Hence X = rg(P ) ⊕ ker P = M ⊕ ker P .

Theorem 5.41. Let X be a Banach space, T ∈ L(X) a compact operator and

λ ∈ C\{0}. Then α(λ−T ) = δ(λ−T ) = p < ∞ and X = ker(λ−T )p ⊕rg(λ−T )p .

The number p = α(λ − T ) = δ(λ − T ) is called the Riesz index of λ − T .

Proof. By Lemma 5.38 it suffices to show that α(T ) and δ(T ) are finite.
Assume that α is not finite. Since in this case ker(λ − T ) ( ker(λ − T )2 ( . . .
we can find a sequence (xn )n∈N ⊆ X such that for all n ∈ N
1
kxn k = 1, xn ∈ ker(λ − T )n , and kxn − zk ≥ for all z ∈ ker(λ − T )n−1 .
2
The last condition can be satisfied by the Riesz lemma (Theorem 1.21) because
ker(λ − T )n is closed for all n ∈ N. Then for all 1 ≤ m < n
1
kT xn − T xm k = kλxn −λxm − (λ − T )xn + (λ − T )xm k ≥ .
| {z } 2
∈ker(λ−T )n−1

Therefore (T xn )n∈N does not contain a convergent subsequence in contradiction

to T being compact.
Assume that δ is not finite. Since in this case rg(λ − T ) ) rg(λ − T )2 ) . . . we
can choose a sequence (xn )n∈N ⊆ X such that for all n ∈ N
1
kxn k = 1, xn ∈ rg(λ − T )n , and kxn − zk ≥ for all z ∈ rg(λ − T )n+1 .
2
The last condition can be satisfied by the Riesz lemma because rg(λ − T )n is
closed for all n ∈ N by Theorem 5.39. Then for all 1 ≤ m < n
1
kT xn − T xm k = kλxn −λxm − (λ − T )xn + (λ − T )xm k ≥ .
| {z } 2
∈rg(λ−T )n+1
Chapter 5. Spectrum of linear operators 115

Therefore (T xn )n∈N does not contain a convergent subsequence in contradiction

to T being compact.

Theorem 5.42 (Spectrum of a compact operator). Let X be a Banach

space. For a compact operator T ∈ L(X) the following holds.

(i) If λ ∈ C \ {0}, then λ either belongs to ρ(T ) or it is an eigenvalue of T ,

that is C \ {0} ⊆ ρ(T ) ∪ σp (T ).

(ii) The spectrum of T is at most countable and 0 is the only possible accu-
mulation point.

(iii) If λ ∈ σ(T ) \ {0}, then the dimension of the algebraic eigenspace Aλ (T )

is finite and Aλ (T ) = ker(λ − T )p where p is the Riesz index of λ − T .

(iv) X = ker(λ − T )p ⊕ rg(λ − T )p for λ ∈ σ(T ) \ {0} where p is the Riesz

index of λ − T and ker(λ − T )p and rg(λ − T )p are T -invariant.

(v) σp (T ) \ {0} = σp (T 0 ) \ {0} and σ(T ) = σ(T 0 ). If H is a Hilbert space then

σp (T ) \ {0} = {λ ∈ C : λ ∈ σp (T ∗ )} \ {0} = σp (T ∗ ) \ {0}, where the bar
denotes complex conjugation, and σ(T ) = {λ ∈ C : λ ∈ σ(T ∗ )} = σ(T ∗ ).

Proof. (i) Let λ ∈ C \ {0}. By Theorem 5.41 the Riesz index p of λ − T is

finite. If p = 0, then X = rg(λ − T ) by the proof of Lemma 5.38 (step 2), hence
λ ∈ ρ(T ). If p 6= 0, then λ ∈ σp (T ).
(ii) It suffices to show that for every ε > 0 the set {λ ∈ σ(T ) : |λ| > ε} is finite.
Assume there exists an ε > 0 such that the set is not finite. Then there exists
a sequence (λn )n∈N such that λn 6= λm for n 6= m and |λn | > ε, n ∈ N. Since
σ(T ) \ {0} consists of eigenvalues, we can choose eigenvectors xn of T with
eigenvalues λn . Note that the xn are linearly independent because λn 6= λm
for n 6= m. Let Un := span{x1 , . . . , xn }. Note that all Un are T -invariant,
closed and that U1 ( U2 ( U3 ( . . . . Using the Riesz Lemma, we can choose a
sequence (yn )n∈N such that for all n ∈ N
1
kyn k = 1, yn ∈ Un , and kyn − zk ≥ for all z ∈ Un−1 .
2
Pn
Let 1 ≤ m < n. Note that T ym ∈ Um . Let yn = j=1 αj xj for some αj ∈ C.
Then
n
X n
X
(λn − T )yn = αn (λn − T )xn + αj (T − λn )xj = αj (λj − λn )xj ∈ Un−1 .
j=1 j=1

Hence
1
kT yn − T ym k = kλn yn −(λn − T )yn − T ym k ≥ . (5.5)
| {z } 2
∈Un−1
116 5.4. Compact operators

Therefore (T xn )n∈N does not contain a convergent which contradicts the as-
sumption that T is compact.
(iii) and (iv) follow from Theorem 5.42.
(v) By Schauder’s theorem T 0 is compact (theorem 5.33) Hence for λ ∈ C it
follows that

λ ∈ ρ(T ) ⇐⇒ ker(λ − T ) = {0} and rg(λ − T ) = X

◦
⇐⇒ rg(λ − T 0 ) = {0} and ◦
ker(λ − T 0 ) = X
⇐⇒ rg(λ − T 0 ) = X 0 and ker(λ − T 0 ) = {0}
⇐⇒ λ ∈ ρ(T )

Theorem 5.43 (Fredholm alternative; Riesz-Schauder theory). Let X

be a Banach space, T ∈ L(X) a compact operator and λ ∈ C \ {0}. Then exactly
one of the following is true:

(i) For every y ∈ X the equation (λ−T )x = y has exactly one solution x ∈ X.
(ii) (λ − T )x = 0 has a non-trivial solution x ∈ X.

Proof. (i) is equivalent to λ ∈ ρ(T ) and (ii) is equivalent to λ ∈ σp (T ). Since λ 6=

0, the latter is equivalent to λ ∈ σ(T ). The assertion follows from Theorem 5.42.

A more precise formulation of the Fredholm alternative is the following.

Theorem 5.44. Let X be a Banach space, T ∈ L(X) a compact operator and

λ ∈ C \ {0}. For x, y, ∈ X and ϕ, η ∈ X 0 consider the equations

(A) (λ − T )x = y, (C) (λ − T 0 )ϕ = η,
(B) (λ − T )x = 0, (D) (λ − T 0 )ϕ = 0.

Then

(i) For y ∈ X the following is equivalent:

(a) (A) has a solution x.

(b) ϕ(y) = 0 for every solution ϕ of (D).

(ii) For η ∈ X 0 the following is equivalent:

(a) (C) has a solution ϕ.

(b) η(x) = 0 for every solution x of (B).

(iii) Fredholm alternative: Exactly one of the following holds:

(a) For all y ∈ X and η ∈ X 0 the equations (A) and (C) have exactly one
solution (in particular (B) and (D) have only the trivial solutions).
Chapter 5. Spectrum of linear operators 117

(b) (B) and (D) have non-trivial solutions. In this case dim(ker(λ−T )) =
dim(ker(λ − T 0 )) > 0 and (A) and (C) have solutions if and only if
ϕ(y) = 0 for all solutions ϕ of (D),
η(x) = 0 for all solutions x of (B).

Definition 5.45. Let X, Y be Banach spaces. T ∈ L(X) is called Fred-

holm operator if rg(T ) is closed and n(T ) := dim(ker T ) < ∞ and d(T ) :=
codimY (rg T ) := dim(Y / rg(T )) < ∞. In this case, χ(T ) := n(T ) − d(T ) is
called the Fredholm index.

Proof of Theorem 5.44. . . . . . . . . . . . .

Now we return to the spectrum of compact operators.

Lemma 5.46. Let H be Hilbert space, 6= {0}, and T ∈ L(H) a selfadjoint

compact operator. Then at least one the values kT k or −kT k is an eigenvalue
of T . In particular, if T 6= 0, then T has at least one eigenvalue distinct from
0.

Proof. If kT k = 0, the assertion is clear. Now assume that kT k 6= 0. Recall

that kT k = sup{|hT x , xi| : x ∈ X, kxk = 1} (Theorem 4.45).
By Lemma 5.27 the numbers m = inf{hT x , xi : x ∈ X, kxk = 1} and M =
inf{hT x , xi : x ∈ X, kxk = 1} belong to the spectrum of T . Since T is compact
and kT k = 6 0, it follows that ∅ =
6 {±kT k} ∩ σ(T ) = {±kT k} ∩ σp (T ).

Theorem 5.47 (Spectral theorem for compact selfadjoint operators).

Let H be a Hilbert space and T ∈ L(H) a compact selfadjoint operator.
(i) There exists an orthonormal system (en )N
n=1 of eigenvectors of T with
eigenvalues (λn )N
n=1 where N ∈ N ∪ {∞} such that
N
X
Tx = λn hx , en i en , x ∈ H. (5.6)
n=1

The λn can be chosen such that |λ1 | ≥ |λ2 | ≥ · · · > 0. The only possible
accumulation point of the sequence (λn )n∈N is 0.
(ii) If P0 is the orthogonal projection on ker T , then
N
X
x = P0 x + hx , en i en , x ∈ H. (5.7)
n=1

(iii) If λ ∈ ρ(T ), λ 6= 0
N
−1 −1
X hx , en i
(λ − T ) x=λ P0 x + en , x ∈ H.
λ −λ
n=1 n
118 5.4. Compact operators

Proof. (i) Let X1 = X and T1 = T . If T 6= 0, then there exists a λ1 ∈ σp (T1 )

such that |λ1 | = kT1 k =
6 0. Let B1 be an orthonormal basis of ker(λ1 −T1 ). Note
that B1 is finite because T is compact (Theorem 5.42). Let X1 := ker(λ1 −T )⊥ =
rg(λ1 − T ) = rg(λ1 − T ). Here we used that T is selfadjoint and consequently
λ ∈ σp (T ) ⊆ R. By Theorem 5.42, X2 is T1 -invariant, hence T2 := T1 |X2 ∈
L(X2 ). Obviously, T2 is selfadjoint and compact. If T2 6= 0, then there exists
a λ2 ∈ σp (T2 ) such that |λ2 | = kT2 k 6= 0. Let B2 be an orthonormal basis
of ker(λ2 − T2 ). Note that B1 is finite because T is compact (Theorem 5.42).
Hence B1 ∪ B2 is an orthonormal basis of span{ker(λ1 − T ), ker(λ2 − T )}. Let
X3 := span{ker(λ1 −T ), ker(λ2 −T )}⊥ and T3 := T2 |X3 . Continuing like this we
obtain a sequence of Banach spaces Xn and a sequence of compact selfadjoint
operators Tn ∈ L(Xn ). Let x ∈ X. Define
X
xn+1 = x − hx , en i en ∈ Xn+1 .
en ∈B1 ∪...Bn

It follows that
X
kT x − T hx , en i en k = kTn+1 xn+1 k ≤ |λn+1 |kxk −→ 0, n → ∞.
en ∈B1 ∪...Bn

This implies that

N
X N
X
Tx = hx , en iT en = λn hx , en i en .
n=1 n=1

(ii) Note that . . .

(iii)

Corollary 5.48. Let H be a Hilbert space and T ∈ L(H) a compact selfadjoint

operator. There exists a sequence (Pn )N
n=1 of pairwise orthogonal projections
with N ∈ N ∪ {∞} and a sequence |λ1 | ≥ |λ2 | ≥ . . . such that
N
X
T = λn Pn (5.8)
n=1

where the series converges to T in the operator norm. If (λn )n is an infinite

sequence, then lim λn = 0. The representation (5.8) is unique if the λn are
n→∞
pairwise distinct.

Proof. If the series is a finite sum, the assertion is clear. Now assume
P∞ that the
series is an infinite. Note that for every k ∈ N the operator n=k λn Pn is
normal and that the norm of a normal operator is equal to maximum of the
moduli of the elements of its spectrum (Theorem 5.19). Since |λk+1 | → 0 for
k → ∞ the claim follows from
Xk
T − λn Pn = sup{|λn | : n ≥ k + 1} = |λk+1 |.

n=1
Chapter 5. Spectrum of linear operators 119

The representation (5.8) allows us to define the root of a positive compact

selfadjoint operator.

Theorem 5.49. Let H be a Hilbert space and K ∈ L(H) a compact operator.

(i) T is positive ⇐⇒ all eigenvalues of T are positive.
T is strictly positive ⇐⇒ all eigenvalues of T are strictly positive.
(ii) If T is positive and k ∈ N then there exists exactly one positive compact
selfadjoint operator R such that Rk = T .

Note that the theorem does not imply that there cannot be non-compact oper-
ators A ∈ L(H) such that A2 = T . In Corollary 5.60 we will show that every
bounded positive selfadjoint operator has a unique positive root.
Proof of Theorem 5.49. Recall that a linear operator T is positive if and only if
hT x , xi ≥ 0 for all x ∈ H. Let P0 , λn and en as in (5.7). Then (i) follows from
DX X E X
hT x , xi = λn hx , en i en , P0 x + λn hx, en , ein = λn |hx , en i|2 ≥ 0.
n n n
P 1/k k
For the proof of (ii) define R = n λn h· , en i en . Obviously R = T . To
show uniqueness, assume that there exists a compact selfadjoint positive linear
k
P S such that S = T . Since S is compact, it has a representation
operator
S = n µn Qn with pairwise orthogonal projections Qn . By assumption
X
T = Sk = µkn Qn .
n

Hence the µn are the kth roots the eigenvalues λn of T , so S = R.

Definition 5.50. Let H be a Hilbert space and T ∈ L(H) a positive selfadjoint

1
compact operator. Then |T | := (T ∗ T ) 2 . The non-zero eigenvalues sn of |T | are
the singular values of T .

Obviously |T | and |T ∗ | are positive selfadjoint compact operators.

Lemma 5.51. (i) k |T |x k = kT xk and k |T ∗ |y k = kT ∗ yk and for x ∈ H1

and y ∈ H2 .
(ii) s is a singular value of T if and only if s2 is an eigenvalue of T ∗ T and
T T ∗.

Proof. (i) For all x ∈ H1

k |T |x k2 = h|T |x , |T |xi = h|T |2 x , xi = hT ∗ T x , xi = kT xk2 .

An analogous calculation shows k |T ∗ |y k = kT ∗ yk and for y ∈ H2 .

(ii) follows from the uniqueness of the representation (5.8).
120 5.4. Compact operators

Note that |T | can be defined more generally for positive selfadjoint operators
on a Hilbert space H, see Definition 5.61.
A representation similar to (5.6) exists for arbitrary compact operators.

Theorem 5.52. Let H1 , H2 be Hilbert spaces and T ∈ L(H1 , H2 ) a compact

operator.
(i) Let s1 ≥ s2 ≥ · · · > 0 be the singular values of T and (ϕn )N
n=1 ⊆ H1 and
(ψn )N
n=1 ⊆ H 2 such that
N
X
Tx = sn hx , ϕn iψn , x ∈ H1 ,
n=1
N
X
T ∗y = sn hy , ψn iϕn , y ∈ H2 .
n=1

If there are infinitely many sn , then lim sn = 0.

n→∞
(ii) The non-zero eigenvalues of |T | and |T ∗ | coincide and are equal to the sn .
The s2n are the eigenvalues of T ∗ T and T T ∗ . Moreover, the ψn = s1n T ϕn
are eigenvectors of T ∗ .

Proof. (i) Let (ϕn )n∈N ⊆ H1 a ONS such that, see Theorem 5.47,
N
X N
X
|T |x = sn hx , ϕn iϕn , T ∗T x = s2n hx , ϕn iϕn .
n=1 n=1

1
Let ψn := xn T ϕn . Then (ψn )n∈N is an ONS in H2 because
1 1 1
hψn , ψm i = hT ϕn , T ϕm i = 2 hT ∗ T ϕn , ϕm i = 2 s2n δnm = δnm .
s2n sn sn
Moreover
1 s2
T T ∗ ψn = T T ∗ T ϕn = n T ϕn = s2n ψn .
sn sn
Hence σp (T ∗ T )\{0} = {s2n : 1 ≤ n ≤ N } ⊆ σp (T T ∗ )\{0}. Similarly the reverse
inclusion can be shown, so that σp (T ∗ T ) \ {0} ⊆ σp (T T ∗ ) \ {0}.
(ii) . . .

Theorem 5.53 (Min-Max-Principle). Let H1 , H2 be Hilbert spaces, K ∈

L(H1 , H2 ) a compact operator with singular values s1 ≥ s2 ≥ s3 ≥ . . . . Then
s1 = kKk and for n ≥ 2
n o
sn+1 = inf sup kKxk : x ∈ H1 , x ⊥ span{x1 , . . . , xn }, kxk = 1 .
x1 ,...xn ∈H1

Proof. . . .
Chapter 5. Spectrum of linear operators 121

5.5 Hilbert-Schmidt operators

Definition 5.54. Let H1 , H2 be Hilbert spaces and K ∈ L(H1 , H2 ). K is called
a Hilbert-Schmidt operator if and only if there exists an ONB (eλ )λ∈Λ of H1 such
that
X
kK eλ k2 < ∞.
λ∈Λ

The set of all Hilbert-Schmidt operators from H1 to H2 is denoted by HS(H1 , H2 ).

Theorem 5.55. Let H1 , H2 be Hilbert spaces.

(i) A operator K ∈ L(H1 , H2 ) is a Hilbert-Schmidt operator if and only if K ∗

is a Hilbert-Schmidt operator. In this case:
X X X
kK eα k2 = kK eβ k2 = kK eλ k2 < ∞
α∈A β∈B λ∈Λ

for all ONBes (eα )α∈A of H1 and (eβ )β∈B of H2 .

(ii) Every Hilbert-Schmidt operator is compact.

(iii) Let K ∈ L(H1 , H2 ) be a compact operator with singular values x1 ≥ x2 ≥

s3 ≥ . . . . Then K is a Hilbert-Schmidt operator if and only if K ∗ is a
Hilbert-Schmidt operator if and only if
X
s2n < ∞.
n

Theorem 5.55 (i) shows that for K ∈ HS(H1 , H2 ) the Hilbert-Schmidt norm
X  12
kKkHS := kK eα k2 for an ONB (eα )α∈A .
α∈A

is well-defined.

Proof of Theorem 5.55. (i)P Let K be a Hilbert-Schmidt operator and (eλ )λ∈Λ
an ONB of H1 such that λ∈Λ kK eλ k2 < ∞. For an arbritray ONB (ψβ )β∈B
of H2 we find, using Parseval’s equality (Theorem 4.31) First we show that K ∗
is also a Hilbert-Schmidt operator.
2
X X X ∗
XX
∗
kK ψβ k =2
hK ψβ , eλ i eλ = |hK ∗ ψβ , eλ i|2



β∈B β∈B λ∈Λ λ∈Λ β∈B
XX X
2
= |hψβ , K eλ i| = kKψλ k2 < ∞.
λ∈Λ β∈B λ∈Λ
122 5.5. Hilbert-Schmidt operators

In particular, the Hilbert-Schmidt norm of K ∗ does not depend on the chosen

ONB of H2 . Applying the same proof to K ∗ , it follows that the Hilbert-Schmidt
norm of K = K ∗∗ does not depend on the chosen ONB of H1 . For the proof of
kKk ≤ kKkHS we observe that every x ∈ H1 with kxk = 1 can be extended to
a ONB of H1 . Hence

kKkHS ≥ kKxk ≥ sup{kKyk : y ∈ H1 , kyk = 1} = kKk.

(ii) Let (eλ )λ∈Λ an ONS of H1 and (en )n∈N a subset containing all eλ with
K eλ 6= 0 (this family is at most countable by Lemma 4.27). For n ∈ N let Pn be
the orthogonal projection on {e1 , . . . , en }. Note that all Pn are compact because
they have finite-dimensional range. Since K is a Hilbert-Schmidt operator, we
find that
∞
X
kK − KPn k2 = kK(id −Pn )k2 ≤ kK(id −Pn )k2HS = kK en k2 −→ ∞,
m=n+1

in particular K is compact because it is the norm limit of compact operators.

(iii) Assume that K is compact. By Theorem 5.52 P we can choose ONSs
(ϕn )λ∈N of H1 and (ψn )λ∈N of H2 such that Kx = n∈N sn hx , ϕn iψn where
s1 ≥ s2 ≥ · · · ≥ 0 are the singular values of K.
If K is a Hilbert-Schmidt operator, then
N
X N
X
s2n = kKϕn k2 ≤ kKk2HS < ∞.
n=1 n=1
PN
Now assume that n=1 s2n < ∞ and choose an arbitrary ONB of H1 containing
(ϕn )n∈N . It follows that

X N
X N
X
kKϕλ k2 = kKϕn k2 ≤ kKk2HS = s2n < ∞,
λ∈Λ n=1 n=1

implying that K is a Hilbert-Schmidt operator.

Lemma 5.56. The finite-rank operators are dense in the Hilbert-Schmidt op-
erators.

Proof. Let H be a Hilbert space and S ∈ HS(H). In particular, S is compact

PN
and there exist ONBs (ϕn )n∈N and (ψn )n∈N such that S = n=1 sn h· , ϕn iψn .
PM
For M ∈ N let us define SM = n=1 sn h· , ϕn iψn . Then kS − SM k2 ≤ kS −
N
SM k2HS = n=M +1 s2n → 0 for M → ∞.
P

An important class of examples is given in the following theorem.

Theorem 5.57. Let H = L2 (0, 1) and T ∈ L(H). Then the following is equiv-
alent:
Chapter 5. Spectrum of linear operators 123

(i) T is a Hilbert-Schmidt operator.

(ii) There exists a k ∈ L2 (0, 1)2 such that
Z 1
(T x)(t) = k(s, t)x(s) ds.
0

In this case we write Tk for T .

If one of the equivalent conditions holds, then
Z 1Z 1 2
kT k = |k(s, t)|2 ds. dt = kkkL2 (0,1)2 .
0 0

Proof. (ii) =⇒ (i) Let (en )n be an ONB of L2 (0, 1). Then also (en )n is an
ONB of L2 (0, 1) (where en denotes the to en complex conjugated function) and
we find
X∞ X∞ Z 1 Z 1 2 ∞ Z 1
X
2
|hk(·, t) , en i|2 dt
 
kT en k = 
 k(s, t) en (s) ds dt =

n=1 n=1 0 0 n=1 0
Z 1X∞
= |hk(·, t) , en i|2 dt (5.9)
0 n=1
Z 1
= kk(·, t)k2 dt (5.10)
0
Z 1 Z 1
= |k(s, t)|2 ds dt = kkkL2 (0,1)2 .
0 0

In (5.9) we have used the monotone convergence theorem to exchange the

sum and the integral (Theorem A.18) and in (5.10) we used Parseval’s equal-
ity (Theorem 4.31). It follows that T is a Hilbert-Schmidt operator and that
kT kHS = kkkL2 (0,1)2 .
(i) =⇒ (ii) By the proof we have an isometry

Ψ : L2 (0, 1)2 → HS(L2 (0, 1)), Ψk = Tk .

We will show that the range of Ψ is dense in HS(H). By Lemma 5.56 it suffices
to show that rg(Ψ) containsPthe finite-rank operators. Let T be of finite rank.
n0
Then T is of the form T = n=1 h· , xn iyn so that for every f ∈ H
n0
X n0 Z 1
X Z 1X n0 
T f (t) = hf , xn iyn (t) = f (s)xn (s)yn (t) ds = xn (s)yn (t) f (s) ds.
n=1 n=1 0 0 n=1

This shows that T ∈ rg Ψ. Fix S ∈ HS(H) and choose a sequence (Sn )n∈N in
the range of Ψ. Since Ψn is an isometry, it follows that (Ψ−1 Sn )n∈N is Cauchy
sequence in H, hence its limit exists. Using the continuity of Ψ we find
 
S = lim Sn = lim ΨΨ−1 Sn = Ψ lim Ψ−1 Sn ∈ rg(Ψ).
n→∞ n→∞ n→∞
124 5.6. Polar decomposition

Theorem 5.58. Let H1 , H2 be Hilbert spaces.

 
(i) HS(H1 , H2 ), k · kHS is a normed spaces. The norm is induced by the
inner product
X
hS , T iHS = hS eα , T eα i, S, T ∈ HS(H1 , H2 ),
α

for an arbitrary ONB (eα )α∈A of H1 .

(ii) Let T ∈ HS(H1 , H2 ) and A a bounded linear operator between appropriate

Hilbert spaces. Then AT and T A are Hilbert-Schmidt operators and

kAT kHS ≤ kAk kT kHS , kT AkHS ≤ kAk kT kHS .

(iii) HS(H) is a two-sided ideal in L(H).

Proof. Note that (a+b)2 = a2 +b2 +2ab = a2 +b2 −(a−b)2 +a2 +b2 ≤ 2(a2 +b2 )
for a, b ∈ R.
(i) Let S, T ∈ HS(H1 , H2 ) and λ ∈ C. Then obviously λS ∈ HS(H1 , H2 ). To
show that S + T ∈ HS(H1 , H2 ) fix an ONS (eλ )λ∈Λ of H1 . Using the above
remark is follows that
X X X
k(S + T ) eλ k2 ≤ (kS eλ k + kT eλ k)2 ≤ 2 kS eλ k2 + kT eλ k2 < ∞.
λ∈Λ λ∈Λ λ∈Λ

It follows that h· , ·iHS is well-defined. The properties of an inner product are

clear. In particular, kT kHS = hT , T iHS for T ∈ HS(H1 , H2 ).
(ii) Note that
X X
kAT eλ k2 ≤ kAk2 kT eλ k2 = kAk2 kT k2HS ,
λ∈Λ λ∈Λ

so AT is a Hilbert-Schmidt operator. It follows that T A = (A∗ T ∗ )∗ is also a

Hilbert-Schmidt operator with norm kT AkHS = k(A∗ T ∗ )∗ kHS = kA∗ T ∗ kHS ≤
kA∗ k kT ∗ kHS ≤= kAk kT kHS .
(iii) is a consequence of (i) and (ii).

5.6 Polar decomposition

Theorem 5.59. Let H be a Hilbert space and T ∈ L(H) a selfadjoint operator
with T ≥ 0. Then there exists exactly one R ∈ L(H) such that R ≥ 0 R2 = T .
In addition, if S ∈ L(H) commutes with T , then S commutes with R.
√
the operator R is called the root of T and is denoted by T.
Chapter 5. Spectrum of linear operators 125

Proof. Without restriction we can assume kT k ≤ 1, hence 0 ≤ T ≤ id. Now

assume that a solution R ∈ L(H) of R2 = T exists. Let A := id −T and
X := id −R. Note that

id −A = T = R2 = (1 − X)2 = id −2X + X 2 .

Note that 0 ≤ R ≤ id if and only if 0 ≤ X ≤ id. Hence R is a non-negative

solution of R2 = T if and only if X is a non-negative solution of

1
X= (A + X 2 ). (5.11)
2
Step 1. Construction of a solution of (5.11).
We define
1 2
X0 := id, Xn := (A + Xn−1 ), n ∈ N.
2
Note that every Xn is a polynomial in A with positive coefficients and that
Xn Xm = Xm Xn for all n, m ∈ N. Since A is positive, this implies that all Xn
are positive. We will show the following properties of the sequence (Xn )n∈N by
induction.

(i) Xn − Xn−1 is a polynomial in A with positive coefficients, so that in

particular Xn − Xn−1 ≥ 0.

(ii) kXn k ≤ 1.

All assertions are clear in the case n = 0 (with X−1 := 0). Now assume that
the assertions are true for some n ∈ N. Note that
1 1 1
Xn+1 − Xn = (A + Xn2 ) − (A + Xn−1
2
) = (Xn2 − Xn−1
2
)
2 2 2
1
= (Xn − Xn−1 )(Xn + Xn−1 ).
2
Since by induction hypothesis both terms in the second line are polynomials in
A with positive coefficients, (i) is proved for n + 1. (ii) follows from kXn+1 k ≤
1
2 (kAk + kXn+1 k) ≤ 1.
Since (Xn )n∈N is uniformely bounded monotonically increasing sequence in,
there exists an X ∈ L(H) such that X = s- lim Xn and kXk ≤ lim inf n→∞ kXn k ≤
n→∞
1 (see Exercise 4.25).
Now let S ∈ L(H) with ST = T S. By definition of A, then also SA = AS and
Xn S = Xn S for all Xn since the Xn are polynomials in A. For every x ∈ H we
therefore obtain

0 ≤ kSXx − XSxk = lim kSXn x − Xn Sxk = lim kSXn x − SXn xk = 0.

n→∞ n→∞
126 5.6. Polar decomposition

Since all Xn commute with T , it follows that Xn X = Xn X for all n ∈ N, so

that for all x ∈ X

k(Xn2 − X 2 )xk = k(Xn − X)(Xn + X)xk ≤ 2k(Xn − X)2 xk −→ 0, n → ∞,

which shows that X 2 = s- lim Xn2 . Therefore X solves (5.11) because

n→∞

1 1 1
X = s- lim Xn = s- lim (A + Xn2 ) = (A + s- lim Xn2 ) = (A + X 2 ).
n→∞ n → ∞2 2 n→∞ 2
Setting R = id −X we obtain a bounded selfadjoint solution of R2 = T with
0 ≤ R ≤ id.
Step 2. Uniqueness of the solution.
Let R0 ∈ L(H) be solution of R2 = T with R0 ≥ 0. Then R and R0 commute
because

R0 A = R0 (R0 )2 = (R0 )2 R0 = AR0 .

It follows that

(R − R0 )R(R − R0 ) + (R − R0 )R0 (R − R0 ) = (R2 − R02 )(R − R0 ) = 0.

Since both operators on the left hand side are non-negative, it follows that both
of them are 0 and therefore

(R − R0 )4 = (R − R0 )R(R − R0 ) − (R − R0 )R0 (R − R0 ) = 0.

Since R − R0 is normal, it follows that k(R − R0 )k4 = k(R − R0 )k4 .

Corollary 5.60. If S, T ∈ L(H) are positive and ST = T S, then also ST is

positive.

Proof. By Theorem 5.59 the root of T exists, is selfadjoiunt and commutes with
S. Hence for all x ∈ H
√ √ √ √ √ √
hST x , xi = hS T T x , xi = h T S T x , xi = hS T x , T xi ≥ 0.
1
Definition 5.61. For T ∈ L(H) we define |T | := (T ∗ T ) 2 .

Definition 5.62. Let H1 , H2 be Hilbert spaces and U ∈ L(H1 , H2 ). U is called

a partial isometry if U |(ker U )⊥ is an isometry. ker U ⊥ is called its initial space.

Note that U is an partial isometry if and only if

U |(ker U )⊥ : (ker U )⊥ → rg(U )

is unitary.
Chapter 5. Spectrum of linear operators 127

Theorem 5.63 (Polar decomposition). Let H1 , H2 be Hilbert spaces and

T ∈ L(H1 , H2 ). Then there exists a partial isometry U ∈ L(H1 , H2 ) such that
T = U |T |. If in addition the initial space of U is (ker T )⊥ , then U is unique.

Proof. Note that k |T |x k2 = kT xk2 for all x ∈ H1 because

1 1
k |T |x k2 = h|T |x , |T |xi = h(T ∗ T ) 2 x , (T ∗ T ) 2 xi = hT ∗ T x , xi = hT x , T xi = kT xk2 .

We define

U : rg(|T |) → rg(T ), U (|T |x) = T x.

U is well-defined because for x, y ∈ H1 with |T |x = |T |y it follows that kT x −

T yk = k|T |x − |T |yk = 0 hence T x = T y. U is and isometry because kT xk =
k |T |x k for all x ∈ H as shown above. In particular, kU k = 1 and has a unique
continuous extension to rg(|T |) → rg(T ). Now we extend U to H1 by setting
⊥
U x = 0 for all x ∈ rg(|T |) = ker(|T |) = ker T .
Chapter A. Lp spaces 129

Appendix A

Lp spaces

Spaces of integrable functions play an important role in applications. As the

norm of a function f : [a.b] → K one could consider
Z b
kf k := |f (t)| dt,
a

or more generally, for some 1 ≤ p < ∞

! p1
Z b
p
kf kp := |f (t)| dt .
a

It can be shown that k·kp is a norm on C([a, ]). However the space of continuous
functions C([a, b]) is not complete for the norm k · k1 . For example, let
(
tn , 0 ≤ t ≤ 1,
fn : [0, 2] → R, fn (t) =
1, 1 < t ≤ 2.

All fn are continuous and it is easy to check that kfn −fm k1 → 0 for n, m → ∞.
So the fn form a Cauchy sequence, but it is not convergent. (If it were, then
there must exist a continuous function g such that
Z 2 Z 1 Z 2
|fn (t) − g(t)| dt = |fn (t) − g(t)| dt + |fn (t) − g(t)| dt −→ 0
0 0 1

for n → ∞. Hence g(t) = 0 for t ∈ (0, 1) and g(t) = 0 for t ∈ (1, 2) which is
impossible for a continuous function by the intermediate value theorem.
If we extend the space of functions to the Riemann integrable functions R([a, b]),
then the sequence above does converge to χ[1,2] . But there are several other
problems with the space of Riemann integrabel functions.
For example, let Q∩[0, 1] = {qn : n ∈ N}. Then all characteristic functions χn :=
χ{q1 , ...,qn } are Riemann integrable, kχn k1 = 0, the χn form a Cauchy sequence,
130 A.1. A reminder on measure theory

the pointwise limit exists and is χQ∩[0,1] which is not Riemann integrabel. This
example shows that k·k1 is only a seminorm on R([a, b]), that Cauchy sequences
do not need to converge and that in general pointwise limit and integral cannot
be exchanged. The pointwise limit of a sequence of Riemann integrable functions
does not need to be Riemann integrable.
Recall that the Riemann integral of a function f : [a, b] → R is obtained as
the limit of Riemann sums when the interval [a, b] is divided in small pieces.
Lebesgue’s approach is to divide the range of the function in small pieces and
then measure the “size” of the pre-image. Hence admissible are functions whose
pre-images of intervals can be measured in some sense.

A.1 A reminder on measure theory

Definition A.1. Let T be a set and Σ ⊂ PT a family of subsets of T .

(i) Σ ⊂ PT is called a ring if for all A, B ∈ Σ also A ∪ B and A \ B belong to

Σ.
S
(ii) Σ ⊂ PT is called a σ-ring if it is a ring and n∈N An ∈ Σ for all (An )n∈N ⊆
Σ.
(iii) Σ ⊂ PT is called an algebra if Σ is a ring and T ∈ Σ, that is

(a) ∅ ∈ Σ,
(b) A ∈ Σ =⇒ T \ A ∈ Σ,
(c) A, B ∈ Σ =⇒ A ∪ B ∈ Σ.
S
(iv) Σ ⊂ PT is called a σ-algebra if it is a algebra and n∈N An ∈ Σ for all
(An )n∈N ⊆ Σ.

Note that for A, B ∈ Σ also A ∩ B = A \ (T \ B) ∈ Σ if Σ is an algebra.

Remark. Σ is indeed a ring in the algebraic sense if one sets A + B := (A ∪

B) \ (A ∩ B) and A · B := A ∩ B.

Definition A.2. Let T be a set with a σ-algebra Σ. A measure on Σ is a

function µ : Σ → [0, ∞] such that

(i) µ(∅) = 0,
S∞ P∞
(ii) (An )n∈N ⊆ A with pairwise disjoint An =⇒ µ( n=1 An ) = n=1 µ(An ).

Obviously, the intersection of rings is again a ring and PT is a ring. Hence, given
a family U of subsets of T , there exists a smallest ring containing U, namely the
intersection of all rings that contain U. This ring is called the ring generated by
U. Analogously the σ-ring, the algebra and the σ-algebra generated by U are
obtained.
Chapter A. Lp spaces 131

Example A.3. The smallest σ-algebra containing all intervals of R is called

the Borel sets.
More generally, let (T, O) be a topological space. Then the Borel sets is the
σ-algebra generated by O.

The aim is to assign a measure µ(U ) to every Borel set U ⊆ R such that the
measure of intervals is its length.

Definition A.4. Let T be a set with a ring Σ of sets. A pre-measure µ on

(T, Σ) is a function

µ : Σ → [0, ∞]

such that µ(∅) = 0 and

[ [ X
(An )n∈N ⊆ Σ, pairwise disjoint and An ∈ Σ =⇒ µ( An ) = µ(An ).
n∈N n∈N n∈N

Note that a pre-measure is monotonic: if A, B ∈ Σ and A ⊆ B, then µ(A) ≤

µ(B).

Example A.5.R Let A be the set of all finite unions of finite intervals and
define µ(A) := R χA dx where χA is the characteristic function of A. Then µ
is a pre-measure on A.
Proof. Obviously A is a ring, for every A ∈ A the characteristic function χA is
S integrable and µ(∅) = 0. S
Riemann Now let (An )Pn∈N ⊆ A be pairwise disjoint
n n
with n∈N An ∈ A. Obviously, µ( n=1 An ) = n=1 µ(An ) for every n ∈ N.
For n ∈ N define Bn := A \ (A1 ∪ · · · ∪ An ). Obviously, Bn ∈ A and
n
[
µ(A) = µ( Ak ) + µ(Bn ).
k=1
S∞
To prove that µ(A) = µ( k=1 Ak ) it suffices to show that lim µ(Bn ) = 0. Fix
n→∞
ε > 0. Since B ∈ A there exists an compact set Cn ⊆ Bn with µ(Bn \ Cn ) <
2−n ε. Let Dn := C1 ∩ · · · ∩ Cn . Then all Dn are compact and Dn ⊆ Cn ⊆ Bn .
By construction, B1 ⊇ B2 ⊇ B3 . . . , hence
n
[ n
[
µ(Bn \ Dn ) = µ(Bn \ (C1 ∩ · · · ∩ Cn ) = µ( (Bn \ Ck )) ≤ µ( (Bk \ Ck ))
k=1 k=1
n
X n
X
≤ µ(Bk \ Ck ) < 2−k ε = ε.
k=1 k=1
T∞ T∞
On the other hand, n=1 Dn ⊆ n=1 Bn = ∅. Since all Dn are compact and
D1 ⊇ D2 ⊇ . . . , there exists an K ∈ N such that Dn = ∅, n ≥ K. Hence
µ(Bn ) = µ(Bn \ Dn ) < ε for all n ≥ k.
132 A.1. A reminder on measure theory

In order to measure all Borel sets, we have to show that the pre-measure of
Example A.5 can be extended to the Borel sets.

Theorem A.6 (Hahn). Let T be a set, A a ring on T and µ̃ a pre-measure

on A. Let Σ(A) be the σ-algebra generated by A. Then
∞
nX ∞
[ o
µ(A) := inf µ̃(An ) : (An )n∈N ⊆ Σ, A ⊆ An .
n=1 n=1
S∞
If µ̃ is σ-finite, i. e., if there exist An ∈ Σ with µ(An ) < ∞ and T = n=1 An ,
then the extension µ is unique.

For the proof, we first show that µ̃ can be extended to an outer measure µ∗ on
PT . Then, by the lemma of Carathéodory, the restriction of the outer measure
to the set of the µ∗ -measurable sets is a measure.

Definition A.7. Let A be a σ-algebra on T . An outer measure on A is a

function µ∗ : A → [0, ∞] such that
(i) µ∗ (∅) = 0,
(ii) A, B ∈ A with A ⊆ B =⇒ µ∗ (A) ≤ µ∗ (B).
S∞ P∞
(iii) (An )n∈N ⊆ A =⇒ µ∗ ( n=1 An ) ≤ n=1 µ∗ (An ).
A set A ∈ A is called µ∗ -measurable if

µ∗ (Z) = µ∗ (Z ∩ A) + µ∗ (Z \ A), Z ∈ A.

Lemma A.8. With the assumptions of Hahn’s theorem (Theorem A.6)

∞
nX ∞
[ o
µ∗ (A) := inf µ̃(An ) : (An )n∈N ⊆ A, A ⊆ An
n=1 n=1

defines an outer measure on PT . In addition, µ(A) = µ∗ (A) for A ∈ A.

Proof. Properties (i) and (ii) of an outer measure are clear. Now let (An )n∈N ⊆
j
S∞ andj ε > 0. Then there exists a family (Bn )n,j∈N ⊆ A such that An ⊆
PT
B
j=1 n and
∞
X ε
µ(Bnj ) ≤ µ∗ (An ) + , n ∈ N.
j=1
2n

Bnj and
S S
By construction A := n∈N An ⊆ n,j∈N

∞
X X
µ∗ (A) ≤ µ(Bnj ) ≤ µ∗ (An ) + ε.
n,j=1 n=1
Chapter A. Lp spaces 133

P∞
Note that (Bnj )n,j∈N is countable, hence we have proved µ∗ (A) ≤ n=1 µ∗ (An ).
∗
Now let A ∈ A. Clearly, µ(A) S ≤ µ (A) holds. Now fix ε > 0 and choose
(An )n∈N ⊆ A such that A ⊆ n∈N An and
∞
X
µ(An ) ≤ µ∗ (A) + ε. n ∈ N.
n=1
S
Since A = n∈N (An ∩ A) and µ is a pre-measure on A, it follows that
∞
X ∞
X
µ(A) ≤ µ(An ∩ A) ≤ µ(An ) ≤ µ∗ (A) + ε. n ∈ N.
n=1 n=1

Since this is true for all ε > 0, it follows that µ(A) ≤ µ∗ (A).

Lemma A.9 (Carathéodory). Let µ∗ be an outer measure on PT . Then the

set M of all µ∗ -measurable sets is a σ-algebra and µ∗ is a measure on M.

Proof. . . . . . .

Proof of Theorem A.6. It suffices to show that the set of the µ∗ -measurable sets
contains A . . .

Hahn’s theorem gives the desired measure on the Borel sets.

Definition A.10 (Lebesgue completion). Let (T, Σ, µ) be a measure space.

A ⊆ T is a zero set if there exists a B ⊆ T with µ(B) = 0 and A ⊆ B (note
that A does not necessarily belong to Σ). The σ-algebra generated by Σ and
the zero sets is called the Lebesgue completion.

The measure on the completion of the Borel sets in R is the Lebesgue measure,
usually denoted by λ. .

A.2 Integration
In the following, I is always an interval in R.

Definition A.11. A function f : I → R is called measurable if for every (a, b) ⊆

R its preimage f −1 ((a, b)) is a Borel set.

More generally, let (T, ΣT , µT ) and (S, ΣS , µS ) be measure spaces. A function

f : T → S is called measurable if for every U ∈ ΣS also f −1 (U ) ∈ ΣT .

Example A.12. Let E be a Borel set. Then the characteristic function χE is

measurable.
134 A.2. Integration

Definition A.13. Let (T, Σ, µ) be a measure space. A function f : T → C is

called a simple function if there are Ek ∈ Σ and αk ∈ C such that
n
X
f= αk χEk .
k=1

It is easy to see that simple functions are measurable. Note, however, that the
sum representation of a simple function is not unique.
The next theorem lists important properties of measurable functions.

Theorem A.14. Let I be an interval in R and fn , f, g : I → R be functions.

(i) If f and g are measurable, then so are f +g, f g, fg (if it exists), max{f, g}
and min{f, g}.
(ii) Every continuous function is measurable.
(iii) If all fn are measurable and f is their pointwise limit (i. e. f (t) = lim fn (t) =
n→∞
f (t), t ∈ I), then f is measurable.
(iv) If f is measurable, then there exists a sequence (ϕn )n∈N of simple functions
that converges pointwise to f . If in addition f ≥ 0, then the sequence can
be chosen such that ϕn (t) % f (t), t ∈ I.
(v) If f is measurable and bounded, then there exists a sequence (ϕn )n∈N of
simple functions that converges uniformly to f .

The theorem says that the set of the measurable functions are a vector space
and that it is stable under taking pointwise limits.
Next we introduce the integral for positive functions.

Definition A.15. Let (T, Σ, µ) be a measure space.

Pn
(i) Let f = k=1 αk χEk with Ek ∈ Σ and αk ∈ [0, ∞] a simple function. We
define its integral as
Z n
X
f dµ = αk µ(Ek ).
T k=1

(ii) Let f : T → [0, ∞] be a measurable function. Choose a sequence (ϕn )n∈N

of simple functions with ϕ1 ≤ ϕ2 ≤ . . . that converges pointwise to f . We
define the integral of f by
Z Z
f dµ = lim ϕn dµ.
T n→∞ T

Of course, it must be proved that the integral in (i) does not depend on the sum
representation of the simple function, and that the limit in (ii) does not depend
on the chosen sequence of simple functions.
Chapter A. Lp spaces 135

Definition A.16. Let I be an interval in R.

R
(i) A function f : I → [0, ∞] is called (Lebesgue) integrable if I
f dλ < ∞.
+
(ii) A function f : I → R is called (Lebesgue) integrable if f := max{f, 0}
and f − := max{−f, 0} are integrable. In this case
Z Z Z
f dλ := f dλ − f − dλ.
+
I I I

(iii) A function f : I → C is called (Lebesgue) integrable if Re(f ) and Im(f )

are integrable. In this case
Z Z Z
f dλ := Re(f ) dλ + i Im(f ) dλ.
I I I

The Lebesgue integral has the following properties.

Lemma A.17. (i) If f, g are Lebesgue integrable and α ∈ K, then

Z Z Z
(αf + g) dλ = α f dλ + g dλ.
I I I

(ii) If f is Lebesgue integrable then

Z  Z
f dλ ≤ |f | dλ.
 

I I

For Lebesgue integrals much stronger convergence theorems hold than for the
Riemann integral. The most important convergence theorems are the following.

Theorem A.18 (Monotone convergence theorem). Let (fn )n∈N be a se-

quence of measurable functions fn : I → [0, ∞] with 0 ≤ f1 ≤ f2 ≤ · · · . Then
f : I → [0, ∞], f (t) := lim fn (t)
n→∞

is measurable and
Z Z
f dλ = lim fn dλ.
I n→∞ I

The monotone convergence theorem is also called Beppo Levi theorem.

A sequence (fn )n∈N converges to f λ-a.e. if the set, where the sequence does
not converge to f , has measure zero.

Theorem A.19 (Dominated convergence theorem). Let (fn )n∈N be a se-

quence of measurable functions and assume that there exists a measurable func-
tion f such that f (t) = lim fn (t) λ-a.e. If there exists an integrable function
n→∞
g with |fn | ≤ g λ-a.e., then f is integrable and
Z Z
f dλ = lim fn dλ.
I n→∞ I
136 A.3. Lp spaces

The dominated convergence theorem is also called Lebesgue’s convergence the-

orem.

Theorem A.20 (Fatou’s lemma). Let (fn )n∈N be a sequence of integrable

functions and assume that there exists a measurable
R function f such that f (t) =
lim fn (t) λ-a.e. If there exists a C such that I fn dλ ≤ C for all n ∈ N, then
n→∞
f is integrable and
Z Z
f dλ ≤ lim inf fn dλ.
I n→∞ I

A.3 Lp spaces
In the following, Ω is always an open subset of Rn .

Definition A.21. For 1 ≤ p < ∞ we define

n Z o
Lp (Ω) := f : Ω → K : f measurable, |f |p dλ < ∞ ,
Ω
Z  p1
kf kp := |f |p , f ∈ Lp (Ω).
Ω

Definition A.22. For a measurable function f : Ω → K we define the essential

supremum
ess sup f := inf{C ∈ R : f (t) ≤ C for λ-almost all t},
ess inf f := − ess sup(−f )
and
L∞ (Ω) := {f : Ω → K : f measurable, ess sup |f | < ∞},
kf k∞ := ess sup |f |, f ∈ L∞ (Ω).

It is easy to see that L∞ (Ω) is a vector space. For 1 ≤ p < ∞ this follows from
Z Z Z
|f + g| dλ ≤ (|f | + |g|) dλ ≤ (2 max{|f |, |g|})p dλ
p p
Ω Ω Ω
Z Z
p
≤2 max{|f | , |g| } dλ ≤ 2p
p p
|f |p + |g|p dλ
Ω Ω
= 2p (kf kpp + kgkpp ) < ∞.
That λf ∈ Lp for λ ∈ K and f ∈ Lp is clear.
That k · kp is a seminorm on Lp follows from the Minkowski inequality:

Theorem A.23 (Minkowski inequality). Let 1 ≤ p ≤ ∞ and f, g ∈ Lp (Ω).

Then
kf + gkp ≤ kf kp + kgkp .
Chapter A. Lp spaces 137

For the proof of the Minkowski inequality, Hölder’s inequality is used.

1 1
Theorem A.24. Let 1 ≤ p ≤ ∞ and q the conjugated exponent, i. e., p+q = 1.
If f ∈ Lp (Ω) and g ∈ Lq (Ω), then f g ∈ L1 (Ω) and

kf gk1 ≤ kf kp kgkq .

Note that Lp (Ω) is only a seminormed space, because there are non-zero func-
tions f with kf kp = 0.

Theorem A.25. (Lp (Ω), k · kp ) is complete.

Definition A.26. Let Np (Ω) := {f ∈ Lp : kf kp = 0}. Then Lp := Lp (Ω)/N (Ω)

is a complete normed space.

Usually an equivalence class [f ] ∈ Lp (Ω) is simply denoted by f .

Often one is interested in dense subspaces of Lp (Ω).

Theorem A.27. Let 1 ≤ p < ∞ and Ω ∈ Rn open. Then the test functions

C0∞ (Ω) := D(Ω) := {ϕ ∈ C ∞ (Ω) : supp(ϕ) ⊂ Ω is compact}.

form a dense subset of Lp (Ω).

Chapter B. Exercises 139

Appendix B

Exercises

Exercises for Chapter 1

1. Banach’s fixed point theorem. Let M be a metric space. A map f :
M → M is called a contraction if there exists a γ < 1 such that

d(f (x), f (y)) ≤ γ d(x, y), x, y ∈ M.

Show that every contraction f on a complete normed space M has exactly

one fixed point, that is, there exists exactly one x0 ∈ M such that f (x0 ) =
x0 .

2. Let X be a normed space. Then the following is equivalent:

(i)X is complete.
(ii)Every absolutely convergent series in X converges in X.

3. Let X be a normed space. Show:

(a) Every finite-dimensional subspace of X is closed.

(b) If V is a finite-dimensional subspace of X and W is a closed subspace
of X, then

V + W := {v + w : v ∈ V, w ∈ W }

is a closed subspace of X.

4. Let T be a set and `∞ (T ) be the space of all functions x : T → K with

kxk∞ := sup{|x(t)| : t ∈ T } < ∞.

Show that (`∞ (T ), k · k∞ ) is a Banach space.

140

5. Let the sequence spaces d, c0 , c be defined as in Example 1.15.

(a) Show that (c0 , k · k∞ ) and (c, k · k∞ ) are Banach spaces.
(b) Show that (d, k · k∞ ) is a normed space, but that it is not complete.

6. Sea X un espacio normado con dim X ≥ 1 y S, T operadores lineales en X

tales que ST − T S = id. Muestre que al menos uno de estos operadores no
es acotado. Ayuda: Muestre que ST n+1 − T n+1 S = (n + 1)T n .

7. Sean X y Y espacios normados con X de dimensión finita. Muestre que

toda función lineal T : X → Y es acotada.

8. (a) Sea X = C([a, b]) con la norma k · k∞ . Muestre que

Z b
T : X → C, Tx = x(t) dt
a

es un operador lineal y acotado. ¿Cuál es su norma?

(b) Ahora considere X con la norma
Z b 1/p
kxkp := |x(t)|p dt , x ∈ X,
a

para 1 ≤ p < ∞. ¿Sigue siendo T acotado? Si es ası́, calcule su norma.

9. Sea 1 ≤ p < ∞. Para z = (zn )n∈N ∈ `∞ sea T : `p → `p definido por

(T x)n = xn zn para x = (xn )n∈N ∈ `p . Muestre que T ∈ L(`p ) y calcule kT k.

Exercises for Chapter 2

1. Demuestre el teorema de Hahn-Banach para espacios vectoriales complejos.

Sugerencia: Para un espacio vectorial sobre los complejos X muestre que:

(a) Sea ϕ : X → R un funcional R-lineal, entonces

Vϕ : X → C, Vϕ (x) := ϕ(x) − iϕ(ix),

es un funcional C-lineal sobre X con Re Vϕ = ϕ.

(b) Sea λ : X → C un funcional C-lineal con Re λ = ϕ, entonces Vϕ = λ.
(c) Sea p un funcional sublineal sobre X y ϕ, Vϕ definido como en el punto
anterior, entonces

|ϕ(x)| ≤ p(x) ⇐⇒ |Vϕ (x)| ≤ p(x), x ∈ X.

(d) kϕk = kVϕ k.

Chapter B. Exercises 141

2. En X = `2 (N) considere el subespacio

U = {(xn )n∈N : xn = 0 excepto para un número finito de ı́ndices n}.

Sea V el complemento algebraico de U en X, i. e., U es un subespacio tal

que U + V = X y U ∩ V = {0}. Muestre que
∞
X
ϕ : X → K, ϕ(x) = un para x = u + v con u ∈ U, v ∈ V.
n=0

es un funcional lineal bien definido y no acotado.

3. (a) Sea c ⊆ `∞ el conjunto de las sucesiones convergentes. Muestre que el

funcional

ϕ0 : c → K, x = (xn )n∈N 7→ lim xn

n→∞

es continuo y calcule su norma.

(b) Sea `∞ (N, R) el conjunto de todas las sucesiones acotadas en R con la
norma del supremo. Muestre que existe ϕ ∈ (`∞ (N, R))0 tal que

lim inf xn ≤ ϕ(x) ≤ lim sup xn , x = (xn )n∈N ∈ `∞ .

n→∞ n→∞

4. Sea X un espacio normado, f : X → K un funcional lineal no nulo y K =

ker f

(a) Muestre que dim(X/K) = 1.

(b) Muestre que f es continuo si y solo si ker f es cerrado.

5. Un isomorfismo entre espacios normados X y Y es un homeomorfismo lineal.

Pruebe las siguientes afirmaciones.

(a) Si T : X → Y es un isomorfismo [isométrico] entre los espacios norma-

dos X y Y , entonces T 0 : Y 0 → X 0 es un isomorfismo [isométrico]. Si
X y Y son espacios de Banach, el converso también vale.
(b) Si un espacio normado Y es isomorfo a un espacio de Banach reflexivo
X, entonces Y es un espacio de Banach reflexivo.

6. Sea X un espacio normado separable y (x0n )n∈N una sucesión acotada en X 0 .

Entonces existe una subsucesión (x0nk )k∈N y x00 ∈ X 0 tal que

lim x0nk (x) = x00 (x), x ∈ X.

k→∞

Es cierto esto sin la hipótesis de que X sea separable?

142

7. Sea X un espacio normado y M un subespacio de X. Sea

L = {f ∈ X 0 | f (x) = 0 para todo x ∈ M }.
Muestre que L es un subespacio cerrado de X 0 y que M 0 es isométricamente
isomorfo a X 0 /L.

8. Sea X un espacio compacto, CR (X) el conjunto de funciones continuas real-

evaluadas sobre X y Y ⊂ X un subconjunto cerrado.
(a) Considere el mapa ρ : CR (X) → CR (Y ) definido por ρ(f ) = f |Y .
Muestre que I := ker(ρ) es un subespacio cerrado de CR (X).
(b) Sea ρe : CR (X)/I → CR (Y ) el mapa inducido en el espacio cociente.
Pruebe que ρe es una isometrı́a.
(c) Demuestre que rg(ρ) es completo.
(d) Use el teorma de Stone-Weierstraß para concluir el teorema de Tietze:
Sea X un espacio compacto de Hausdorff y Y ⊆ X un subconjunto
cerrado. Entonces cada función continua f : Y → R tiene una extensión
continua fe : X → R con kfekC(X) = kf kC(Y ) .

9. Muestre que en l1 la convergencia débil y la convergencia en norma coinciden.

Exercises for Chapter 3

1. (a) Todo espacio métrico completo con infinitos elementos y ningún punto
aislado es no enumerable.
(b) Toda base algebraica de un espacio de Banach infinito dimensional es
no enumerable.

2. (a) Sea X un espacio de Banach, Y un espacio normado y (Tn )n∈N ⊆

L(X, Y ). Suponga que para todo x ∈ X el lı́mite T x := lim Tn x existe.
n∈N
Entonces T ∈ L(X, Y ).
(b) Sean X, Y espacios de Banach, Y reflexivo, y (Tn )n∈N ⊆ L(X, Y ) tal
que (ϕ(Tn x))n∈N converge para todo x ∈ X y ϕ ∈ Y 0 . Entonces existe
w
un T ∈ L(X, Y ) tal que Tn −
→ T.

3. Muestre que la hipótesis de completitud en el principio de acotación uniforme

es necesaria.
(n) (n) (n)
4. Sea [a, b] ⊆ R, n ∈ N y tome a ≤ t1 < · · · < tn ≤ b y αk ∈ K,
k = 1, . . . , n. Para f ∈ C([a, b]) se define
n
(n) (n)
X
Qn (f ) := αk f (tk ).
k=1

Muestre que los siguientes enunciados son equivalentes:

Chapter B. Exercises 143

Z b
(a) Qn (f ) → f (t) dt, n → ∞, para todo f ∈ C[a, b].
a
Z b
(b) Qn (p) → p(t) dt, n → ∞, para todo polinomio p : [a, b] → K y
Pn a (n)
sup k=1 |αk | < ∞.
n∈N

Sean X, Y, Z espacios de Banach y T : X ⊇ D(T ) → Y un operador lin-

eal.
(a) Sea S : X ⊇ D(S) → Y un operador lineal. Entonces la suma de
operadores S + T se define como

D(S + T ) := D(S) ∩ D(T ), (S + T )x := Sx + T x.

(b) Sea R : Y ⊇ D(R) → Z un operador lineal. Entonces el producto de

operadores o composición RT se define como

D(RT ) := {x ∈ D(T ) : T x ∈ D(R)}, (RT )x := R(T x).

5. Sean X, Y, Z espacios de Banach, R ∈ L(X, Y ), T : X ⊇ D(T ) → Y ,

S : Y ⊇ D(S) → Z operadores lineales cerrados. Muestre que:

(a) R + T es un operador lineal cerrado.

(b) SR es cerrado.
(c) Si S es contı́nuamente invertible (i. e., S −1 : rg(S) → Y existe y es
contı́nuo), entonces ST es cerrado.

Muestre además que estas afirmaciones siguen siendo válidas cambiando

“cerrado” por “clausurable”

6. Sea X = `2 (N) y

T : X ⊇ D(T ) → X, T x = (nxn )n∈N para x = (xn )n∈N .

Diga si T es cerrado con:

(a) D(T ) = {x = (xn )n∈N ∈ `2 (N) : (nxn )n∈N ∈ `2 (N)},
(b) D(T ) = d = {x = (xn )n∈N ∈ `2 (N) : xn 6= 0 para solo finitos n}.

7. Sea X un espacio de Banach, n ∈ N y T un operador lineal densamente

definido de X en K n . Muestre que T es cerrado si y solo si T ∈ L(X, Kn ).

8. Sean X y Y espacios normados y T : X ⊇ D(T ) → Y un operador lineal

cerrado.
144

(a) Sea K ⊂ X compacto. Muestre que T (K) es cerrado en Y .

(b) Muestre que si F es un compacto en Y entonces T −1 (F ) es cerrado en
X.
(c) ¿Si A es cerrado en X, es cierto que T (A) es cerrado?

9. Sea X un espacio normado. Una sucesión (xn )n∈N ⊆ X es una sucesión débil
de Cauchy si para todo ϕ ∈ X 0 la sucesión (ϕ(xn ))n∈N es una sucesión de
Cauchy en K.

(a) Sea x = (xn )n∈N una sucesión acotada en X. Muestre que x es una
sucesión débil de Cauchy si y solo si existe un subconjunto denso U 0 de
X 0 tal que (ϕ(xn ))n∈N es una sucesión de Cauchy para todo ϕ ∈ U 0 .
(b) Toda sucesión débil de Cauchy en X es acotada.

10. Sea X un espacio de Banach, (xn )n∈N ⊆ X, (ϕn )n∈N ⊆ X 0 , y x0 ∈ X,

k·k w∗
ϕ0 ∈ X 0 tal que xn −−→ x0 y ϕn −−→ ϕ0 . Muestre que lim ϕn (xn ) = ϕ0 (x0 ).
n→∞

11. Sea X un espacio normado.

(a) Muestre que (X, k · k)0 = (X, σ(X, X 0 ))0 . Es decir: un funcional lineal
ϕ : X → K es continua con respecto a la topologı́a inducida por k · k si
y sólo si es continua con respecto a la topologı́a débil.
w
(b) Sean (xn )n∈N ⊆ X, x0 ∈ X y (ϕn )n∈N ⊆ X 0 , ϕ0 ∈ X 0 tal que xn −
→ x0
w∗
y ϕn −−→ ϕ0 . Muestre

kx0 k ≤ lim inf kxn k, kϕ0 k ≤ lim inf kϕn k.

n→∞ n→∞

(c) Sean S = {x ∈ X : kxk = 1} la esfera unitaria y K = {x ∈ X : kxk =

1} la bola unitaria cerrada en X. ¿Siempre son débilmente cerradas
(prueba o contraejemplo)?

12. Para n ∈ N sea en = (0, . . . , 1, 0, . . . ) la sucesión que tiene 1 en la posición

n y 0 en el resto.

(a) Muestre que (en )n∈N no es convergente débilmente en `1 .

(b) Muestre que (en )n∈N es w∗ convergente en `1 .

13. Sea X un espacio vectorial y M ⊆ X un subconjunto convexo, balanceado

y absorbente. Muestre que el funcional de Minkowski pM es una seminorma
en X.
Chapter B. Exercises 145

Exercises for Chapter 4

1. Sea X un espacio pre-Hilbert, U ⊆ H un subespacio denso y x0 ∈ X tal que
hx0 , ui = 0 para todo u ∈ U . Muestre que x0 = 0.

2. Sea w ∈ C([0, 1], R). Para x, y ∈ C([0, 1]) se define

Z 1
hx , yiw := x(t)y(t)w(t) dt.
0

Halle una condición necesaria y suficiente spobre w para que h· , ·iw sea
un producto interno. Bajo qué condición la norma inducida por h· , ·iw es
equivalente a la norma usual de L2 ?

3. Let H be Hilbert space, (xn )n∈N ⊆ H and x0 ∈ H. Then the following is

equivalent:
(a) xn → x0 .
w
(b) kxn k → x0 and xn −
→ x0 .

4. Ejemplo de una proyección no acotada. Sea H = l2 y ei el vector usual

eji = δij . Defina
L1 := span{e2n+1 : n ∈ N0 }
y
 
1 1 1
L2 := span e1 + e2 , e3 + 2 e4 , e5 + 3 e6 , . . . .
2 2 2

(a) Muestre que L1 ∩ L2 = {0}.

(b) Muestre que L1 ⊕ L2 = H.
(c) Muestre que L1 ⊕ L2 6= H.
(d) Defina el operador P0 : L1 ⊕ L2 → L1 ⊕ L2 , P0 (x + y) = x. Muestre
que P0 es una proyección no acotada.

5. Para λ ∈ R defina fλ : R → C, fλ (s) = eiλs y sea X = span{fλ : λ ∈ R}.

Muestre que
Z T
1
hf , gi := lim f (s)g(s) ds
T →∞ 2T −T

define un producto interior en X. Muestre que la completación de X no es

separable. (kfλ − fλ0 k =?)
Los elementos en la completación de X se llaman funciones casi periódicas.

6. ¿Existe algún producto interno h· , ·i en C[0, 1] tal que hx, xi = kxk2∞ para
todo x ∈ C[0, 1] ?
146

7. Sea X un espacio pre-Hilbert. Muestre los siguientes resultados

(a) Sean x, y ∈ X con x ⊥ y, entonces

kx + yk2 = kxk2 + kyk2 .

¿El converso es cierto en general? ¿Hay algún caso para el que se tenga?
(b) Si x 6= 0, y 6= 0 y x ⊥ y muestre que el conjunto {x, y} es linealmente
independiente.
¿Como se puede generalizar este resultado?
(c) x ⊥ y, si y solo si kx + αyk ≥ kxk para todo escalar α.

8. Let H be a Hilbert space, Y ⊆ H a subspace and ϕ0 ∈ Y 0 . Show that there

exists exactly one extension ϕ ∈ H 0 of ϕ0 with kϕ0 k = kϕk.

9. Sea X un espacio pre-Hilbert y U ⊆ X un subespacio. Muestre que

(a) U 6= U ⊥⊥ . ¿Se tiene alguna contenencia?
(b) U ⊕ U ⊥ 6= X

10. Sea 1 ≤ p ≤ ∞. Para f ∈ Lp (R) y s ∈ R defina Ts : Lp (R) → Lp (R) como

(Ts f )(t) := f (t − s). Claramente los Ts son isometrias lineales.
s
(a) Sea 1 ≤ p < ∞. Muestre que Ts −
→ id para s → 0. Los Ts convergen
en norma?
(b) Los Ts convergen en norma o convergen fuertemente en el caso p = ∞?

11. Muestre que W m (Ω), H m (Ω) y H0m (Ω) son espacios de Hilbert.
Para el problema 4.10: Para Ω ⊆ R definimos el conjunto de funciones de
prueba

D(Ω) := {ϕ ∈ C ∞ (Ω) : supp(ϕ) ⊆ Ω es compacto}.

Para un multi-ı́ndice α = (α1 , . . . , αn ) ∈ Nn se define |α| = α1 + · · · + αn y

Dα ϕ = ∂1α1 . . . ∂nαn ϕ si la derivada existe.
Sea f ∈ L2 (Ω). Una función g ∈ L2 (Ω) se llama la derivada débil α-ésima
de f si

hg , ϕi = (−)|α| hf , Dα ϕi, ϕ ∈ D(Ω).

Note que la derivada débil es única si existe; se denota por D(α) f .

Para m ∈ N definimos el espacio de Sobolev

W m (Ω) := {f ∈ L2 (Ω) : D(α) f ∈ L2 (Ω), |α| ≤ m}.

Chapter B. Exercises 147

W m (Ω) es un producto interior con

X
hf , giW m := hD(α) f , D(α) gi2 .
|α|≤m

Además, definimos los espacios

H m (Ω) := C m (Ω) ∩ W m (Ω) and H0m (Ω) := D(Ω)

donde la clausura es tomada con respecto a la norma inducida por h· , ·iW m .

p B : H × H → K sesquilineal. En H × H
12. Sea H un espacio de Hilbert and
considere la norma k(x, y)k := kxk2 + kyk2 .

(a) Muestre que las siguientes son equivalentes:

(i) B es contı́nua.
(ii) B is parcialmente contı́nua, es decir, para cada x0 fijo, y 7→ B(x0 , y)
es contı́nua para cada y0 fijo, x 7→ B(x, y0 ) es contı́nua.
(iii) B es acotado, es decir, existe M ∈ R tal que kB(x, y)k ≤ M kxkkyk
para todo x, y ∈ H.
(b) Si B es contı́nuo, entonces existe T ∈ L(H) tal que

B(x, y) = hT x , yi, x, y ∈ H.

(c) Si además existe m > 0 tal que B(x, x) ≥ mkxk2 , x ∈ H, entonces T

es invertible y kT −1 k ≤ m−1 .

13. Sea H un espacio de Hilbert. Muestre que para toda sucesión (xn )n ⊆ H
acotada, existe una subsucesión (xnk )k tal que la sucesión (ym )m donde,
m
1 X
ym = x nk ,
m
k=1

converge.

14. Sea X un espacio normado, (xn )n∈N ⊆ X y x ∈ X. Las siquientes son

equivalentes:
P
(a) n∈N xn converge incondicionalmente a x.
(b) Para todo ε > 0 existe un conjunto finito A ⊆ N tal que para todo
conjunto finito B con A ⊆ B ⊆ N
X
xb − x < ε.


b∈B

15. Sea H un espacio de Hilbert. Si P : H → H es un operador lineal, las

siquientes son equivalentes:
148

(a) P es una proyección ortogonal.

(b) P 2 = P y hP x , yi = hx , P yi.

16. Sea H un espacio de Hilbert, V, W ⊆ H subespacios cerrados y PV , PW sus

correspondientes proyecciones ortogonales.
(a) Muestre que
V ⊆W ⇐⇒ P V = PV PW = PW PV .

(b) Muestre que las siguientes afirmaciones son equivalentes:

(i) PV PW = 0.
(ii) V ⊥ W .
(iii) PV + PW es una proyección ortogonal.
Muestre que rg(PV + PW ) = V ⊕ W si alguna de las condiciones anteriores
se tiene.

17. Sea H un espacio de Hilbert y P0 , P1 las proyecciones ortogonales sobre

H0 , H1 ⊆ H. Entonces las siguientes afirmaciones son equivalenteas:
(a) H0 ⊆ H1 ,
(b) kP0 xk ≤ kP1 xk, x ∈ H.
(c) hP0 x , xi ≤ hP1 x , xi, x ∈ H.
(d) P0 P1 = P0 .

18. Sea H un espacio de Hilbert separable, (xn )n∈N una base ortonormal de H,
y, (yn )n∈N una sucesión tal que:
∞
X
||xn − yn || < 1
n=1

y z ⊥ yn , para todo n ∈ N, entonces z = 0.

19. Sea H un espacio de Hilbert complejo y T : H → H un operador lineal

acotado. Muestre que T es normal si y solo si ||T ∗ x|| = ||T x|| para todo
x ∈ H. En este caso, muestre que ||T 2 || = ||T ||2 .

20. Haar functions. Let ψ = χ[0, 1/2) − χ[1/2, 1) . For n, k ∈ Z define

ψn,k : R → R, ψn,k (t) = 2k/2 ψ(2k t − n).

For k ∈ N0 and n ∈ {0, 1, 2, . . . , 2k − 1} let


h2k +n (t) = ψk,n (t),
 for t ∈ [0, 1),
h2k +n : [0, 1] → R,
h2k +n (1) = lim ψk,n (t).

t→1−

and h0 (t) = 1, t ∈ [0, 1].

Chapter B. Exercises 149

(a) (hj )j∈N0 is a orthonormal system in L2 [0, 1] and (ψn,k )n,k∈Z is a or-
thonormal system in L2 (R).
P2k −1
(b) T : L2 [0, 1] → L2 [0, 1], T f = j=0 hf , hj ihj is a orthonormal projec-
tion on the subspace

U = {f ∈ L2 [0, 1] : f const. in intervals [r2−k , (r + 1)2−k ) with r ∈ N0 }.

P∞
(c) For f ∈ C[0, 1], the series j=0 hf , hj ihj converges uniformely to f .
(d) (hj )j∈N0 is an orthonormal basis of L2 [0, 1].
(e) (ψk,n )k,n∈Z is an orthonormal basis of L2 (R).

21. Sea H un espacio de Hilbert, V, W ⊆ H subespacios cerrados y PV , PW sus

correspondientes proyecciones ortogonales.
(a) Muestre que

V ⊆W ⇐⇒ PV = PV PW = PW PV .

(b) Muestre que las siguientes afirmaciones son equivalentes:

(i) PV PW = 0.
(ii) V ⊥ W .
(iii) PV + PW es una proyección ortogonal.
Muestre que rg(PV + PW ) = V ⊕ W si alguna de las condiciones anteriores
se tiene.

22. Sea H un espacio de Hilbert y P0 , P1 las proyecciones ortogonales sobre

H0 , H1 ⊆ H. Entonces las siguientes afirmaciones son equivalenteas:
(i)H0 ⊆ H1 ,
(ii)kP0 xk ≤ kP1 xk, x ∈ H.
(iii)hP0 x , xi ≤ hP1 x , xi, x ∈ H.
(iv)P0 P1 = P0 .

23. Sea H un espacio de Hilbert separable, (xn )n∈N una base ortonormal de H,
y, (yn )n∈N una sucesión tal que:
∞
X
||xn − yn || < 1
n=1

y z ⊥ yn , para todo n ∈ N, entonces z = 0.

24. Sea H un espacio de Hilbert complejo y T : H → H un operador lineal

acotado. Muestre que T es normal si y solo si ||T ∗ x|| = ||T x|| para todo
x ∈ H. En este caso, muestre que ||T 2 || = ||T ||2 .
150

25. Sea H un espacio de Hilbert y (Tn )n∈N una sucesión acotada y monótonamente
creciente de operadores autoadjuntos. Muestre que la sucesión converge en
el sentido fuerte a un operador autoadjunto.

26. Sea (Pn )n∈N una sucesión monótona de proyecciones ortogonales en un es-
pacio de Hilbert H. Muestre que (Pn )n∈N converge en el sentido fuerte a
una proyección ortogonal P y además
S
(a) rg P = n∈N rgPn si Pn es creciente.
T
(b) rg P = n∈N rgPn si Pn es decreciente.

27. Sean H1 , H2 y H3 espacios de Hilbert y S(H1 → H2 ) y T (H2 → H3 )

operadores lineales densamente definidos.

(a) Si T ∈ L(H2 , H3 ) entonces T S es densamente definido y (T S)∗ = S ∗ T ∗ .

(b) Si S es inyectivo y S −1 ∈ L(H2 , H1 ) entonces T S es densamente

definido y (T S)∗ = S ∗ T ∗ .
(c) Si S es inyectivo y S −1 ∈ L(H2 , H1 ) entonces S ∗ es inyectivo y (R∗ )−1 =
(R−1 )∗

28. Sean H1 , H2 espacios de Hilbert y U : H1 × H2 → H2 × H1 , U (x, y) =

(−y, x). Entonces
(a) U es unitario.
(b) Si T (H1 → H2 ) es densamente definido,
G(T ∗ ) = [U (G(T ))]⊥ = U (G(T )⊥ ).

(c) T ∗ es cerrado.
(d) Si T es clausurable, T ∗ es densamente definido y T ∗∗ = T .

Exercises for Chapter 5

1. (a) Sea X = C([0, 1]) y a ∈ C([0, 1]). Muestre que
A : X → X, (Ax)(t) = a(t)x(t)
es un operador lineal acotado. Encuentre kAk, σ(A), σp (A), σc (A) y
σr (A).
(b) Sea H = {f ∈ C([0, 1]) : x(0) = 0} y
Z t
S : H → H, (Sx)(t) = x(s) ds.
0

Encuentre σ(S), σp (S), σc (S) y σr (S).

Chapter B. Exercises 151

2. Sea (λn )n∈N ⊆ C una sucesión acotada, y,

T : `1 → `1 , T ((xn )n∈N ) = (λn xn )n∈N .

Encuentre σ(T ), σp (T ), σc (T ) y σr (T ). Muestre además que, para todo

K ⊆ C compacto no vacı́o, existe un operador T ∈ L(`1 ) cuyo espectro es
K.

3. Sea X un espacio de Banach S, T ∈ L(X). Muestre que σ(ST ) \ {0} =

σ(T S) \ {0}.
Hint. Muestre que id −ST es invertible si y solo si id −T S es invertible, en-
contrando una relación entre (id −T S)−1 y (id −ST )−1 . Suponga ||T || ||S|| <
1 y mire si la relación en este caso es válida en general.

4. Encuentre el espectro puntual, el espectro continuo y el espectro residual de

los operadores:

R : `2 (N) → `2 (N), R(x1 , x2 , x3 , . . . ) = (0, x1 , x2 , x3 , . . . ),

L : `2 (N) → `2 (N), L(x1 , x2 , x3 , . . . ) = (x2 , x3 , x4 , . . . ),
T : `∞ (N) → `∞ (N), T (x1 , x2 , x3 , . . . ) = (x2 , x3 , x4 , . . . ).
Bibliography

[Den04] Robert Denk. Funktionalanalysis I. Universität Konstanz, http://

www.mathematik.uni-regensburg.de/broecker/index.html, 2004.

[DS88] Nelson Dunford and Jacob T. Schwartz. Linear operators. Part I. Wiley
Classics Library. John Wiley & Sons Inc., New York, 1988. General
theory, With the assistance of William G. Bade and Robert G. Bartle,
Reprint of the 1958 original, A Wiley-Interscience Publication.

[Hal98] Paul R. Halmos. Introduction to Hilbert space and the theory of spectral
multiplicity. AMS Chelsea Publishing, Providence, RI, 1998. Reprint
of the second (1957) edition.

[Heu06] Harro Heuser. Funktionalanalysis. Mathematische Leitfäden. [Math-

ematical Textbooks]. B. G. Teubner, Stuttgart, fourth edition, 2006.
Theorie und Anwendung. [Theory and application].

[Kat95] Tosio Kato. Perturbation theory for linear operators. Classics in Math-
ematics. Springer-Verlag, Berlin, 1995. Reprint of the 1980 edition.

[Les] Jaime Lesmes. Análisis funcional. 2 edición.

[Mac09] Barbara D. MacCluer. Elementary functional analysis, volume 253 of

Graduate Texts in Mathematics. Springer, New York, 2009.

[RS80] Michael Reed and Barry Simon. Methods of modern mathematical

physics. I. Academic Press Inc. [Harcourt Brace Jovanovich Publish-
ers], New York, second edition, 1980. Functional analysis.

[Rud87] Walter Rudin. Real and complex analysis. McGraw-Hill Book Co., New
York, third edition, 1987.

[Rud91] Walter Rudin. Functional analysis. International Series in Pure and

Applied Mathematics. McGraw-Hill Inc., New York, second edition,
1991.

[Wei80] Joachim Weidmann. Linear operators in Hilbert spaces, volume 68

of Graduate Texts in Mathematics. Springer-Verlag, New York, 1980.
Translated from the German by Joseph Szücs.

153
154 Bibliography

[Wer00] Dirk Werner. Funktionalanalysis. Springer-Verlag, Berlin, extended

edition, 2000.
[Yos95] Kōsaku Yosida. Functional analysis. Classics in Mathematics. Springer-
Verlag, Berlin, 1995. Reprint of the sixth (1980) edition.

[Zei95] Eberhard Zeidler. Applied functional analysis, volume 109 of Applied

Mathematical Sciences. Springer-Verlag, New York, 1995. Main prin-
ciples and their applications.