F UNCTIONAL ANALYSIS

Georg-August University Göttingen

Summer Semester 2019
Prof. Dr. Max Wardetzky

Blackboard notes by:

Andreas Wagenblast

Version:

June 11, 2019

Disclaimer:

WARNING! This is no official script. These notes were taken live during the lectures without
much post-editing. Therefore it is just natural that there will be typos and errors. So use these
notes with care. Corrections would be appreciated. So feel free to contact me.
k: [email protected]
1 CONTENTS

Contents
1 Metric spaces 2

2 Complete metric spaces 4

3 Mapping spaces as metric spaces 7

4 Compactness 8

5 Topological vector spaces 11

6 Quotient spaces 14

7 Theorems of Hahn-Banach 15

8 The dual space 19

9 Principle of uniform boundness 21

10 Open mapping theorem & closed graph theorem 23

11 Hilbert spaces (Part I: basics) 27

12 Orthonormal bases in Hilbert spaces 33

13 Hermitian (self-adjoint) operators 39

14 L p -spaces 42
14.1 Integration of step functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
14.2 Summable functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
14.3 Applications of monotone convergence (Beppo-Levi) . . . . . . . . . . . . . . . . 48
14.4 Measurable and summable sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2 1 METRIC SPACES

1 Metric spaces
Definition 1.1. A pair (X , d ), where X is a set and d : X × X → R+ is called metric space, if

(i) d (x, y) = 0 ⇔ x = y (definiteness).

(ii) d (x, y) = d (y, x), for all x, y ∈ X (symmetry)

(iii) d (x, z) ≤ d (x, y) + d (y, z), for all x, y, z ∈ X (triangle-inequality)

Example 1.2. (A) Rn with d `2 (x, y) =

pP
i |x i − y i |2 (euclidean metric)

(B) Rn with d `∞ (x, y) = supi |x i − y i |

(C) Q with d (x, y) = |x − y|

(D) Z with d (x, y) = |x − y|

(E) A graph(V, E ) where V =vertices and E =edges s.t.

• between two vertices ∃ at most one edge
• no edge starting and ending at the same vertex
• every vertex is accessible
Then d (x, y) = minimal number of edges connecting x and y

(F) . . .

Lemma 1.3 (−inequality). For all a, a 0 , b, b 0 the inequality

|d (a, a 0 ) − d (b, b 0 )| ≤ d (a, b) + d (a 0 , b 0 )

holds.

Proof.

d (a, a 0 ) ≤ d (a, b) + d (b, a 0 )

≤ d (a, b) + d (b, b 0 ) + d (a 0 , b 0 )

⇒ d (a, a 0 ) − d (b, b 0 ) ≤ d (a, b) + d (a 0 , b 0 )

Likewise: d (b, b 0 ) − d (a, a 0 ) ≤ d (a, b) + d (a 0 , b 0 )
3 1 METRIC SPACES

Definition 1.4. Let x ∈ X , ε > 0, ε ∈ R. The ε−neighborhood of x is defined by

Uε (x) := {y ∈ X | d (x, y < ε)}.

Lemma 1.5. (i) x ∈ Uε (x)

(ii) Uε1 (x) ∩Uε2 (x) = Uε (x), where ε = min(ε1 , ε2 )

(iii) y ∈ Uε (x), then Uε−d (x,y) (y) ⊆ Uε (x)

Proof. (i ) and (i i ) are trivial. For (i i i ), let z ∈ Uε−d (x,y) (y). Then d (z, y) ≤ ε − d (x, y)
⇒ d (x, z) ≤ d (x, y) + d (y, z) < ε − d (x, y) + d (x, y)
| {z }
=ε
⇒ z ∈ Uε (x)

Definition 1.6. U ⊆ X is called open, iff ∀x ∈ U ∃ε > 0 s.t. Uε (x) ⊆ U .

Remark 1.7. By (i i i ) of the previous lemma, all the Uε (x) are open.

Recall: A topological space is Hausdorff, if ∀x, y ∈ X with x 6= y ∃ open U ,V ⊆ X s.t. x ∈

U , y ∈ V and U ∩ V = ;

Theorem 1.8. Let (X , d ) be a metric space. Then the set of open sets provides a topology for X
which is Hausdorff.

Proof. We need to show:

(i) union of open sets are open

(ii) intersection of two open sets are open

(iii) Hausdorff property

Well (i ) and (i i ) are obvious. For (i i i ) let x, y ∈ X with x 6= y. Let ε := d (x, y)/2. Then Uε (x) ∩
Uε (y) = ;. Suppose not, then ∃z ∈ Uε (x) ∩Uε (y).
⇒ d (x, y) ≤ d (x, z) + d (z, y) < 2ε = d (x, y)

Definition 1.9. Let (X , d X ) and (Y , d Y ) be two metric spaces. A map f : X → Y is called an

isometric embedding if
d Y ( f (x 1 ), f (x 2 )) = d X (x 1 , x 2 ).

We call X and Y isometric if f is surjective.

4 2 COMPLETE METRIC SPACES

Remark 1.10. Any isometric embedding is injective.

Recall: If X and Y are topological spaces, then f : X → Y is called continuous if f −1 (V ) is

open in X for all open V ⊆ Y .

Definition 1.11. Let (X , d X ) and (Y , d Y ) be metric spaces. Let f : X → Y . Then f is continuous

at x ∈ X if: ∀ε > 0 ∃δ > 0 s.t. f (Uδ (x)) ⊆ Uε ( f (x)).

B δ may depend on ε and on x.

Theorem 1.12. f : (X , d X ) → (Y , d Y ) is continuous (in the topological sense) iff f is continuous
at every x ∈ X .

Proof. ⇒ Let x ∈ X and fix ε > 0. Then f −1 (Uε ( f (x))) is open in X . But x ∈ f −1 (Uε ( f (x))).
Hence ∃δ > 0 s.t. Uδ (x) ⊆ f −1 (Uε ( f (x))).
⇐ Let V ⊆ Y be open. We have to show that f −1 (V ) is open in X . Pick x ∈ f −1 (V ). Let ε > 0
s.t. Uε ( f (x)) ⊆ V . Then by assumption ∃δ > 0 s.t. f (Uδ (x)) ⊆ Uε ( f (x)).
⇒ Uδ (x) ⊆ f −1 (Uε ( f (x))) ⊆ f −1 (V )
⇒ f −1 (V ) is open.

Definition 1.13. A map f : (X , d X ) → (Y , d Y ) is called uniformly continuous is ∀ε > 0 ∃δ > 0

s.t. d X (x 1 , x 2 ) < δ ⇒ d Y ( f (x 1 ), f (x 2 )) < ε ∀x 1 , x 2 ∈ X .

Remark 1.14. Let (X , d X ) and (Y , d Y ) be two metric spaces. Then (X × Y , d X ×Y ) becomes a

metric space via
d X ×Y ((x 1 , y 1 ), (x 2 , y 2 )) := d X (x 1 , x 2 ) + d Y (y 1 , y 2 ).

Exercise 1.15. Show that d X : X × X → R+ is uniformly continuous.

2 Complete metric spaces

Definition 2.1. Let (X , d ) be a metric space. Then a sequence (x n )n∈N with x n ∈ X converges
to x ∈ X if ∀ε > 0 ∃N ∈ N s.t. x n ∈ Uε (x) ∀n ≤ N . The element x is called the limit of (x n )n∈N in
n→∞
this case. We usually write lim x n = x or simply x n −−−−→ x.
n→∞

Lemma 2.2. Limits are unique.

Proof. Suppose not. Then ∃x 6= y s.t. x n → x and x n → y. Let ε := 12 d (x, y > 0). Then ∃N1 ∈ N
s.t. x n ∈ Uε (x) ∀n ≥ N and ∃n 2 ∈ N s.t. x n ∈ Uε (y) ∀n ≥ N2 .
⇒ x n ∈ Uε (x) ∩Uε (y) ∀n ≥ max{N1 , n 2 } .
5 2 COMPLETE METRIC SPACES

Definition 2.3. A set A ⊆ X is called closed iff X \ A is open.

Definition 2.4. Let A ⊆ X be any set. The closure of A is

\
A := B.
B ⊆X closed
A⊆B

Remark 2.5. A is closed.

Lemma 2.6. Let (X , d ) be a metric space. Let A ⊆ X be a subset. Then

A = Ã := {x ∈ X | ∃ sequence (a n )n∈N with a n ∈ A s.t. a n → x}.

Proof. A ⊆ à It suffies to show that à is closed and containes A. Clearly A ⊆ à by constant

sequences. We now show: X \ Ã is open. Let x ∈ X \ Ã. Then ∃ a sequence (a n )n∈N with a n ∈ A
n→∞
s.t. a n −−−−→ x. but then x ∈ Ã . Therefore ∃ε > 0 s.t. Uε (x) ∩ Ã = ;.
⇒ Uε (x) ⊆ X \ Ã, hence it is open.
à ⊆ A Let x ∈ Ã. It suffies to show that x ∈ B ∀ closed sets B that contain A. Suppose not.
Then x ∈ X \ B (open). ⇒ ∃ε > 0 s.t. Uε (x) ⊆ X \ B .
⇒ Uε (x) ∩ B = ;, i.e. Uε (x) ∩ A = ; to the fact that x ∈ Ã, i.e. x is a limit point.

Definition 2.7. A sequence is (x n )n∈N in a metric space (X , d ) us a Cauchy Sequence, if

∀ε > 0 ∃N ∈ N s.t. ∀n, m ≥ N : d (x m , x m ) < ε.

n→∞
Remark 2.8. Suppose x n −−−−→ x ∈ X , then (x n )n∈N is Cauchy (because d ( x n , x m ) ≤ d (x, x n ) +
d (x, x m )).

Definition 2.9. A metric space (X , d ) is called complete if every Cauchy-sequence in X con-

verges in X .

Example 2.10. 1. (Q, | · |) is not complete.

2. (Rd \{0}, d `2 ) is not complete.

Lemma 2.11. Let (X , d ) be a complete metric space. Then a subspace A ⊆ X is closed iff A is
complete as a metric space.

Proof. ⇒ A is closed. Let (a n )n∈N be a Cauchy-sequence in A. Since X is complete, ∃x ∈ X

s.t. lim a n = x.
n→∞
6 2 COMPLETE METRIC SPACES

⇐ A is complete. To show: A = A. Let x ∈ A. Then ∃ sequence (a n )n∈N in A s.t. lim a n = x.

n→∞
But (a n )n∈N is a Cauchy-sequence. Since A is complete, x ∈ A.

Theorem 2.12. Let (X , d ) be a metric space. Then ∃ a complete metric space ( X̂ , dˆ) and an
isometric embedding i : X → X̂ s.t. i (X ) = X̂ . The space ( X̂ , dˆ) is unique up to isometries. The
space ( X̂ , dˆ) is called a completion of X .

Proof. (sketch) Suppose ∃ another metric space ( X̃ , d˜) and an isometric embedding j : X → X̃
s.t. j (X ) = X̃ . To show: ∃ isometry h : X̂ → X̃ s.t. the following diagram is commutative

X̂
i

X h

j
X̃

Let x̂ ∈ X̂ . We need to define h(x̂) ∈ X̃ . Since i (X ) = X̂ ⇒ ∃ a sequence (x n )n∈N in X s.t.

n→∞
i (x n ) −−−−→ x̂. Since i is an isometric embedding, (x n )n∈N is a Cauchy-sequence.
¡¡ ¢ ¢ n→∞
⇒ j (x n ) n∈N is Cauchy-sequence in X̃ . Since X̃ is complete, ∃!x̃ ∈ X̃ s.t. j (x n ) −−−−→ x̃. Now
set h(x̂) := x̃ := lim j (x n ).
n→∞

Exercise 2.13. Show that h is well-defined.

Now consider ŷ = lim i (y n ), (y n )n∈N is Cauchy-sequence in X . Then (since i , j are iso-

n→∞
metric embeddings):

dˆ(x̂, ŷ) = lim d X (x n , y n ) = d˜(x̃, ỹ) = d˜(h(x̂), h( ŷ))

n∈N

⇒ h is an isometric embedding. Easy: h is surjective. Hence h : X̂ → X̃ is an isometry.

Existence: Let C S(X ) be the set of all Cauchy-sequences in X . Let (x n )n∈N , (y n )n∈N ∈ C S(X ).
We say that (x n )n∈N ∼ (y n )n∈N if lim d (x n , y n ) = 0. Define: X̂ := C S(X )/ ∼. Require dˆ on X̂ .
n→∞
to this end, let (x n )n∈N , (y n )n∈N ∈ C S(X ) be two representatives of the classes x̂ and ŷ in X̂ ,
respectively. Define: dˆ(x̂, ŷ) := lim d X (x n , y n ).
n→∞

Exercise 2.14. To show:

1. (d X (x n , y n ))n∈N is convergent in R.

2. dˆ(x̂, ŷ) is well-defined (i.e., independent of the choise of representatives).

7 3 MAPPING SPACES AS METRIC SPACES

3. dˆ is indeed a metric.

In order to finish the proof, define i : X → X̂ by i (x) = [(x, x, . . .)] ∈ X̂ . By construction,

i : X → X̂ is an isometric embedding.

Exercise 2.15. Show:

1. i (X ) = X̂

2. ( X̂ , dˆ) is indeed complete.

This completes the proof.

3 Mapping spaces as metric spaces

Let E be some set, and let (X , d ) be a metric space. Let F (E , X ) := { f : E → X }. Goal: Equip
F (E , X ) with a metric.

¡ ¢
Let f , g ∈ F (E , X ). Define: D( f , g ) := supt ∈E f (t ), g (t )

Exercise 3.1. D is a metric on F (E , X ) is we allow D to take on value +∞. (Note: This is called
∞-metric).

Remark 3.2. i : X → F (E , X ) where x 7→ f x with f x (t ) ≡ x (a constant map), is an isometric

embedding.

Lemma 3.3. The image i (X ) is closed in F (E , X ).

Proof. We show that F (E , X ) \ i (X ) is open. Let f ∈ F (E , X ) \ i (X ). Then ∃t 6= t 0 ∈ E s.t. f (t ) 6=

f (t 0 ) ∈ X . Let ε := 21 d X ( f (t ), f (t 0 )) > 0.
Claim: Uε ( f ) ∩ i (X ) = ;. Suppose not. Then ∃x 0 ∈ X s.t. i (x 0 ) ∈ Uε ( f ). Then

d (i (x 0 )(t ), f (t )) < ε ⇔ d (x 0 , f (t )) < ε,

and
d (i (x 0 )(t 0 ), f (t 0 )) < ε ⇔ d (x 0 , f (t 0 )) < ε.

Hence 2ε = ( f (t ), f (t 0 )) ≤ d (x 0 , f (t )) + d (x 0 , f (t 0 )) < 2ε

Corollary 3.4. If F (E , X ) is complete, then X is complete.

8 4 COMPACTNESS

Proof. Since i (X ) is closed, i (X ) is complete. Since i is an isometric embedding, X is com-

plete.

The converse also holds.

Lemma 3.5. Let X be a complete metric space. Then F (E , X ) is complete.

n→∞
Proof. Let ( f n )n∈N be a Cauchy-sequence in F (E , X ). To show: ∃ f ∈ F (E , X ) s.t. f n −−−−→ f . By
definition of D (sup-metric): Fix ε > 0, then d ( f n (t ), f m (t )) < ε ∀t ∈ E ∀m, n > N . Therefore
∀t ∈ E , ( f n (t ))n∈N is a Cauchy-sequence in X . Since X is complete, ∃ limit x = lim f n (t ).
n→∞
Define: f (t ) := lim f n (t ). Hence f ∈ F (E , X ). Now d ( f n (t ), f m (t )) < ε ∀t ∈ E ∀m, n ≥ N . So,
n→∞
d ( f (t ), f n (t )) ≤ ε ∀n ≥ N .
⇒ D( f , f n ) ≤ ε ∀n ≥ N .
n→∞
⇒ f n −−−−→ f in (F (E , X ), D).

Let now E be a topological space. Let C (E , X ) := { f : E → X | f continuous}.

Definition 3.6. We call f ∈ F (E , X ) continuous in t ∈ E if ∀ε > 0 ∃ open U ⊆ E s.t.

(i) t ∈ U

(ii) f (U ) ⊆ Uε ( f (t ))

We say f ∈ C (E , X ) if f is continuous ∀t ∈ E .

Remark 3.7. f ∈ C (E , X ) ⇔ f −1 (V ) is open in E ∀ open V ⊆ X .

Theorem 3.8. C (E , X ) is closed in F (E , X ).

Proof. Let f ∈ F (E , X ) \ C (E , X ). i.e. ∃t ∈ E and ∃ε > 0 s.t. ∀U ⊆ E open with t ∈ U and ∃t 0 ∈ U

s.t. d ( f (t ), f (t 0 )) ≥ ε.
Claim: Uε/3 ( f ) ∩C (E , X ) = ;. Let g ∈ Uε/3 ( f ) ⇒ d (g (t ), g (t 0 )) ≥ ε/3.
⇒ ∃t ∈ E and ε > 0 s.t. ∀U ⊆ E open with t ∈ U we have d (g (t ), g (t 0 )) ≥ ε/3 for some t 0 ∈ U .
⇒ g is not continuous in t .
⇒ Uε/3 ( f ) ∩C (E , X ) = ;.

4 Compactness
Definition 4.1. A metric space (X , d ) is called
9 4 COMPACTNESS

(a) compact if for all coverings {U j } j ∈ J , where U j ⊆ X open, there exists a finite sub-covering
X = U j 1 ∪U j 2 ∪ . . . ∪U j N , for N ∈ N

(b) sequentially compact if every sequence (x n )n∈N contains a convergent sub-sequence, i.e.
k→∞
there exists a sub-sequence (x nk )k∈N s.t. x nk −−−−→ x ∈ X .
N
(c) totally bounded if ∀ε > 0 ∃N ∈ N and points x 1 , x 2 , . . . , x N in X s.t.
S
Uε (x i ) = X .
i =1

Theorem 4.2. In every metric space (X , d ) the following conditions are equivalent:

(i) (X , d ) is compact.

(ii) (X , d ) is sequentially compact.

(iii) (X , d ) is totally bounded and complete.

Proof. We will proof it in the following order: (i ) ⇒ (i i ) ⇒ (i i i ) ⇒ (i ).

(i ) ⇒ (i i ) Suppose that (X , d ) were not sequentially compact. Then ∃ sequence (x n )n∈N that
does not have a convergent sub-sequence. So ∀x ∈ X ∃ε(x) > 0 s.t. Uε(x) (x) contains only
S ¡ ¢
finitely many members of (x n )n∈N . We have X = Uε(x) (x), i.e. Uε(x) (x) x∈X is a open cov-
x∈X
ering. By assumption there is a finite sub-covering Uε(y 1 ) (y 1 ) ∪ . . . ∪Uε(y N ) (y N ) = X . But every
Uε(y i ) (y i ) contains only finitely many members of (x n )n∈N .
(i i ) ⇒ (i i i ) Completeness: Let (x n )n∈N be a Cauchy-sequence. By assumption (i i ) ∃ conver-
n→∞ n→∞
gent subsequence (x kk )k∈N , i.e. x nk −−−−→ x ∈ X . Hence x n −−−−→ x.
N
Totally boundness: Suppose ∃ε > 0 s.t. ∀N ∈ N f or al l (x 1 , . . . , x N ) we have
S
Uε (x i ) á X .
i =1
2
S
This means ∃x 1 ∈ X s.t. Uε (x 1 ) á X . But then ∃x 2 ∈ X \ Uε (x 1 ) s.t. Uε (x i ) á X . Then
i =1
3
S
∃x 3 ∈ X \ (Uε (x 1 ) ∪Uε (x 2 )) s.t. Uε (x i ) á X . And so on.
i =1
In this way, we obtain a sequence (x n )n∈N where d (x i , x j ) > ε ∀i 6= j . This contradicts the ex-
istence of a convergent sub-sequence .
S
(i i i ) ⇒ (i ) Suppose ∃ covering (U j ) j ∈J with open U j ’s and U j = x, that does not contain a
j ∈J
n
S
finite sub-covering. By assumption ∃ finite covering U1/2 (x i ) = X .
i =1
W.l.o.g.: We have that Uε (x 1 ) cannot be covered by finitely many elements U j . Again by as-
sumption U1/2 (x 1 ) can be covered by finitely many balls of radius 14 . Hence ∃x 2 ∈ U1/2 (x 1 ) s.t.
U1/4 (x 2 ) cannot be covered by finitely many U j . Then ∃x 3 ∈ U1/4 (x 2 ) s.t. U1/8 (x 3 ) cannot be
covered by finitely many U j , and so on . . .
10 4 COMPACTNESS

This gives rise to a sequence (x n )n∈N s.t. x n ∈ U1/2n−1 (x n−1 ). Hence (x n )n∈N is a Cauchy-
n→∞
sequence. By completeness ∃x ∈ X s.t. x n −−−−→ x. But ∃ j ∈ J s.t. x ∈ U j . Therefore U1/2n (x n ) ⊆
U j for n large enough .

Definition 4.3. (i) A subset Y in a metric space (X , d ) is called dense if Y = X .

(ii) (X , d ) is called separable if ∃ countable dense subsets.

Theorem 4.4. Let (X , d ) be totally bounded, then (X , d ) is separable.

Proof. By assumption we have that ∀n ∈ N ∃ finite subset C n s.t.

S
U1/n (x i ) = X . Let C :=
S x i ∈C n
C n . Then C is countable and dense.
n∈N

Definition 4.5. Let (X , d ) be a metric space. A subset A ⊆ X is called relatively compact if A

is compact.

Lemma 4.6. (i) A is totally bounded ⇔ A is totally bounded.

(ii) A is relatively compact ⇔ A is totally bounded and A is complete.

n
S
Proof. (i ): ⇒ ∀ε > 0 ∃a 1 , . . . , a N ∈ A s.t. Uε/2 (a i ) = A. Therefore
i =1

N
[ N
[ N
[
A= Uε/2 (a i ) ⊆ Uε/2 (a i ) ⊆ Uε/2 (a i ).
i =1 i =1 i =1

Hence A is totally bounded.

N
S
⇐ Let A be totally bounded. So ∀ε > 0 ∃b 1 , . . . , b N with b i ∈ A s.t. A ⊆ Uε/2 (b 1 ). pick
i =1
N
S
c i ∈ A ∩Uε/2 (b i ). Then A ⊆ Uε (c i ). This means that A is totally bounded.
i =1

(i i ) ⇒ A relatively compact ⇒ A is compact. So by Theorem ??? A is totally bounded and

complete. By (i ) A is totally bounded (and A is complete).
⇐ By (i ) A is totally bounded (and complete). So A is compact. Hence A is relatively com-
pact.
11 5 TOPOLOGICAL VECTOR SPACES

5 Topological vector spaces

Let K ∈ {R, C}.

Definition 5.1. Let X be a K −vector space that is also a topological space. Then X is called a
topological vector space if

X ×X → X K ×X → X
and
(x, y) 7→ x + y (a, x) 7→ a · x

are continuous.

Here a subset W ⊆ X × Y is open iff ∀(x, y) ∈ W ∃ open sets U ,V in X s.t. x ∈ U and y ∈ V

and U × V ⊆ W . Likewise for K ×X .
Then for addition to be continuous means that ∀(x 0 , y 0 ) ∈ X × X and ∀ open sets V s.t.
(x 0 + y 0 ) ∈ V we find open sets U ,U 0 s.t. x 0 ∈ U and y 0U 0 and U +U 0 ⊆ V .
Here
U +U 0 := }x + y | x ∈ U , y ∈ U 0 ⊆ X .
© ª

Lemma 5.2. Let X be a topological vector space. Let L ⊆ X be a linear subspace. Then L is also
a linear subspace.

Proof. Let x 0 , y 0 ∈ L. We show: x 0 + y 0 ∈ L. It suffies to show that for any open neighborhood
V of (x 0 + y 0 ) we have V ∩ L 6= ;. Let U ,U 0 be open neighborhoods of x 0 and y 0 respectively,
s.t. U +U 0 ⊆ V (by continuity). Let x ∈ U ∩ L and y ∈ U 0 ∩ L. Then x + y ∈ (U +U 0 ) ∩ L ⊆ V ∩ L.
Likewise let a ∈ K and x ∈ L. Then we show ax ∈ L. Consider M a : X → X , y 7→ a y. Then
M a is a homeomorphism. Therefore L = M a (L) = M a (L), because images of closed sets under
homeomorphisms are closed. Hence ax ∈ L.

Lemma 5.3. Let X , Y be two topological K-vector spaces. Let T : X → Y be a linear map be-
tween them. Then the following are equivalent:

(i) T is continuous in 0 ∈ X .

(ii) T is continuous in x ∈ X ∀x ∈ X .

(iii) T is continuous.

Proof. Clearly the implication (i i ) ⇒ (i ) is trivial and the equivalence (i i ) ⇔ (i i i ) is obvious.

So whats left to show is
12 5 TOPOLOGICAL VECTOR SPACES

(i ) ⇒ (i i ) Let x 0 ∈ X . We need to show taht T is continuous in x 0 . Let L y : X → X , z 7→ z + x.

So, L y is a homeomorphism. Consider the following commutative diagram:

T
X Y
L (−x0 ) L T (x0 )
T
X Y

Since L (−x0 ) and L (T (x0 )) are homeomorphisms and T is continuous in 0 ∈ X , the map T is
continuous in x 0 .

Definition 5.4. Let X , Y be topological K-vector spaces. Then L(X , Y ) is the set of linear
continuous maps. If Y = K, then L 0 = L(X , K) is called the dual space of X .

Goal: We want to understand L(X , Y ) as a topological vector space.

Definition 5.5. Let X be a K-vector space. Then k · k : X → R is called a norm if:

(i) kxk ≥ 0 and kxk = 0 ⇔ x = 0 ∈ X .

(ii) kaxk = |a|kxk ∀a ∈ K, ∀x ∈ X .

(iii) kx + yk ≤ kxk + kyk (triangle inequality ∀x, y ∈ X ).

Lemma 5.6. Let (X , k · k) be a normed K-vector space. Then (X , d ) is a metric space with
d k·k (x, y) := kx − yk and (X , d ) is in fact a topological K-vector space.

Proof. (i) To show: d is a metric. This is obvious.

(ii) Addition is continuous: Follows from the triangle-inequality:

k(x + y) − (x 0 + y 0 )k ≤ kx − x 0 k + ky − y 0 k.

Hence B ε (x 0 ) + B ε (y 0 ) ⊆ B 2ε (x 0 + y 0 ).

(iii) Multiplication by scalars is continuous: Let a, a 0 ∈ K, x, x 0 ∈ X . Then

kax − a 0 x 0 k = kax − ax 0 + ax 0 − a 0 x 0 k
≤ kax − ax 0 k + kax 0 − a 0 x 0 k
|a|kx − xc 0 k + |a − a 0 |kx 0 k
13 5 TOPOLOGICAL VECTOR SPACES

Lemma 5.7. Let X , Y be normed K-vector spaces. Let T : X → Y be linear. Then T is continuous
⇔ there exists C > 0 s.t.
kT xk ≤ C kxk ∀x ∈ X

N.B.: C is independend of x ∈ X

Proof. ⇐ T is continuous in 0, so T is continuous.

⇒ Let t be continuous in° 0.³ Let ε´°:= 1. By continuity ∃δ > 0 s.t. T (B δ (0)) ⊆ B 1 (0) ⊆ Y . So
∀x ∈ X with x 6= 0 we have °T δ2 kxk
x °
° ≤ 1. Hence δ2 kxk
x
kT (xk ≤ 1) and C = δ2 .
°

Remark 5.8. Let T be continuous between normed K-vector spaces X and Y , then

kT xk
≤ C ∀x ∈ X with x 6= 0
kxk

Therefore continuous linearmaps between normed spaces are bounded.

Definition 5.9. Let L : X → Y be continuous linear between normed K-vector spaces. Then

kT xk
kT k := sup
x6=0 kxk

is called the norm (or operator norm) of T .

Lemma 5.10. Let X , Y be normed K-vector spaces. Then L(X , Y ) is a normed K-vector space
with the operator norm.

Proof. (i) kT k = 0 ⇔ T = 0.

xk kT xk
(ii) Let a ∈ K. Then kaT k = sup kaT
kxk = sup |a| kxk = |a|kT k.
x6=0 x6=0
³ ´
(iii) kT1 + T2 k = sup kT1 x+T
kxk
2 xk
≤ sup kT1 xk+kT2 xk
kxk
1 xk
≤ sup kTkxk 2 xk
+ sup kTkxk = kT1 k + kT2 k.
x6=0 x6=0 x6=0 x6=0

Remark 5.11. By construction: kT xk ≤ kT kkxk ∀x ∈ X .

Definition 5.12. A complete normed K-vector space is called a Banach space.

Example 5.13. Let x be a compact topological space. Then C (X , R) is a Banach space with
norm defined by sup-norm.
14 6 QUOTIENT SPACES

6 Quotient spaces
Let X be a normed K-vector space. Let M ⊆ X be a closed subspace. Define an equivalence
relation x ∼ y if (x − y) ∈ M . Let X/M := X/∼. Define kx̃k := inf kx − yk.
y∈M

Lemma 6.1. This renders X/M into a normed K-vector space. Moreover, if X is a Banach space,
then also X/M is a Banach space. Also the natural projection map Π : X → X/M is continuous
and open, and kΠk ≤ 1
n→∞
Proof. (i) Obviously kx̃k ≥ 0. Suppose kx̃k = 0. Then ∃ sequence (y n )n∈N in M s.t. y n −−−−→
x, where x is a representant of x̃. Since M is closed, x ∈ M . Hence x̃ = 0 ∈ X/M .

(ii) a ∈ K, then kaxk = inf kax − a yk = |a| inf kx − yk = |a|kx̃ + ỹ.

y∈M y∈M

(iii) kx̃ + ỹk = inf

0
kx + y − z − z 0 k ≤ inf kx − zk + inf
0
ky − z 0 k.
z,z ∈M z∈M z ∈M

(iv) Clearly Π is linear.

(v) kΠ(x)k = kx̃k = inf kx − yk ≤ kxk. Hence kΠk ≤ 1 and Π is constant.

y∈M

(vi) We need to show: Π is open, i.e. ∀ open U ⊆ X , the image Π(U ) is open in X/M . Let
x̃ ∈ Π(U ) and consider a representant x of x̃ and ε > 0 s.t. B ε (x) ⊆ U . Consider a point
ỹ ∈ B ε (x). Then inf kx − y + zk < ε
z∈M
⇒ (y − z) ∈ B ε (x) ⊆ U
⇒ ỹ ∈ Π(U ). And since ỹ ∈ B ε (x̃) we have B ε (x̃) ⊆ Π(U ). Hence Π(U ) is open.

(vii) Let X be a Banach space. Let (x̃ n )n∈N be a Cauchy sequence in X/X . Pick representant
x 1 of x̃ 1 . Pick ε > 0. Since kx̃ 1 − x̃ 2 k = inf kx 1 − x 2 − yk. There exists representant x 2 of
y∈m
x̃ 2 s.t. kx 1 − x 2 k ≤ kx̃ 1 x̃ 2 k + ε. Since ε > 0 is arbitrary there ∃ representant x 2 of x̃ 2 s.t.
kx 1 − x 2 k ≤ kx̃ 1 − x̃ 2 k + 21 .
. . . etc.
∃ representant x n+1 of x̃ n+1 s.t. kx n+1 − x n k ≤ kx̃ n+1 − x n k+ 21n . Since (x̃ n )n∈N is a Cauchy-
n→∞
sequence, (x n )n∈N is Cauchy-sequence. Since X is complete, ∃x ∈ X with x n −−−−→ x.
n→∞
Hence x̃ n = Π(x n ) −−−−→ Π(x).
15 7 THEOREMS OF HAHN-BANACH

7 Theorems of Hahn-Banach
Definition 7.1. Let X be a R-vector space. A mapping p : X → R is called sublinear if

(i) p(λx) = λp(x) ∀λ ≥ 0, λ ∈ R.

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

Example 7.2. Let (X , k · k) be a normed vector space. Then p(x) := kxk is sublinear.

Let σ := {p : X → R | p is sublinear}. We put an order (≤) on σ by p ≤ p 0 iff p(x) ≤ p 0 (x) ∀x ∈

X

Definition 7.3. Let I be any set. An order "≤" on I is given if

(1) a ≤ a ∀a ∈ I (reflexivity)

(2) a ≤ b and b ≤ a, then a = b (antisymmetry)

(3) a ≤ b and b ≤ c, then a ≤ c (transitivity)

We call I totally ordered if ∀a, b ∈ I we have a ≤ b or b ≤ a. An element x is called maximal

(or minimal) if y ≥ x (or y ≤ x), then y = x. Let J ⊆ I be a subset. x ∈ I is called lower (upper)
bound for J if a ≥ x (or a ≤ x) ∀a ∈ J . I is called inductively ordered downwards or (upwards)
if every totally ordered subset of I has a lower (upper) bound.

Theorem 7.4 (Zorn’s Lemma). Let I be inductively ordered downwards (upwards). Then I has
a minimal (maximal) element.

Lemma 7.5. σ is inductively ordered downwards.

Proof. Let (p i )i ∈I with p i ∈ σ be totally orderd. Define p(x) := inf p i (x). Then ∀x ∈ X we have
i ∈I
p(x) > −∞. And since p is the infimum of a totally ordered family of sublinear functions, p is
sublinear.

Lemma 7.6. Let p ∈ σ. Then p is minimal iff p is linear.

Proof. ⇐ This is obvious by convexity arguments.

⇒ Let a ∈ X . Define p a : X → R by

¡ ¢
p a (x) := inf p(x + t a) − t p(a)
t ≥0
16 7 THEOREMS OF HAHN-BANACH

We show: p a (x) > −∞.

We have −p(−x) ≤ p(x + t a) − t p(a) ∀t ≥ 0, because

t p(a) = p(t a) = p(t a + x − x) ≤ p(t a + x) + p(−x).

⇒ p a (x) ≥ −p(−x) > −∞.

We now show that p is sublinear:

(i) Let λ ≥ 0. Then

¡ ¢
p a (λx) = inf p(λx + t a) − t p(a)
t ≥0
t t
µ µ µ ¶ ¶¶
= inf λ p x + a − p(a)
t ≥0 λ λ
t t
µ µ ¶ ¶
= λ inf p x + a − p(a)
t ≥0 λ λ
= λp a (x)

(ii) Let x, y ∈ X . We know ∀ε > 0 ∃t 1 > 0, t 2 > 0 s.t.

p a (x) ≥ p(x + t 1 a) − t 1 p(a) − ε

p a (y) ≥ p(y + t 2 a) − t 2 p(a) − ε

Set t := t 1 + t 2 . Then

p a (x) + p a (y) ≥ p(x + t 1 a) + p(y + t 2 a) − t 1 p(a) − t 2 p(a) − 2ε

≥ p(x + y + t a) − t p(a) − 2ε
≥ p a (x + y) − 2ε

Since this holds ∀ε > 0 we have p a (x) + p a (y) ≥ p a (x + y) and hence p a is sublinear.

Remark 7.7. Setting t = 0 in the definition of p a we have p a ≤ p. But since p is minimal by

assumption we have p a = p.
Then we have p(x) = p a (x) |{z}
≤ p(x + a) − p(a) and so p(x) + p(y) ≤ p(x + y) ≤ p(x) + p(y).
t =1
⇒ p is linear.
17 7 THEOREMS OF HAHN-BANACH

Theorem 7.8 (Hahn-Banach). Let X be a R vector space. Let p : X → R be sublinear. Then ∃

linear f : X → R s.t. f ≤ p.

Proof. Let σp = {p 0 ∈ σ | p 0 ≤ p}. Then σp is inductively ordered downwards. So by Lemma of

Zorn ∃ minimal element in σ. By definition of σp ⇒ f is minimal in σ. So by the lemma f is
linear. Hence f ≤ p.

Theorem 7.9. Let X be a R-vector space and p : X → R be sublinear. Let L ⊆ X be a linear

subspace and let f : L → R be linear s.t. f ≤ p|L . Then ∃ linear F : X → R with F ≤ p and
F |L = f .

Proof. Define p̃ : X → R via

¡ ¢
p̃(x) := inf p(x − y) + f (y) .
y∈L

The infimum is greater then −∞ because ∀y ∈ L

p(x − y) + f (y) ≥ p(−y) − p(−x) − f (−y) ≥ −p(−x).

And p̃ is sublinear:

(i) Let λ ≥ 0.

¡ ¢
p̃(λy) = inf p(λx − y) + f (y)
y∈L

λp(x − y) + λ f (y 0 )
¡ ¢
= inf
0y ∈L

= λp̃(x).

(ii) Let x, y ∈ X . Then ∀ε > 0 ∃y 1 , y 2 s.t.

p̃(x) ≥ p(x − y 1 ) + f (y 1 ) − ε
p̃(z) ≥ p(z − y 2 ) + f (y 2 ) − ε

So for y = y 1 + y 2

p̃(x) + p̃(z) ≥ p(x − y 1 ) + p(z − y 2 ) + f (y) − 2ε

≥ p(x + z − y) + f (y) − 2ε
≥ p̃(x + z) − 2ε.
18 7 THEOREMS OF HAHN-BANACH

This holds ∀ε > 0 therefore p̃ is sublinear. Note: The definition of p̃ implies p̃|L ≤ f . But
p̃|L must be linear, so p̃|L = f . By Hahn-Banach ∃ linear F : X → R s.t. F ≤ p̃ and since
p̃ ≤ p we have F ≤ p.

Remark 7.10. Reminder: Let X be a R-vector space and let p : X → R be sublinear. Then ∃
linear f : X → R s.t. f (x) ≤ p(x) ∀c ∈ X .

Theorem ((last time)). Setting as above. Additionally let L ⊆ X be a linear subspace and f : L →
R linear. Then ∃ extension F : X → R, F linear and F |L = f and F (x) ≤ p(x) ∀x ∈ X .

So far: K = R. What about K = C?

Observation: Let F : X → C be C-linear. Let F 1 := ℜ(F ). Then F (x) = F 1 (x) + i F 1 (i x).

Theorem 7.11. Let X be a K-vector space, where K ∈ {R, C}. Let p : X → R≥0 be a (semi) norm.
Let L ⊆ X be a linear subspace. Let f : L → K be linear with | f (x)| ≤ p(x) ∀x ∈ L. Then ∃ linear
extension F : X → K s.t. |F (x)| ≤ p(x) ∀x ∈ X .

Proof. Let K = R. Then any (semi) norm is sublinear. Let f : L → R be linear s.t. | f (x)| ≤
p(x) ∀x ∈ L. Then f (x) ≤ p(x)
f or al l x ∈ L. Therefore ∃ linear extension F : X → R s.t. F (x) ≤ p(x) ∀x ∈ X .
⇒ −F (x) = F (−x) ≤ p(−x) = p(x)
⇒ |F (x)| ≤ p(x) ∀x ∈ X .

Now let K = C. Let f 1 := ℜ( f ). Then | f 1 (x)| ≤ | f (x)| ≤ px() ∀x ∈ L. So by the above ∃

linear extension F 1 : X → R where F 1 is R-linear s.t. |F 1 (x)| ≤ p(x) ∀x ∈ X . Define F (x) :=
F 1 (x)+i F (i x). ⇒ F is C-linear. We need to show: |F (x)| ≤ p(x) ∀x ∈ X . For a properly choosen
a ∈ C with |a| = 1 we have that |F (x)| = aF (x).
⇒ |F (x)| = aF (x) = F (ax) = F 1 (ax) ≤ p(ax) < |a|p(x) = p(x).
| {z } | {z }
∈R+ ∈R+

Corollary 7.12 ((often known as Hahn-Banach)). Let (X , k · k) be a normed K-vector space. Let
L ⊆ X be a linear subspace and f : L → K be linear and continuous (i.e. k f k < ∞). Then ∃
extension F : X → K s.t. F |L = f and kF k ≤ k f k. Hence kF k = k f k.

Proof. Let p(x) := k f k · kxk. Then p is a norm (if f 6= 0). (previous thm.) ⇒ ∃ K-linear exten-
sion F : X → K s.t. |F (x)| ≤ p(x) = kxkk f k.
|F (x)| k f kkxk
⇒ kF k = sup kxk ≤ sup kxk ≤ k f k.
x∈X ,x6=0 x6=0
19 8 THE DUAL SPACE

8 The dual space

Remark 8.1. Reminder: Let X , Y be normed K-vector spaces. Then L(X , Y ) is a normed linear
xkY
space with norm T (X ) = sup kT
kxk X
.
x6=0
Goal: Study L(X , K), the "dual space".

Lemma 8.2. Let X , Y be normed. Assume that Y is a Banach space. Then L(X , Y ) is a Banach
space.

Proof. Let {Tn }n∈N be a Cauchy sequence in L(X , Y ). Fix ε > 0. Then kTn − Tm k ≤ ε ∀n, m ≥ N .
Fix x ∈ X . Then kTn x−Tm xk ≤ kTn −Tm kkxk ∀n, m ≥ N . Hence {Tn x}n∈N is a Cauchy sequence
in Y . Since Y is a Banach space ∃y ∈ Y s.t. lim (Tn x) = y. Define T x := lim (Tn x).
n→∞ n→∞
Claim: T ∈ L(X , Y ).
¡ ¢ ¡ ¢
linearity: T (x + y) = lim Tn (x + y) = lim (Tn x) + (Tn y) = T x + T y.
n→∞ n→∞
T is bounded (i.e. kT k < ∞): Fix x ∈ X . Then

kTn xk ≤ k(Tn − Tm )xk + kTm xk

= kTn − Tm kkyk + kTm kkxk.

Therefore
kTn k ≤ kTn − Tm k + kTm k,

and so
|kTn k − kTm k| ≤ kTn − Tm k ≤ ε ∀n, m ≥ N .

So (kTn k)n∈N is a Cauchy sequence in R. So ∃ M ∈ R+ s.t. M = lim kTn k < ∞. So

n→∞

n→∞
kTn xk ≤ kTn kkxk −−−−→ M kxk.

So kT xk ≤ M kxk and hence kT k ≤ M .

⇒ T ∈ L(X , Y ).
n→∞
It remains to proof that Tn −−−−→ T in L(X , Y ).
Fix x ∈ X . Then kTn x − Tm xk ≤ ε · kxk ∀n, m ≥ N .

kT x − T nX k ≤ kTn x − Tm xk + kTm x − T xk
≤ εkxk + kTm x − T xk
| {z }
m→∞
20 8 THE DUAL SPACE

Hence kT x − Tn X k ≤ εkxk ∀n ≥ N and therefore

kT x − Tn xk
kT − Tn k = sup ≤ ε ∀n ≥ N .
x6=0 kxk

Let X be a normed K-vector space. Then the dual space

X 0 := L(X , K)

is a Banach space. Note: X must not be Banach!

Remark 8.3. (1) T1 ∈ L(X , Y ) and T2 (Y , Z ). Then (T2 ◦ T1 ) ∈ L(X , Z ) and kT2 ◦ T1 k ≤ kT1 kkT2 k.

(2) If X is Banach, then End(X ) := L(X , X ) is Banach.

Definition 8.4. Let X , Y be normed K-vector spaces. Let T ∈ L(X , Y ). Define the dual map:
T 0 : Y 0 → X 0 by g 7→ T 0 (g ), where (T 0 (g ))(x) := g (T (x)).

Lemma 8.5. T 0 ∈ L(Y 0 , X 0 ). And we have in addition kT 0 k = kT k.

Proof. Linearity of T 0 follows from the definition. For the boundedness let g ∈ Y 0 . Then
kT 0 g k
kT 0 g k = kg ◦ T k ≤ kT kkg k. So kT 0 k = sup kxk ≤ kT k.
x6=0
To show: We also have kT 0 k ≥ kT k.
Let x 0 ∈ X . Let L := {λT x 0 | λ ∈ K} ⊆ Y . Let g̃ (T x 0 ) := kT x 0 k where x 0 is fixed, and extend g̃ to
L(L, K).
⇒ kg̃ k = 1. By Hahn-Banach ∃ extension g ∈ L(Y , K) s.t. g |L = g̃ , kg k = 1.
⇒ kT x 0 k = |g (T x 0 )| = |g ◦ T (x 0 )| ≤ kg ◦ T kkxk ≤ kT kkx 0 k ≤ kg ◦ T kkx 0 k = kT 0 (g )kkx 0 k ≤
kT 0 kkx 0 k, since kg k = 1.

0
xk
⇒ kT k = sup kT kT kkxk 0
kxk ≤ kxk ≤ kT k.
x6=0

Let X 00 := (X 0 )0 . Define i X : X → X 00 by i X (x)( f ) := f (x) ∀ f ∈ X 0 . Then i X is linear and

bounded, since |i X (x)( f )| = | f (x)| ≤ k f kkxk. Hence ki X k ≤ 1.

Lemma 8.6. i X : X → X 00 is an isometric embedding.

Proof. Since ki X k ≤ 1, we have ki X (x)k ≤ kxk. To show: ki X (x)k ≥ kxk.

By Hahn Banach we have f ∈ X 0 s.t. f (x 0 ) = kx 0 k for some fixed but arbitrary x 0 . Clearly
k f k = 1. Therefore ki X (x 0 )k ≥ |i X (x 0 )( f )| = | f (x 0 )| = kx 0 k.
Hence ki X k ≥ 1.
21 9 PRINCIPLE OF UNIFORM BOUNDNESS

Remark 8.7. X 00 is a Banach space. So, we can always embed X isometrically in a Banach
space.

Remark 8.8. Let i X 0 : X 0 → X 000 , and (i X )0 : X 000 → X 0 , then (i X )0 ◦ i X 0 = id X 0 and i X 0 ◦ (i X )0 = id X 000 .

Remark 8.9. A Banach space is uniquely determind by the geomentry of the unit sphere.

9 Principle of uniform boundness

Let X be a topological space. A set A ⊆ X is called dense if A = X . A set A is nowhere dense if
(X \ A) is dense. I.e. A is nowhere dense if A has no interior points.
B A nowhere dense ⇔ X \ A = X . But in general X \ A 6= X \ A.
Let (X , d ) be a metric space. And define

U (x, ε) := B ε (x) = {y ∈ X | d (x, y) < ε},

K (x, ε) := B ε (x) = {y ∈ X | d (x, y) ≤ ε}.

B in general B (x) 6= K (x, ε).

ε

Lemma 9.1. Let (X , d ) be a metric space. Then:

(i) A ⊆ X is nowhere dense if A does not contain open balls.

(ii) If A = A, then A is nowhere dense iff

(X \ A) ∩U (x, ε) 6= ; ∀x ∈ X , ε > 0

Proof. Obvious.

Lemma 9.2. Let (X , d ) be a complete metric space. Let K i = K (x i , εi ), x i ∈ X , ε1 > 0 s.t. K 1 ⊃

i →∞
K 2 ⊃ . . . and ε1 −−−→ 0.
∞
T
Then ∃!x ∈ X with {x} = Ki .
i =1

Proof. (x i )i ∈N is Cauchy sequence in X , since ∀ j ≥ i we have d (x i , x j ) ≤ qεi → 0. Let x :=

∞
K i . By continuity of d , we get that d (x 1 , x) ≤ εi ∀i ∈ N.
T
lim x i then x ∈
i →∞ i =1
22 9 PRINCIPLE OF UNIFORM BOUNDNESS

∞
T
To show: K i = {x}.
i =1
∞
T
Let y ∈ Ki . Then
i =1
i →∞
d (x, y) ≤ d (x, x i ) + d (x i , y) ≤ 2εi −−−→ 0

Hence x = y.

Theorem 9.3 (Baire). Let (X , d ) be a complete metric space. Let {A n }n∈N be a collection of closed
∞
A i . Then ∃n ∈ N s.t. A n contains an open ball.
S
subset. Suppose ∃ open ball U ⊆
i =1
Equivalent formulation: A countable union of closed nowhere dense subset has no interior
point.

Proof. Let A 1 , A 2 , . . . be a countable collection of closed nowhere dense subsets. Suppose ∃

∞
open ball U0 = U (x 0 , ε0 ) s.t. U0 ⊆
S
A i . Since A 1 is nowhere dense and closed, it follows that
i =1
(X \ A 1 ) ∩U0 6= ; and open. Therefore ∃ closed K 1 = K (x 1 ε1 ) with 0 < ε1 < 1 s.t. K 1 ⊆ (X \ A 1 ) ∩
U0 . Since A 2 is closed and nowhere dense. By (i i ) from lemma 9.2 ; 6= (X \ A 2 ) ∩U (x 1 , ε1 ) is
open. So ∃ K 2 = K (x 2 , ε2) with 0 < ε2 < 21 and ε2 < ε1 s.t. K 2 ⊂ (X \ A 2 ) ∩U (x 1 , ε1 ).
By construction K 2 ⊆ U (x 1 , ε1 ) ⊆ K 1 and so forth.
1
This gives K i = K (x 1 , ε1 ) with 0 < εi < i s.t. K i +1 ⊂ K i and K i +1 ⊂ (X \ A i +1 ) ∩ U (x i , εi ). The
collection {K i }i ∈N satisfies the assumption of the previous lemma. µ ¶
∞
T T S S
So ∃!x ∈ X s.t. {x} = K i . By construction, x ∈ (X \ A i ) = X \ A i .Therefore x ∉ Ai .
i =1 i ∈N i ∈N i ∈N
S
But x ∈ K 1 ⊂ U0 ⊂ Ai .
i ∈N
Definition 9.4. Let X be a set. Let F be a set of functions X → R. Then F is called pointwise
uniformely bounded if ∀x ∈ X ∃ constant K x > 0 s.t. F (x) ≤ K x ∀ f ∈ F .

Theorem 9.5. Let (X , d ) be a complete metric space. Let F ⊂ C (X , R) be a family of continuous

functions that are pointwise uniformely bounded. Then ∃ open ball U ⊂ X and C > 0,C ∈ R s.t.
f (x) ≤ C ∀ f ∈ F ∀x ∈ U .

Proof. Let n ∈ N. Define A n := {x ∈ X | f (x) ≤ n ∀ f ∈ F } Then A n = f −1 ((−∞, n]) is closed.

T
f ∈F
i
S
Also: Since F is pointwise uniformely bounded we have n f t y A n = X . By the first for-
n=1
mulation of Baire ∃m ∈ N s.t. A m contains an open ball U . By definition of A m we have
f (x) ≤ m ∀x ∈ U , ∀ f ∈ F .

Theorem 9.6. Let X be a Banach space and let Y be a normed space. Consider F ⊂ L(X , Y )
s.t. {x 7→ kT xk | T ∈ F } is pointwise uniformely bounded. Then ∃ constant C > 0,C ∈ R s.t.
kT k ≤ C ∀T ∈ F .
23 10 OPEN MAPPING THEOREM & CLOSED GRAPH THEOREM

Proof. The previous theorem implies ∃ open ball B 2δ (x 0 ) and C > 0 s.t.

kT xk ≤ C ∀x ∈ B 2δ (x 0 ) ∀T ∈ F.

Let x ∈ X with kxk = 1. Then

1 1 1 1 2C
kT xk = kT (δx)k = kT (δx) + T x 0 − T x 0 k ≤ (kT (δx + x 0 )k + kT x 0 k) ≤ (C +C ) = .
δ δ δ δ δ

Hence kT k = sup kT xk ≤ 2C
δ =: C̃ .
x∈X ,kxk=1

Corollary 9.7. Let X be a normed space, and let M ⊂ X be a subset. Suppose: ∀ f ∈ X 0 ∃K f > 0
s.t. k f (x)k ≤ K f ∀x ∈ M . Then M is bounded.

Proof. Define F := i X (M ) ⊂ X 00 := L(X 0 , R). Then F is pointwise uniformely bounded. There-

fore by the last theorem ki X (x)k ≤ C ∀x ∈ M . But since ki X (x)k = kxk.

Remark
µ 9.8. ¶Vice-versa, suppose M ⊂ X is bounded. Let f ∈ X 0 . Then | f (x)| ≤ k f kkxk ≤
k f k sup kxk .
x∈M
| {z }
Kf

Summary: M is bounded in X iff M is weakly bounded

Corollary 9.9. Let X be a Banach space and Y be a normed space. Let F ⊂ L(X , Y ) s.t. ∀ f ∈ Y 0
and ∀x ∈ X ∃K f ,x > 0 s.t. | f (T x)| ≤ K f ,x ∀T ∈ F . Then ∃C > 0 s.t. kT k ≤ C ∀T ∈ F .

Proof. Fix x ∈ X . Let M x := {T x | T ∈ F } ⊆ Y . By the last corollary M x is bounded in Y . There-

fore {x 7→ kT xk | T ∈ F } is pointwise uniformely bounded. By using theorem 9.6 kT k ≤ C for
some C > 0.

Remark 9.10. Principle of uniform boundedness is also called Banach-Steinhaus theorem.

10 Open mapping theorem & closed graph theorem

Recall: Let X be a Banach space and M = M ⊂ X be closed. Then (X /M ) is Banach. Moreover,

Π : X → X /M

is continuous, open and surjective.

24 10 OPEN MAPPING THEOREM & CLOSED GRAPH THEOREM

Plan for today: Now consider X , Y Banach spaces and a surjective T ∈ L(X , Y ). Then Y ∼
=
X /(ker T ) is an isometry and T : X → Y is open.
Remark 10.1. Look at the short exact sequence:

0 ker T X X / ker T 0
∼
=
Y

And X ∼
= ker T × Y is an isometry.

Lemma 10.2. Let (X ; k · k1 ) be a Banach space. Let k · k2 be another norm on X s.t.

id : (X , k · k1 ) → (X , k · k2 )

k·k2
is continuous. If ∃r > 0 s.t. U2 (0, 1) ⊂ U2 (0, r ) . Then U2 (0, 1) ⊂ U1 (0, 2r ) and k · k1 & k · k2
| {z } | {z }
unit ball in k · k2 r -ball in k · k1
are equivalent.

Proof. By assumption we have that U2 (0, 1) ⊂ U1 (0, r ) + U2 (0, 21 ). Rescaling gives: U2 (0, 21n ) ⊂
U1 (0, 2rn ) +U2 (0, 2n+1
1
)
f r or al l n ∈ N. Let y ∈ U2 (0, 1). To show: kyk1 < 2r .
Write y = x 1 +y 1 s.t. kx 1 k1 < r and ky 1 k2 < 21 . Iteration gives y n ∈ U2 (0, 21n ) with y n = x n+1 +y n+1
∞
and kx n k1 < 2rn , ky n+1 k2 < 2n+1
1 P
. Since X is Banach ∃x ∈ X with x = x n . Since : (X , k · k1 ) →
n=1
N
P N →∞
(X , k · k2 ) is continuous, we have x n −−−−→ x with respect to k · k2 .
n=1
N N →∞
1 P
Now y = x 1 + y 1 = x 1 + x 2 + y 2 = x 1 + x 2 + x 3 + y 3 = . . . with ky n k2 < 2n
. Then x n −−−−→ y
n=1
∞
P
with respect to k · k2 . Hence x = y and kyk1 = kxk1 ≤ kx n k < 2r . Therefore kyk1 < 2r .
n=1
We have just shown: y ∈ U2 (0, 1) ⇒ kyk1 < 2r . Hence U2 (0, 1) ⊂ U1 (0, 2r ). In other words
kyk1 ≤ 2r kyk2 ∀y ∈ X .
By continuity of id : (X , k·k1 ) → (X , k·k2 ) ∃C > 0 s.t kyk2 ≤ C kyk1 ∀y ∈ X . So the two norms
are equivalent.

Theorem 10.3. Let (X , k · k1 )&(X , k · k2 ) be two Banach spaces. If id : (X , k · k1 ) → (X , k · k2 ) is

continuous, then k · k1 & k · k2 are equivalent.

Proof.
∞
[ k·k2
X= U1 (0, n) .
n=1
25 10 OPEN MAPPING THEOREM & CLOSED GRAPH THEOREM

k·k2
By Baire ∃x 0 ∈ X , r 0 > 0, n 0 ∈ N s.t. U2 (x 0 , r 0 ) ⊂ U1 (0, n 0 ) .
k·k2
⇒ U2 (0, r 0 ) ⊂ U1 (−x 0 , n 0 )
k·k2
⇒ U2 (0, r 0 ) ⊂ U1 (0, n 0 + kx 0 k1 ) by triangle inequality
k·k2
⇒ U2 (0, r 0 ) ⊂ U1 (0, r10 (n 0 + kx 0 k1 ))
⇒ k · k1 & k · k2 are equivalent by the preceding lemma.

Theorem 10.4 (of bounded inverse). Let X , Y be two Banach spaces and T ∈ L(X , Y ). Suppose
that T is bijective. Then T −1 ∈ L(X , Y ).

Proof. Define another norm on the space Y :

kykT := kT −1 yk X

Then k · kT is a norm on Y . To show: (Y , k · kT ) is Banach.

Consider Cauchy sequence (y n )n∈N with respect to k · kT .

ky n − y m kT = kT −1 y n − T −1 y m k X

Then (T −1 y n )n∈N is Cauchy with respect to k · k X .

n→∞
Since X is Banach x 0 := lim (T −1 y n ) ∈ X . kT x 0 − y n kT = kx 0 − T −1 y n k −−−−→ 0. Therefore
n→∞
lim y n = T x 0 with respect to k · kT . Hence (Y , k · kT ) is Banach.
n→∞
But kykY = kT (T −1 y)kY ≤ kT kkT −1 yk X . Therefore id : (Y , k·kT ) → (Y , k·kY ) is continuous and
by the preceding theorem the two norms are equivalent. So ∃C > 0 s.t.

kT −1 yk X = kykT ≤ C kykY .

Hence (T −1 ) is indeed bounded.

Theorem 10.5 (Open mapping theorem). Let X , Y be two Banach spaces. Let T ∈ L(X , Y ) be
surjective. Then T is open.

Proof. T is continuous so ker T = T −1 (0) is closed. Therefore X / ker T is Banach and Π : X →

X / ker T is open.
Let T̂ : X / ker T → Y be the induced mapping. Then T̂ ∈ L(X / ker T, Y ) (this is easy). Clearly: T̂
is bijective. So T̂ −1 is continuous. Therefore T̂ is open and T = T̂ ◦ Π is open.

Remark 10.6. It follows from the proof, that T̂ : (X / ker T, k·k X / ker T ) → (Y , k·kY ) is an isometry.
26 10 OPEN MAPPING THEOREM & CLOSED GRAPH THEOREM

Definition 10.7. Let X , Y be noremd spaces. Then X × Y becomes a normed space via

k(x, y)k := kxk X + kykY

Remark 10.8. If X , Y are Banach, then X × Y is Banach.

Definition 10.9. Let X , Y be Banach. Let T ∈ L(X , Y ). Then the graph of T is G(T ) := {(x, T x) | x ∈
X} ⊂ X ×Y .

Theorem 10.10 (Closed graph theorem). Let X , Y be Banach and T : X → Y be linear. Then
T ∈ L(X , Y ) iff G(T ) is closed.

Proof. ⇒ Let ((x n , T x n )n∈N ) be a Cauchy sequence in G(T ). I.e. ∀ε > 0 ∃N ∈ N s.t.

kx n − x m k X + kT x n − T x m kY ≤ ε ∀n, m ≥ N .

So ∃x ∈ X , y ∈ Y with x = lim x n and y = lim T x n . By continuity of T we have T x = y.

n→∞ n→∞
n→∞
Therefore (x n , T x n ) −−−−→ (x, y). Hence G(T ) is closed.
n→∞
⇐ Let x = lim x n for some (x n )n∈N in X . We need to show: T x n −−−−→ T x.
n→∞
Let Π X : X ×Y → X with (x, y) 7→ x be the natural projection. Then Π X ∈ L(X ×Y , X ). And since
G(T ) is Banach we have that
Π X |G(T ) : G(T ) → X

is continuous and bijective.

Then it follows from the inverse mapping theorem, that

¢−1
Π X |G(T )
¡
: X → G(T )

is continuous. But:
¢−1
Π X |G(T ) )(x) = (x, T x) ∈ G(T ).
¡

Hence
¢−1 n→∞ ¡ ¢−1
Π X |G(T ) (x n ) −−−−→ Π X |G(T )
¡
(x).
n→∞
And hence T x n −−−−→ T x.

Corollary 10.11. Let X be Banach. Let X 1 , X 2 be two linear closed subspaces. If X 1 ∩ X 2 = {0}
and X = X 1 + X 2 , then
Φ : X 1 × X 2 → X with (x 1 , x 2 ) 7→ x 1 + x 2
27 11 HILBERT SPACES (PART I: BASICS)

is an isometry.

Proof. Φ is linear.
Φ is injective: Suppose (x 1 + x 2 ) = (x̃ 1 , x̃ 2 ), then (x 1 − x̃ 1 ) = (x 2 − x̃ 2 ) so x i = x̃ i for i = 1, 2.
Φ is surjective by assumption.
Φ is bounded: kx 1 + x 2 k ≤ kx 1 k X + kx 2 kx = k(x 1 + x 2 )k X 1 ×X 2 . And for kΦk ≤ 1 let x 2 = 0 so
Φ(x 1 , x 2 ) = x 1 so kΦk = 1.

Summary: Let X , Y be Banach, T ∈ L(X , Y ) surjective. Theorem:

(i) Y ∼
= X / ker T .

(ii) X ∼
= (ker T ) × X

Remark 10.12. Correction: Suppose T : (X , k·k X ) → (Y , k·kY ) be linear. Then T is called quasi-
isometric embedding if ∃C > 0 s.t. kxk X ≤ kT xkY ≤ C kxk X ∀x ∈ X . Then T is continuous
and injective. If T is also surjective, then T is called a quasi − i somet r y. In this case we
write X ∼
= Y . And T is called an isometry if C = 1. Last time: Let X be Banach and X 1 , X 2 ⊂ X
are closed linear subspaces s.t. X 1 ∩ X 2 = {0} and X 1 ∪ X 2 = X . Then Phi : X 1 × X 2 → X with
(x 1 , x 2 ) 7→ x 1 + x 2 is a quasi-isometry of Banach spaces.

Remark 10.13. Two norms k · k1 & k · k2 on X are equivalent iff id : (X , k · k1 ) → (X , k · k2 ) is a

quasi-isometry

11 Hilbert spaces (Part I: basics)

Definition 11.1. Let X be a K-vector space (K ∈ {R, C}). A hermitian form on X is a mapping

〈·, ·〉 : X × X → K

s.t. ∀x, x 0 , y ∈ X and α ∈ K we have:

(i) 〈x + x 0 , y〉 = 〈x, y〉 + 〈x 0 , y〉,

(ii) 〈αx, y〉 = α〈x, y〉,

(iii) 〈x, y〉 = 〈y, x〉.

Remark 11.2. • 〈x, y + y 0 〉 = 〈x, y〉 + 〈x, y 0 〉 by (i ) and (i i i ).

28 11 HILBERT SPACES (PART I: BASICS)

• 〈x, αy〉 = |α|〈x, y〉 by (i i ).

• 〈x, x〉 ∈ R.
Remark 11.3. If K = R then 〈·, ·〉 is symmetric bilinear form.

Definition 11.4. 〈·, ·〉 is called positiv semi-definit if 〈0, 0〉 ≥ 0 ∀x 6= 0. Scalar product := inner
product := positive hermitian form.

Theorem 11.5 (Cauchy-Schwarz-inequality). Let X be a K-vector space, and 〈·, ·〉 be positive

semi-definit. Then
k〈x, y〉k2 ≤ 〈x, x〉〈y, y〉
−〈x,y〉
Proof. 0 ≤ 〈x +λy, x +λy〉 = 〈x, x〉+λ〈y, x〉+λ〈x, y〉+λλ〈y, y〉. If 〈y, y〉 6= 0. Then set λ := 〈y,y〉
.
Then
k〈x, y〉k2 k〈x, y〉k2 k〈x, y〉k2
0 ≤ 〈x, x〉 − − +
〈y, y〉 〈y, y〉 〈y, y〉
k〈x,y〉k2
Therefore 0 ≤ 〈x, x〉〈y, y〉 − 〈y,y〉 . If 〈x, x〉 6= 0 proceed analorously.
If 〈x, x〉 = 〈y, y〉 = 0, set λ := −〈x, y〉. Then 0 ≤ −2|〈x, y〉|2 , hence 〈x, x〉 = 0 and then |〈x, y〉|2 =
0 = 〈x, x〉〈y, y〉.

Theorem 11.6. Let X be a K-vector space with a positive definite 〈·, ·〉. Then (X , k · k) is a
noremed space with
p
kxk := 〈x, x〉.

Proof. ⇒ x = 0.
(i) kxk = 0 |{z}
pos. def.

(ii) kαxk = |α|kxk is immediate.

(iii) (Triangle-inequality):

kx + yk2 = 〈x + y, x + y〉
= kxk2 + kyk2 + 〈x, y〉 + 〈y, x〉
≤ kxk2 + kyk2 + 2kxkkyk by CS-inequality
= (kxk + kyk)2

Lemma 11.7. Let X be normed with positiv definit 〈·, ·〉. Then 〈·, ·〉 : X × X → K is continuous
with respect to k · k.
29 11 HILBERT SPACES (PART I: BASICS)

Proof.

|〈x, x〉 − 〈x 0 , y 0 〉| = 〈x − x 0 , y〉 − 〈x 0 , y 0 − y〉
≤ |〈x − x 0 , y〉| + |〈x 0 , y 0 − y〉|
≤ kx − x 0 kkyk + kx 0 kky − y 0 k

Question: Given a norm k · k, when does it come from some 〈·, ·〉?

Lemma 11.8. Let (X , k · k) be a normed space. Then ∃ inner product 〈·, ·〉 on X that induced k · k
iff the parallelogram law holds:

kx + yk2 + kx − yk2 = 2kxk2 + 2kyk2 .

Proof. ⇒ Suppose kxk2 = 〈x, x〉, then the parallelogram law holds (straight forward compu-
tation).
⇐ 1st case: K = C :Define

1¡
kx + yk2 − kx − yk2 + i kx + i yk2 − kx − i yk2 .
¢
〈x, y〉 :=
4

Claim: This is an inner product inducing k · k.

(i) We show: 〈x + x 0 , y〉 = 〈x, y〉 + 〈x 0 , y〉.

⇔ kx + x 0 + yk2 − kx − x 0 − yk2 + +i kx + x 0 + i yk2 − kx + x 0 − i yk2

= kx + yk2 − kx − yk2 + i kx + i yk2 − kx − i yk2
+ kx 0 + yk2 − kx 0 − yk2 + i kx 0 + i yk2 − kx 0 − i yk2
30 11 HILBERT SPACES (PART I: BASICS)

The real part of left hand side & right hand side of previous equation are equal because

kx + x 0 − yk2 = 2kx + yk2 + 2kx 0 k2 − kx + y − x 0 k2 (A)

= 2kx 0 + yk2 + 2kxk2 − kx 0 + y − xk2 (B )
1 1
= kx + yk2 + kx 0 k2 + kx 0 + yk2 + kxk2 − kx + y − x 0 k2 − kx 0 + y − xk2
|{z} 2 2
1
2 ((A)+(B ))

= kx + yk2 + kx 0 + yk2
µ ¶
0 2 1 0 2
+ kx k − kx + (y − x) < )
2
µ ¶
2 1 0 2
+ kxk − kx + (y − x ) < )
2
= kx + yk2 + kx 0 + yk2
µ ¶
1 0 2 2
+ kx − (y − x)k − ky − xk
2
µ ¶
1
+ kx − (y − x 0 )k2 − ky − x 0 k2
2
= kx + yk2 + kx 0 + yk2 + kx + x 0 − yk2 − kx − yk2 − kx 0 − yk2

This shows equality of real parts. Similarily for imaginary parts.

Therefore
〈x + x 0 , y〉 = 〈x, y〉 + 〈x 0 , y〉.

(ii) From (i ) follows 〈αx, y〉 = α〈x, y〉 ∀α ∈ Z. Let α := m

n ∈ Q. Then

m 1 m 1 m 1 m
〈αx, y〉 = 〈 x, y〉 = n〈 x, y〉 = 〈 nx, y〉 = 〈mx, y〉 = 〈x, y〉 = α〈x, y〉.
n n n n n n n

By construction 〈·, ·〉 : X × X → X is continuous. Hence 〈αx, y〉 = α〈x, y〉 ∀α ∈ R.

Easy to show: 〈i x, y〉 = i 〈x, y〉 And so 〈αx, y〉 = α〈x, y〉 ∀α ∈ C.

(iii) 〈x, y〉 = 〈y, x〉 follows easily from the definition of 〈·, ·〉.

This finishes K = C.
2nd case: K = R : Define
1¡
kx + yk2 − kx − yk2
¢
〈x, y〉 :=
4
and proceed analogously to case K = C.

Definition 11.9. A Hilbert soace is a pair (X , 〈·, ·〉) of a K-vector space X and inner product
31 11 HILBERT SPACES (PART I: BASICS)

〈·, ·〉 s.t. (X , k · k) is complete. If (X , k · k) is not complete then X is called pre-Hilbert.

Lemma 11.10. Let (X , 〈·, ·〉) be a pre-Hlbert space. then the completion X̂ with respect to k · k is
a Hilbert space.

Proof. Let i : X → X̂ be the cannonical embedding. Since the parallelogram law holds in X
and since i (X ) is dense in X̂ , the parallelogram law holds in i (X ) = X̂ .

Theorem 11.11 (Riesz representation). Let X be Hilbert, and let j : X → X 0 , where X 0 = L(X , K)
is the dual space, defined as
j (y)(x) := 〈x, y〉 ∀x ∈ X .

Then:

(i) j is additive and anti-linear, i.e. j (αy) = ᾱ j (x).

(ii) j is an isometric embedding, hence injective.

(iii) j is surjective.

Proof. (i) is obvious.

(ii) • | j (y)(x)| = |〈x, y〉| ≤ kxkkyk ∀x ∈ X . So k j (y)k ≤ kyk.

³ ´
y
• j (y) kyk = kyk. Therefore k j (y)k ≥ kyk. Hence k j (y)k = kyk and j is an isometric
embedding.

(iii) To show: ∀F ∈ X 0 = L(X , K) ∃y ∈ X s.t. j (y) = F ⇔ F (x) = 〈x, x〉 ∀x ∈ X .

W.l.o.g.: kF k = 1. Then ∃ sequence (y n )n∈N in X s.t. ky n k = 1 ∀n, and lim |F (y n )| = 1.
n→∞
Multiply each y n with a carefully choosen αn ∈ K, |αn | = 1, and obtain F (αn , y n ) ∈ R≥0 .
Clearly: kαn y n k = 1. W.l.o.g.: y n are such that 0 ≤ F (y n ) ≤ q1 ∀n and lim F (y n ) = 1.
n→∞
Therefore ∀ε > 0 ∃N (ε) ∈ N s.t. F (y n ) ≥ 1 − 8ε ∀n ≥ N (ε).
⇒ F (y n + y m ) ≥ 2 − 4ε ∀m, n ≥ N (ε)
⇒ ky n − y m k2 = 2ky n k2 + 2ky m k2 − ky n + y m k2 = 4 − ky n + y m k2 ≤ 4 − (F (y n + y m ))2 ≤
2
4 − (2 − 4ε ) = ε − 16
ε
≤ ε.
Therefore (y n )n∈N is Cauchy. Since X is complete ∃y ∈ X with lim y n = y. By construc-
n→∞
tion kyk = 1 & F (y) = 1.
Claim: F = j (y).
1st case: K = R. Since kF k = 1
 
1 1¡
F (y − λx) − F (y) |{z} ky + λxk − kyk
¢
− ≤
λ | {z } λ
=1 kF k=1
32 11 HILBERT SPACES (PART I: BASICS)

⇒ F (x) ≤ λ1 ky + λxk − kyk . Therefore

¡ ¢

1¡ 1¡
ky − λxk − kyk ≤ − F (y − λx) − F (y)
¢ ¢
−
λ λ

1¡ 1¡
ky − λxk − kyk ≤ F (x) ≤ ky − λxk − kyk .
¢ ¢
⇒ −
λ λ
Now consider λ → 0. LHS & RHS then converge to
¯ ¯
d ¯¯ d ¯¯
ky + λxk = 〈y + λx, y + λx〉1/2 (because y 6= 0)
¡ ¢
d y λ=0
¯ d y λ=0
¯
1
= 〈y, y〉−1/2 2〈x, y〉
2
= 〈x, y〉.

Hence F (x) = 〈x, y〉.

This finishes K = R.
2nd case: K = C : Notice that any C-Hilbert space becomes an R-Hilbert space by con-
sidering 〈·, ·〉R := ℜ(〈·, ·〉). Let f := ℜ(F ). Then F (x) = f (x) − i f (i x). We have k f k = 1 and
f (y) = 1 = F (y).
By 1st case, we have that ℜ(〈x, y〉) = f (x) ∀x ∈ X , where y is exactly the one from 1st case.
Therefore

F (x) = f (x) − i f (i x)
= ℜ(〈x, y〉) − i ℜ(〈i x, y〉)
= ℜ(〈x, y〉) + i ℑ(〈x, y〉)
= 〈x, y〉.

Corollary 11.12. X is reflexive, i.e. the isometric embedding i X : X → X 00 is an isometry.

Proof. Consider the following commutative diagram:

iX
X X 00
∼
= j ∼
= j0
∼
=
X0 X0

where j is the Riesz map and j 0 is dual to j , and i X (x)(x 0 ) = x 0 (x).

33 12 ORTHONORMAL BASES IN HILBERT SPACES

Indeed, j 0 : X 00 → X 0 is an isomorphism, because:

Let T ∈ L(X , Y ) be an isomorphism, then T 0 ∈ L(Y 0 , X 0 ) is also an isomorphism:
Injectivity of T 0 : Let T 0 (y 0 )(x) = 0 ∀x ∈ X . Then y 0 (T X ) = 0 and so is y 0 (since T is surjective).
Surjectivity of T 0 : Let x 0 ∈ X 0 . We need to show: ∃y 0 ∈ Y 0 s.t. T 0 (y 0 ) = x 0 .

⇔ T 0 (y 0 )(u) = x 0 (u) ∀u ∈ X
⇔ y 0 (Tu) = x 0 (u) ∀u ∈ X
⇔ y 0 (v) = x 0 (T −1 v) ∀v ∈ Y since T is bijective

Set y 0 := x 0 ◦ T −1 .

12 Orthonormal bases in Hilbert spaces

Definition 12.1. Let X be a Hilbert space. Then x, y ∈ X are called orthogonal (x ⊥ y) if
〈x, y〉 = 0. Let M 1 , M 2 ⊆ X be two subsets. Then

M 1 ⊥ M 2 ⇔ 〈x, y〉 = 0 ∀x ∈ M 1 , ∀y ∈ M 2 .

Let M ⊂ X be a subset. Define

M ⊥ := {y ∈ X | 〈x, y〉 = 0 ∀x ∈ M }

is the orthogonal complement. A sequence {x n }n∈N is called orthonormal if 〈x i , x j 〉 = δi j .

Lemma 12.2 (Pythagoras). If x ⊥ y, then

kx + yk2 = kxk2 + kyk2 .

Proof. Obviously.

Lemma 12.3. Let M ⊆ X be a closed subspace. Then M ⊥ ∩ M = {0} and M + M ⊥ = X .

Proof. M ∩ M ⊥ = {0} is obvious.

Let y ∈ X . Consider the Riesz map j X : X → X 0 and restrict this map to M . Since M is closed,
M is Hilbert. Therefore j X (y)|M ∈ M 0 and by Riesz ∃y 1 ∈ M s.t. J X (y, x) = 〈x, y 1 〉 ∀x ∈ M .
34 12 ORTHONORMAL BASES IN HILBERT SPACES

Therefore

〈x, y〉 = j X (y)(x) = 〈x, y 1 〉 ∀x ∈ M

⇔ x ⊥ (y − y 1 ) ∀x ∈ M
⇔ y − y1 ∈ M ⊥

Therefore y = y 1 + (y − y 1 ) ∈ M + M ⊥ .
P
Goal: Make sence of an expression of the form xi .
i ∈I

Definition 12.4. Let X be a Hilbert space. We call°a family {x °i }i ∈I summable to x ∈ X if ∀ε > 0 ∃

° < ε. We then write
° P °
finite J ε ⊆ I s.t. ∀ finite J ⊆ I with J ε ⊆ J we have °
° x − x i °
i ∈J

X
x= xi .
i ∈I

Lemma 12.5. (i) {x i }i ∈I is summable to x ∈ X iff ∀ε > 0 ∃ finite J 0 ⊂ I s.t. ∀ finite J ⊆ I with
J ∩ J 0 = ; we have ° °
°X °
° x j ° < ε.
° °
° j ∈J °

(ii) {x i }i ∈I is summable to x ∈ X iff only countably many x i , i ∈ I are non-zero and for every
∞
P
order x 1 , x 2 , . . . of these (non-zero elements) we have that x i = x.
n=1
° °
°P °
Proof. (i) ⇒ Let {x i }i ∈I be summable to x. Then ∀ε > 0 ∃J 0 (finite) s.t. ° x j − x ° ≤ 2ε .
° °
° j ∈J 0 °
Now let J be finite with J ∩ J 0 = ;. Then
° ° ° °
°X ° ° X X °
° xj° = ° xj − xj°
° ° ° °
° j ∈J ° ° j ∈J ∪J j ∈J 0
°
0
° ° ° °
° X ° °X °
≤° x j − x° + ° x j − x°
° ° ° °
° j ∈J ∪J ° ° j ∈J °
0 0
ε ε
< + = ε.
2 2
35 12 ORTHONORMAL BASES IN HILBERT SPACES

° °
° P °
1
⇐ Pick ε = n. Then ∃J n (finite) s.t. J ∩ J n = ; implies ° x ° < 1 . Therefore
° °
° j ∈J n j ° n

à !
X
xn
j ∈J 1 ∪ j 2 ∪...∪J n

is a Cauchy sequence. Hence this sequence converges to x ∈ X . By construction: x =

P
xi .
i ∈I

∞
1
∀n ∈ N. Therefore x i = 0. So only
S
(ii) ⇒ Let J 1 be as in (i ). Let i ∉ J n . Then kx i k < n
n=1
countably many x i can be non-zero.
The rest is left as an exercise.

Remark 12.6. Recall:

1. If {x i }i ∈I is summable to x ∈ X then only countable many x i ’s are different from zero.

P
We write x = x i .
i ∈I

2. {x i }i ∈I is called orthonormal (ON-system), if 〈x 1 , x 2 〉 = δi j ∀i , j ∈ I .

Lemma 12.7. If {x i }i ∈I and {y i }i ∈I are summable, and α ∈ K, z ∈ X , then

(i) α αx i .
P P
xi =
i ∈I i ∈I
P P P
(ii) xi + yi = (x i + y i ).
i ∈I i ∈I i ∈I
¿ À
P P
(iii) xi , z = 〈x i , z〉.
i ∈I i ∈I

Proof. (i ) and (i i ) follow immediately. For (i i i ) let J ⊂ I be a countable set s.t.

X X ∞
X
xi = xi = x n = x.
i ∈I i ∈J n=1

Then ° ° ° ° ° °
° n ° ° X ∞ ° ° X ∞ °
x j ° kzk ≤ εkzk
X
°〈x, z〉 − 〈x j , z〉° = ° 〈x j , z〉° ≤ °
° ° ° ° ° °
j =1 j =n+1 j =n+1
° ° ° ° ° °

for n large enough.

36 12 ORTHONORMAL BASES IN HILBERT SPACES

Lemma 12.8. Let {x i }i ∈I be s.t. x i ⊥ x j ∀i 6= j . Then {x i }i ∈I is summable iff kx i k2 is summable

© ª

in the Hilbert space R. Then we have

°X °2 X
° °
° x i ° = kx i k2 .
° °
°i ∈I ° i ∈I

Proof. Let J ⊂ I be finite. By Pythagoras we have

° X °2
° °
2
kx j k = ° x j ° < ε2
X ° °
j ∈J j ∈J
° °

iff kx i k2 i ∈I is summable. This shows equivalence.

© ª
P
Let x = x i . Then
i ∈I

* + * + Ã !
kx i k2 .
X X X X X X X X
〈x, x〉 = x, xi =
|{z} 〈x, x i 〉 = x j , xi = 〈x i , x j 〉 |{z}
= 〈x i , x i 〉 =
i ∈I Lemma i ∈I i ∈I j ∈J i ∈I j ∈J x i ⊥x j i ∈I i ∈I

Theorem 12.9. Let X be a Hilbert space, and let {x i }i ∈I be an ON-system. Then

|〈x i , x〉|2 ≤ kxk2 ∀x ∈ X (Bessel).

P
(i)
i ∈I

|〈x i , x〉|2 = kxk2 iff x =

P P
(ii) 〈x, x i 〉x i (Parseval).
i ∈I i ∈I
° °2
° °
Proof. (i) Let J ⊂ I be finite. Then 0 ≤ °x − 〈x, x j 〉x j ° = kxk2 − |〈x, x j 〉|2 . Therefore
° P ° P
° j ∈J ° j ∈J
|〈x, x j 〉|2 ≤ kxk2 . So |〈x, x i 〉|2 exists and |〈x, x i 〉|2 ≤ kxk2 .
P P P
j ∈J i ∈I i ∈I
° °
2 2
° P ° P P
(ii) °x − 〈x, x i 〉x i °
°
° = kxk − |〈x, x i 〉| = 0 iff x = 〈x, x i 〉x i .
i ∈I i ∈I i ∈I

Corollary 12.10. Let X be Hilbert and let S = {x i }i ∈I be an ON-system. Fix x ∈ X . Then the set

S x = {x i ∈ S | 〈x i , x〉 6= 0}

is at most countable.
37 12 ORTHONORMAL BASES IN HILBERT SPACES

Proof. From Bessel we get S x,n = {x i ∈ S | |〈x, x i 〉| ≥ n} is finite ∀n ∈ N. Therefore S x =

S
S x,n
n∈N
is at most countable.

Corollary 12.11. Let X be Hilbert, and let {x n }n∈N be a countable ON-system. Then ∀x, y ∈ X

∞
X
|〈x, x n 〉〈y, x n 〉| ≤ kxkkyk.
n=1

µ ¶1 µ ¶1
∞
P ∞
P 2 ∞
P 2
Proof. |〈x, x n 〉〈y, x n 〉| ≤ |〈x, x n 〉| |〈y, x n 〉| ≤ kxkkyk.
n=1 |{z} n=1 n=1 |{z}
Hölder with p=q=2 B essel

Remark 12.12. Recall: Let X be a Hilbert space, M ⊆ X a closed linear subspace, then

M ∩ M ⊥ = {0} and X = M + M ⊥ .

This allows for defining orthogonal projectior P : X → M with (x 1 +x 2 ) 7→ x 1 ∀x 1 ∈ M , x 2 ∈ M ⊥ .

Clearly P ∈ L(X , M ) and kP k = 1.

Lemma 12.13. Let X be a Hilbert space, and S = {x i }i ∈I be an ON-system. Then the mapping
P
X 3 x 7→ 〈x, x i 〉x i is the orthogonal projection onto
i ∈I

span(S) = {finite linear combinations of elements of S}.

Proof. S 0 = {x i ∈ S | 〈x i , x〉 6= 0}. This is at most countable. Let x j ∈ S. Then

X X
〈x − P x, x j 〉 = 〈x − 〈x, x i 〉x i , x j 〉 = 〈x, x j 〉 − 〈x, x i 〉〈x i , x j 〉.
x i ∈S 0 x i ∈S 0

If x j ∈ S \ S 0 , then the right hand side is zero. If x j ∈ S 0 , then the right hand side is zero by
⊥
construction. Therefore 〈x − P x, x j 〉 = 0 ∀x j ∈ S. And so (x − P x) ∈ (span(S))⊥ = (span(S)) .
The last equality holds, because for M ⊆ X linear subspace we have M ⊆ M . So (M )⊥ ⊆ M ⊥ .
But if y ∈ M ⊥ ⇔ 〈y, x〉 = 0 ∀x ∈ M . Let x̃ ∈ M , then ∃x n → x̃. Therefore 〈x̃, y〉 = lim 〈x n , y〉 = 0.
n→∞
Therefore y ∈ (M )⊥ and we have (M )⊥ = M ⊥ . Hence (x − P x) ∈ (span(S))⊥ and P is indeed an
orthogonal projection onto span(S).

Definition 12.14. Let X ne a Hilbert space, and let {x i }i ∈I be an ON-system. Then {x i }i ∈I us

called complete (or maximal) if for any other ON-system {y j } j ∈J that contains every element
x i ∈ {x i }i ∈I , we have that {x i }i ∈I and {y j } j ∈J are equal sets. Maximal ON-systems are called
ON-bases.
38 12 ORTHONORMAL BASES IN HILBERT SPACES

Theorem 12.15. Let X be a Hilbert space and S = {x i }i ∈I and ON-system in X . Then the follow-
ing are equivalent:

(i) S is ON-bases.

(ii) S ⊥ = {0}.

(iii) X = span(S).
P
(iv) ∀x ∈ X : x = 〈x, x i 〉x i (Fourier decomposition).
i ∈I
P
(v) ∀x, y ∈ X : 〈x, y〉 = 〈x, x i 〉〈y, x i 〉.
i ∈I

(vi) ∀x ∈ X : Parseval holds, i.e. kxk2 = |〈x, x i 〉|2 .

P
i ∈I
n o
X
Proof. (i ) ⇒ (i i ) Suppose ∃x ∈ S ⊥ s.t. x 6= 0. Then S ∪ kxk
would be an extended ON-system
(to maximality of S).
(i i ) ⇒ (i i i ) Follows from (M ⊥ )⊥ = M (this is true for all linear subspaces M ⊆ X ). Apply this
to M =span(S).
(i i i ) ⇒ (i v) Last theorem implies x − P x = 0.
(i v) ⇒ (v) Simple calculation.
(v) ⇒ (vi ) Set x = y. n o
X
(vi ) ⇒ (i ) Suppose S is not maximal. Then ∃x 6= 0 s.t. S ∪ kxk is an extended ON-system. In
particular, x ⊥ x i ∀x i ∈ S. By (i v) we have 0 6= kxk2 = |〈x i , x〉|2 = 0 .
P
i ∈I

Theorem 12.16 (Existence of ON-bases). Every Hilbert space contains an ON-bases.

Proof. Order the set of ON-systems in X by inclusion. Apply Zorn’s lemma. Then maximal
elements are ON-bases.

Theorem 12.17 (Dimension theorem). (1) All ON-bases in a Hilbert space have the same car-
dinality.

(2) If X and Y are Hilbert spaces that have bases with the same cardinality, then they are iso-
metric.

(3) A Hilbert space X is separable iff ∃ countable ON-bases in X .

39 13 HERMITIAN (SELF-ADJOINT) OPERATORS

Proof. We show only (1), ((2)&(3) are exercise):

© ª
Let S and T be two ON-bases of X . Then Te = f ∈ T | 〈e, f 〉 6= 0 is (at most) countable ∀e ∈ S.
Then card(Te ) ≤ card(N). If card(T
µ e¶) < ∞ then we are done (by linear algebra). If card(Te ) =

Te ≤ card(S × N) = card(S).
S
card(N), then card(T ) ≤ card
e∈S
Likewise card(S) ≤ card(T ).

13 Hermitian (self-adjoint) operators

Remark 13.1. Recall Riesz theorem: j X : X → X 0 with x 7→ 〈·.x〉 is an isometry.

Definition 13.2 (& Theorem). Let X and Y be two Hilbert spaces and suppose A ∈ L(X , Y ).
Let the adjoint be defined by

〈Ax, y〉 = 〈x, A ∗ y〉 ∀x, y.

Then A ∗ ∈ L(Y , X ). We have kA ∗ k = kAk.

Proof. Construct A ∗ by considering the following commutative diagram:

A0
Y0 X0
jY j X−1

Y X
A∗

Notice that 〈Ax, y〉 = 〈x, A ∗ y〉 is equal to ( j X (A ∗ y))(x) = ( j Y (y))(Ax) |{z}

= (A 0 ( j Y ))(x) ∀x, y.
def. of A 0
∗ 0
⇔ j X (A y) = A ( j Y (y)) ∀y
⇔ j X ◦ A∗ = A0 ◦ jY
⇔ commutativity of the diagram.
Therefore A ∗ = j X−1 ◦ A 0 ◦ j Y ∈ L(Y , X ). Notice: k j X±1 k = k j Y±1 k = 1 and kAk = kA 0 k.
⇒ kA ∗ k = k j X−1 ◦ A 0 ◦ j Y k ≤ kA 0 k = kAk and
kAk = kA 0 k ≤ k j X ◦ A ∗ ◦ j Y−1 ≤ kA ∗ k. Hence kAk = kA ∗ k.

Lemma 13.3. ∗ : L(X , Y ) → L(Y , X ) with A 7→ A ∗ is an antilinear isometry. We have (A ∗ )∗ = A.

For A ∈ L(X , Y ) and B ∈ L(Y , Z ) we have (B ◦ A)∗ = A ∗ ◦ B ∗ .

Proof. Straight forward.

40 13 HERMITIAN (SELF-ADJOINT) OPERATORS

Lemma 13.4. Let A ∈ L(X , Y ) satisfy |〈Ax, y〉| ≤ C kxkkyk ∀x ∈ X , ∀y ∈ Y for some C > 0. Then
kAk ≤ C . Furthermore

kAk = inf C ∈ R ¯ |〈Ax, y〉| ≤ C kxkkyk ∀x, y| = sup |〈Ax, y〉| ¯ kxk = kyk = 1 .
© ¯ ª © ¯ ª

Proof.

|〈Ax, Ax〉| ≤ C kxkkAxk

≤ C kAkkxk2 .

p p
Therefore kAxk ≤ C kAkkxk. And so kAk ≤ C kAk. Hence kAk ≤ C .
Rest is definition of the operator norm.

Lemma 13.5. Let A ∈ L(X , X ) be such that |〈Ax, y〉| ≤ C kxk2 ∀x ∈ X . Then ∀x, y ∈ X we have:

|〈Ax, y〉 + 〈Ay, x〉| ≤ 2C kxkkyk. (∗)

If K = C, we have the stricter version:

|〈Ax, y〉| + |〈Ay, x〉| ≤ 2C kxkkyk.

Proof.
¡ ¢
〈A(x + y), x + y〉 − 〈A(x − y), x − y〉 = 2 〈Ax, y〉 + 〈Ay, x〉 .

2|〈Ax, y〉 + 〈Ay, x〉| ≤ C kx + yk2 + kx − yk2

¡ ¢

2 2
¡ ¢
≤
|{z} 2C kxk + kyk .
parallel. law

x
Replace x by α and y by αy for some α > 0:
µ ¶
1 2 2 2
|〈Ax, y〉 + 〈Ay, x〉| ≤ C 2 kxk + α kyk .
α

kxk
If y = 0, then (∗) is obvious. So let y 6= 0. Let α2 = kyk to obtain (∗).
This finishes the case K = R.
Case: K = C: Replace x by e i ψ and multiply the left-hand-side of (∗) by 1 = |e i φ |.
41 13 HERMITIAN (SELF-ADJOINT) OPERATORS

Then
|e i φ 〈Ae i ψ x, y〉 + e i φ 〈Ay, e i ψ x〉| ≤ 2C kxkkyk

and
| e i (φ+ψ) 〈Ax, y〉 + e i (φ−ψ) 〈Ay, x〉 | ≤ 2C kxkkyk
| {z } | {z }
β1 β2

Choose ψ and φ s.t.: both β1 and β2 are real & positive. Then the claim follows.

Definition 13.6. Let X be a Hilbert space and A ∈ L(X , Y ). Then A is called self-adjoint (or
hermitian) if A ∗ = A (i.e. 〈Ax, y〉 = 〈x, A ∗ y〉 ∀x, y).

Lemma 13.7. Let X be a C-Hilbert space. Then A ∈ L(X , X ) is hermitian iff 〈Ax, y〉 ∈ R ∀x ∈ X .

Proof. ⇒ We have 〈Ax, x〉 = 〈x, Ax〉 = 〈Ax, y〉.

⇐ And we have 〈Ax, x〉 = 〈Ax, x〉 = 〈x, Ax〉 = 〈A ∗ x, x〉. Therefore 〈(A − A ∗ )x, x〉 = 0 ∀x ∈ X . By
the preceding lemma we have 〈(A − A ∗ x, y〉 = 0 ∀x, y) Hence A = A ∗ .

Remark 13.8. K = R : Let A be hermitian. Then 〈Ax, y〉 = 〈x, Ay〉 = 〈Ay, x〉 = . . .

Lemma 13.9. Let X be a Hilbert space and A ∈ L(X , X ) be hermitian, then

kAk = inf C ∈ R ¯ |〈Ax, x〉| ≤ C kxk2 ∀x .

© ¯ ª

Proof. If |〈Ax, x〉| ≤ C kxk2 ∀x, then we have by the preceding lemma that

|〈Ax, y〉| ≤ C kxkkyk ∀x, y ∈ X .

Definition 13.10. Let A ∈ L(X , X ) be hermitian, then we say A ≥ 0 iff 〈Ax, x〉 ≥ 0 ∀x ∈ X . And
A ≥ B iff (A − B ) ≥ 0.

Lemma 13.11. Let A ∈ L(X , X ) be hermitian and A ≥ 0, then

|〈Ax, x〉|2 ≤ 〈Ax, x〉〈Ay, y〉 ∀x, y ∈ X .

Proof. Observe that X × X → K with (x, y) 7→ 〈Ax, y〉 is a positive semi-definit hermitian inner
product. Then apply the Cauchy-Schwarz inequality.
42 14 L P -SPACES

Theorem 13.12. Let {A i }i ∈N be a monotonially increasing (or decreasing) bounded sequence

n→∞
of hermitian operators. Then A n −−−−→ A (weakly) to a hermitian operator A ∈ L(X , X ), i.e.
i →∞
A i x −−−→ Ax ∀x ∈ X .

Proof. By construction, we have that {〈A i x, x〉}i ∈N converges for all fixed x ∈ X . Then we have
that also {〈A i x, y〉}i ∈N N converges for all fixed x, y ∈ X beacuse of polarization:
K = R : 〈Ax, y〉 = 41 〈A(x + y), x + y〉 − 〈A(x − y, x − y〉) .
¡ ¢

K = C : 〈Ax, y〉 = 14 〈A(x + i y), x + i y〉 − 〈A(x − i y, x − i y〉) .

¡ ¢
i →∞ i →∞
Therefore A i x −−−→ Ax (weakly), i.e. 〈A i x, y〉 −−−→ 〈Ax, y〉 ∀y.
Straight forward: A ∈ L(X , X ) and hermitian.
i →∞
To show: A i x −−−→ Ax ∀x ∈ X .
By construction, A m ≥ A n ∀m ≥ n. W.l.o.g.: A n ≤ 21 ∀n ∈ N (by rescaling).
Then we have kA m − A n k ≤ 1 ∀m ≥ n. And

〈(A m − A n )x, (A m − A n )x〉2 ≤

|{z} 〈(A m − AN )x, x〉〈(A m − A n )2 x, (A m − A n )x〉
last lemma

≤
|{z} 〈(A m − A n )x, x〉kxk2
kA m −A n k≤1

≤ εkxk2

for m, n large enough by weak convergence (A m − A n )x →

− 0.
Therefore k(A m − A n )xk4 ≤ εkxk2 ∀m, n large enough. So (A m − A n )x → 0.
i →∞
Hence A i x −−−→ Ax ∀x ∈ X .

14 L p -spaces
Goal: Study L p -spaces, i.e. declare: | f |p .
R

We call a subset I ⊆ Rd an interval if I = I 1 × I 2 × · · · × I d , where each I k is open, closed,

half-open, unbounded, bounded interval in R. We thus include:

I k ∈ {[a, b], (a, b), (a, ∞), {a}, (a, b], [a, b)} .

Define σd := bounded intevals i Rd .

© ª

Definition 14.1. An interval-function is a mapping ϕ : σd → R. We call ϕ

43 14 L P -SPACES

• monotone, if ∀I 1 , T2 ∈ σd with I 1 ⊆ I 2 we have ϕ(I 1 ) ≤ ϕ(I 2 ).

• additive, if ∀I 1 , I 2 , I ∈ σd with I 1 ∩ I 2 6= ; and I = I 1 t I 2 , then ϕ(I ) = ϕ(I 1 ) + ϕ(I 2 ).

Remark 14.2. (1) By definition we require that ; ∈ σd .

(2) If ϕ is monotone and additive, then ϕ(I ) = ϕ(I ) + ϕ(;) ⇒ ϕ(;) = 0. And since ; ⊆ I ⇒
ϕ(I ) ≥ 0 ∀I ∈ σd .

Definition 14.3. (1) ϕ is called regular, if ∀ε > 0, ∀I ∈ σd ∃ open I ∗ ∈ σd s.t. I ⊆ I ∗ and ϕ(I ) ≤
ϕ(I ∗ ) < ϕ(I ) + ε.

(2) A measure is an additive, monotone and regular interval function.

Example 14.4. (1) Let v d (·) be a volume-form on Rd , i.e., if I = 〈a 1 , b 1 〉×〈a 2 , b 2 〉×· · ·×〈a d , b d 〉.
d
Q
Then v d (I ) = |b i − a i |. Here 〈∈ {(, [} and 〉 ∈ {), ]}.
i =1

(2) Let N ⊆ Rd be a set without accumulation points. Let m : N → R≥0 . Define:

X
m(I ) := m(x).
x∈N ∩I

Definition 14.5. Let ϕ be a measure. We say that M ⊆ Rd has measure zero, if ∀ε > 0 ∃ se-
quence {I k }k∈N with I k ∈ σd s.t. M ⊂ ϕ(I k ) < ε.
S P
I k and
k∈N k∈N

Remark 14.6. (1) By regularity, one can choose all I k to be open previous definition.
S
(2) Let {M k }k∈N be a sequence of measure zero sets, then M := M k is a measure zero set.
k∈N

∞
ϕ(I n,k ) < ε/2n+1 and M n ⊆
P S
Proof. Choose intervals {I n,k }k∈N s.t. I n,k . Therefore
k=1 n∈N
ϕ(I n,k ) < ε.
S S P P
M⊆ I n,k and
n∈N k∈N n∈N k∈N

Definition 14.7. A function f is called ϕ-defined if ∃ a measure zero set M ⊆ Rd , s.t. f is

defined on Rd \M .

Remark 14.8. Let { f n }n∈N be a sequence of ϕ-defined functions. then by the previous remark
(2), ∃ common measure zero set ∀ f n , n ∈ N.
44 14 L P -SPACES

Definition 14.9. Two ϕ-defined functions f and g are called equal almost everywhere (a.e.)
if they agree outside a measure zero set. We write

f =ϕ g .

Analogously, f ≤ϕ g . We call a sequence { f n }n∈N ϕ-pointwise-convergent if ∃M ⊆ Rd , ϕ(M ) =

0 s.t. ∃ ϕ-definded f with f (x) = lim f n (x) ∀x ∈ Rd \M . Analogously for ϕ-monotonicity.
n→∞

14.1 Integration of step functions

Definition 14.10. A mapping f : Rd → R is called a step function is ∃ finitely many disjoint
bounded
µ interfals
¶ I 1 , . . . , I n ∈ σd s.t. f |I k ≡ c k is constant for all k ∈ {1, . . . , n} and f (x) = 0 ∀x ∈
∞
Rd \
S
Ik .
k=1

Let I0 Rd be the R-algebra of step functions on Rd

¡ ¢

Remark 14.11.

Therefore two intervals always admit a common finite subdivision.

Remark 14.12. If f , g ∈ I0 Rd , then inf f , g , sup f , g , | f | ∈ I0 Rd .

¡ ¢ © ª © ª ¡ ¢

Remark 14.13. The set of intervals defining step functions is not unique! For example let f
take the values α and β on two intervals:

α β
45 14 L P -SPACES

If we subdivide one of the intervals into smaller intervals:

β β
α β β
β
β

Then they define the same step functions.

Definition 14.14. Let f ∈ I0 Rd be a step function with defining intervals I 1 , . . . , I n . Let f |I k ≡

¡ ¢

c k ∈ R. Then the Lebesgue-integral of f is defined as

Z n
f d ϕ := c k ϕ(I k ).
X
k=1

Remark 14.15. This definition is independent of the choice (of subdivision) of the intervals I k .

Remark 14.16. We have

f d ϕ ≤ g d ϕ.
R R
(1) f ≤ g , then

(2) ¯ f d ϕ¯ ≤ | f |d ϕ.
¯R ¯ R

(3) The mapping I0 Rd → R with f 7→ f d ϕ is R-linear.

¡ ¢ R

14.2 Summable functions

© ª
Lemma 14.17. Let f n n∈N be a monotonically increasing sequence of step functions, s.t.
Z
f n d ϕ ≤ A ∀n ∈ N

for some A ∈ R. Then ∃ function f : Rd → R s.t. f n →ϕ f . (∗)

Remark 14.18. f is only ϕ-defined in gerneral.

Proof (sketch). 1. Part: (Assume f n ≥ 0 ∀n). Let k ∈ N, and let M k,l := x ∈ Rd ¯ f l (x) > k . Then
© ¯ ª

each M k,l is union of finitely many disjoint bounded intervals. Moreover, ϕ(M k,l ) < kA ∀k, l ∈ N
½ ¾
d ¯
because of (∗). Define M k,l := x ∈ R
¯
sup f l (x) > k . Then M k,1 ⊆ M k,2 ⊆ . . . ⊆ M k . by
l ∈N
monotonicity of f n . W.l.o.g.: M k,l +1 is opbtained froim M k,l by augmentating M k,l by a finite
set of intervals. Then these intervals form a sequence {I k, j } j ∈N for every fixed k. We have:
46 14 L P -SPACES

½ ¾
∞ ∞ ∞
A d
ϕ(I k, j ) ≤ ⊇ M k . But: M := x ∈ R
P S ¯ T
k
and I k, j ¯ sup f l (x) = ∞ = M k . Therefore M is
j =1 j =1 l ∈N k=1
a measure zero set.

2. Part: (General case). Define f˜n := f n − f 1 . Then f˜n ≥ 0 ∀n ∈ N. Obtain limit f˜ as in part 1
and define f := f˜ − f 1 .

Definition 14.19. Let I1 Rd be the set of functions obtained as limits of step functions as in
¡ ¢

the previous lemma. Define: f d ϕ := lim f n d ϕ ∀ f ∈ I1 Rd . Let

R R ¡ ¢
n→∞

³ ´ n ³ ´o
I2 Rd := f − g ¯ f , g ∈ I1 Rd
¯
.

© ª
Recall: Let f n n∈N be a sequence of step functions s.t.
Z
f n d ϕ ≤ A < ∞ ∀n ∈ N . (∗)

n→∞
Then ∃ f : Rd → R s.t. f n −−−−→ f . Then define
Z Z
f d ϕ := lim f n d ϕ.
n→∞

Let I1 Rd , ϕ := f : Rd → R arising from (∗) . Define: I2 Rn , ϕ as the vector space generated

¡ ¢ © ª ¡ ¢

by I1 Rd , ϕ , i.e.
¡ ¢
³ ´ n ³ ´o
I2 Rd , ϕ = f | f = f 1 − f 2 , f i ∈ Ii Rd , ϕ .

Definition 14.20. Elements of I2 Rd , ϕ are called summable.

¡ ¢

Definition 14.21. Let f ∈ I2 Rd , ϕ , i.e. f = f 1 − f 2 for f i ∈ Ii Rd , ϕ . Then

¡ ¢ ¡ ¢

Z Z Z
f d ϕ := f1d ϕ − f 2 d ϕ.

Remark 14.22. (1) This is well-defined.

(2) f , g ∈ I2 Rd , ϕ , then inf( f , g ), sup( f , g ), | f |, |g | ∈ I 2 Rd , ϕ .

¡ ¢ ¡ ¢

(3) The mapping I2 Rd , ϕ → R defined by f 7→ f d ϕ is R-linear and

¡ ¢ R

(i) f ≤ϕ g ⇒ f d ϕ ≤ g d ϕ.
R R

(ii) ¯ f d ϕ¯ ≤ | f |d ϕ.
¯R ¯ R
47 14 L P -SPACES

© ª
Theorem 14.23 (Beppo Levi, monotone convergence). Let f n n∈N be a sequence of summable
functions s.t.

f n d ϕ ≤ A < ∞ ∀n ∈ N.
R
(i)
© ª
(ii) fn n∈N is monotonically increasing.
n→∞
Then ∃ f ∈ I2 Rd , ϕ s.t. f n −−−−→ϕ f and f d ϕ = lim f n d ϕ.
¡ ¢ R R
n→∞

n→∞
Theorem 14.24 (Lebesgue). Let f n n∈N ⊆ I2 Rd , ϕ s.t. f n −−−−→ f . Suppose that ∃g ∈ I2 Rd , ϕ
© ª ¢ ¡ ¡ ¢

s.t. ¯ f n ¯ ≤ϕ g . Then f ∈ I2 Rd , ϕ and f d ϕ = lim f n d ϕ.

¯ ¯ ¡ ¢ R R
n→∞

Definition 14.25. A function f : Rd → R is called measurable if f =ϕ lim f n , where f n is a

n→∞
step function ∀n ∈ N.

(i) f ∈ I2 Rd , ϕ ⇒ f is measurable.
¡ ¢
Remark 14.26.

(ii) f ∈ C 0 Rd , R ⇒ f is measurable.
¡ ¢

(iii) f , g measurable, then ( f + g ), ( f · g ) are measure.

The set of measurable functions is closed under limits.

Lemma 14.27. (1) Let f n n∈N be a ϕ-convergent sequence of measurable functions. Then
© ª

f =ϕ lim f n is measurable.
n→∞
© ª © ª © ª
(2) Let f n n∈N be a sequence of measurable functions. Then inf f n n∈N and sup f n n∈N are
n∈N n∈N
measurable.

Theorem 14.28. If f is measurable, g ∈ I2 Rd , ϕ , and | f | ≤ϕ g , then f ∈ I2 Rd , ϕ .

¡ ¢ ¡ ¢

Proof. We have f = sup( f , 0) + inf( f , 0). So w.l.o.g. let f ≥ϕ 0. Let f = lim f n , where each f n is
| {z } | {z } n→∞
≥0 ≤0
a non-negative step function. Since f ≤ϕ g we have f = inf ( f n , g ). Since f n , g ∈ I2 Rd , ϕ , we
¡ ¢
n→∞ ¡
have inf( f n , g ) ∈ I2 Rd , ϕ and therefore inf( f n , g ) ⊆ I2 Rd , ϕ s.t., ¯inf( f n , g )¯ ≤ϕ g . The
¡ ¢ © ª ¢ ¯ ¯
n∈N
result then follows from Lebesgue’s theorem.

Corollary 14.29. Let f be measurable and let | f | ∈ I2 Rd , ϕ . Then f ∈ I2 Rd , ϕ .

¡ ¢ ¡ ¢

Corollary 14.30. Let f ∈ I2 Rd , ϕ and let g be measurable s.t. |g | is ϕ-bounded (from above).
¡ ¢

Then f · g ∈ I2 Rd , ϕ .
¡ ¢
48 14 L P -SPACES

14.3 Applications of monotone convergence (Beppo-Levi)

f d ϕ = 0.
R
Lemma 14.31. Let f ≥ϕ 0. Then f =ϕ 0 iff. f is summable and

Proof. ⇒ Integration is independend of the choise of the function up to a set of measure

zero.
⇐ Suppose f ∈ I2 Rd , ϕ , f d ϕ = 0. Let f k := k f , ∀k ∈ N. Then f k k∈N is monotonically
¡ ¢ R © ª

increasing and f k ∈ I2 Rd , ϕ (since f k d ϕ = 0) ∀k ∈ N. By Beppo-Levi ∃ f˜ ∈ I2 Rd , ϕ with

¡ ¢ R ¡ ¢

(i) f˜ =ϕ lim f k .
k→∞

f˜d ϕ = 0. Therefore the set M := x ∈ Rd | f˜(x) = ∞ has measure zero.

R © ª
(ii)

Theorem 14.32 (Fatou’s Lemma). Let f n ⊆ I2 Rd , ϕ s.t.

© ª ¡ ¢

(i) f n ≥ 0 ∀n ∈ N.

(ii) ∃ f =ϕ lim f n .
n→∞

If f n d ϕ ≤ A < ∞ ∀n ∈ N, then f d ϕ ≤ A.
R R

Proof. Define h k = inf f k , f k+1 , f k+2 , . . . . Then h k is measurable and h k ≤ |h k | ≤ f k ∀k ∈ N.

© ª

The sequence {h k }k∈N is monotonically increasing and h k d ϕ ≤ A < ∞ ∀k ∈ N, so by Beppo-

R

Levi ∃ f˜ ∈ I2 Rd , ϕ s.t. f˜ =ϕ lim h k and f˜d ϕ ≤ A.

¡ ¢ R
k→∞
Claim: f˜ =ϕ f .
Since we know f =ϕ lim f n , for almost all x ∈ Rd and ∀ε > 0 ∃N ∈ N s.t. ¯ f (x) − f n (x)¯ < ε ∀n ≥
¯ ¯
n→∞
N . So by construction of the h k we have ¯h k (x) − f (x)¯ ≤ ε ∀k ≥ N . Hence f =ϕ lim h k and so
¯ ¯
k→∞
f =ϕ f˜.

Remark 14.33. Fatou’s lemma is usually stated as

Z ³ ´ µZ ¶
lim inf f n d ϕ ≤ lim inf fn d ϕ .
n n

14.4 Measurable and summable sets

Let E ⊆ Rd . The characteristic function χE is defined as

1, x ∈ E ,
χE (x) :=
0, else.
49 14 L P -SPACES

Definition 14.34. We call E ⊂ Rd ϕ-measurable if χE is measurable. We call E ⊆ Rd ϕ-summable

if χE ∈ I2 Rd , ϕ , and we define ϕ(E ) := χE d ϕ.
¡ ¢ R

T S
Remark 14.35. (i) Let {E n }n∈N be measurable. Then E n and E n are both measurable,
n∈N n∈N
since
χ T
En = inf χE n , χ S
En = sup χE n .
n∈N n∈N n∈N n∈N

(ii) E is measurable iff. E Ù = Rd \E is measurable, since χE Ù = χRd − χE and vise versa.

(iii) Open and closed sets are measurable (use that continuous functons are measurable).

(iv) Measurable subsets of summable sets are measurable.