Functional Analysis

Winter Term 2021/22

After the lecture notes by Prof. Dr. Gero Friesecke

February 11, 2022

Technical University of Munich

Distribution of the notes or the link to the Gitlab repository beyond the participants of the course is not permitted.
Disclaimer
These lecture notes have been thoroughly written according to the lecture and handwritten notes
“Functional Analysis” by Prof. Friesecke. Nevertheless, there can and will be typos and errors which
we cannot take any responsibility for. Please take everything you read with a grain of salt and when
in doubt, refer to the official notes by Prof. Friesecke. If you should find any errors, please kindly
report them to [Link]@[Link].

1
Contents
1 Normed Vector Spaces and Banach Spaces 4

2 Basic Mathematics in Banach Spaces 15

2.1 Continuity of Algebraic Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 Notion of Basis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3 Failure of compactness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

3 Hilbert Spaces 22

4 Examples of Linear Functionals and Linear Operators 32

5 Open Mapping Theorem 35

5.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5.2 The Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.3 The Proof . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.3.1 Proof of Open Mapping Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 38
5.3.2 Translation Argument . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
5.3.3 Cauchy Sequence Argument . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.3.4 Translation Argument . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.4 Related Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

6 Invertibility, Notion of Spectrum 41

6.1 Invertibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
6.2 Analogue of Eigenvalues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

7 More about Invertibility: Index Theory 43

7.1 Definition of the Index, Main Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
7.2 Proof of the Main Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

8 Spectral Theory of Compact Operators 52

8.1 Comparison with Jordan Normal Form . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

9 Duality in Hilbert Spaces 65

10 Dual Spaces 70

11 Weak Convergence 79

2
Basic idea of Functional Analysis
View functions y : Ω ⊂ Rn → R (or C) as points in a suitable infinite-dimensional vector space Y .
View ordinary differential equations, partial differential equations, integral equations ??? as equations
of the form F (y) = b, where F is a map between infinite-dimensional vector spaces. Focus of lecture:
linear equations
Ay = b (0.1)
A linear map between infinite-dimensional vector spaces.
Goal: systematic theory of such equations, and linear maps A, which answers questions like

ˆ existence

ˆ uniqueness

ˆ continuous dependence on data

Methods: Linear Algebra + Analysis + new ideas

Example (Ordinary differential equations).

y ′′ (x) + a1 (x) y ′ (x) + a0 (x) y ′ (x) + a0 (x) y (x) = b (x) x ∈ I = [a, b]. (0.2)

Unknown: y : I → R, given: b : I → R, ??? a1 , a0 : I → R

Rewrite in form (0.1) Ay = b, A : Y → Z linear operator, Y, Z vector spaces.

Define the vector spaces

Y = {y : I → R : y 2 times continuously differentiable}

Z = {b : I → R : b continuous}

???: a1 , a0 continuous.
A : Y → Z linear map between infinite-dimensional vector spaces

(Ax) (x) := y ′′ (x) + a1 (x) y ′ (x) + a0 (x) y (x)

Then (0.2) ⇐⇒ (0.1), solve ODE ⇐⇒ solve the functional analysis ???
Many other examples: partial differential equations, quantum physics, science, engineering...

Literature
ˆ Tablet Script (enough for exam)

ˆ [5, Werner: Funktionalanalysis] (nice proofs, not so conceptional)

ˆ [1, Alt: Linear Functional Analysis] (good for PDEs)

ˆ [2, Dieudonné: Foundations of Modern Analysis] (9 volumes, conceptional)

ˆ [3, Hirzebruch, Scharlau: Einführung in die Funktionalanalysis] (lean, elegant)

ˆ [6, Yoshida: Funktionalanalysis] (comprehensive)

ˆ [4, Reed, Simon: Methods of Modern Mathematical Physics] (4 volumes, good for quantum
mechanics)

3
1 Normed Vector Spaces and Banach Spaces
Let K = R or C, X vector space over K
Definition 1.1. A map n : X → [0, ∞) is called norm if
(a) Homogeneity: n (λx) = |λ| n (x) ∀ x ∈ X, λ ∈ K

(b) Triangle inequality: n (x + y) ≤ n (x) + n (y)

(c) Positivity: n (x) ≥ 0, equality ⇐⇒ x = 0

Notation n (x) = ∥x∥. The pair (X, ∥·∥) is called a normed vector space. If the norm is clear from
context, one often speaks of X itself as a normed vector space.


Definition 1.2. Let X, ∥·∥ be a normed vector space. A sequence (xj )j∈∈N , xj ∈ X is called

(a) convergent to x ∈ X, notation: xj → x or limj→∞ xj = x, if ∥xj − x∥ → 0 (j → ∞).

| {z }
∈R

(b) convergent if ∃ x ∈ X : xj → x.

(c) Cauchy if ∀ ε > 0 ∃ N ∈ N : ∥xj − xk ∥ < ε ∀ j, k ≥ N .

Definition 1.3. A normed vector space X, ∥·∥ in which every Cauchy sequence is convergent is
called complete normed vector space or Banach space.
Examples.
1. Rn , Cn

2. Basic examples in infinite dimensions

3. The spaces lp

4. function spaces

Real and Complex Vector Spaces

Rn endowed with any norm is Banach. Canonical norm: euclidean norm
q
x = (x1 , . . . , xn ) ∈ Rn ∥x∥ = x21 + · · · + x2n .

Same for Cn : q
z = (z1 , . . . , zn ) ∈ C n
∥z∥ = |z1 |2 + · · · + |zn |2 .

Basic Examples in Infinite Dimensions

Simple spaces of sequences (idea: Rn ⇝ “R∞ ”)


(a) d = a1 , a2 , · · · = (an )n∈N : an ∈ K, an ̸= 0 only for finitely many n ∈ N

(b) c0 = (an )n∈N : an ∈ K, limn→∞ an = 0 null sequences

(c) c = (an )n∈N : an ∈ K, limn→∞ an exists convergent series

(d) l∞ = (an )n∈N : an ∈ K, supn∈N |an | < ∞ bounded sequences



all endowed with the supremum norm

∥(an )∥∞ := sup |an | (1.1)

n∈N

are normed vector spaces.

4
Proof that ∥·∥∞ is a norm. We prove the triangle inequality:

∥(an ) + (bn )∥∞ = sup |an + bn | ≤ sup |an | + sup |bn | = ∥(an )∥∞ + ∥(bn )∥∞
n | {z } n n
≤|an |+|bn |

Claim: c0 , c, l∞ are Banach spaces, d is not.

Proof that d is not a Banach space. We need to find a Cauchy sequence a(N ) , a(N ) ∈ d which


N ∈N
is not convergent. Take a = (an ) = n1 n∈N ̸∈ d.

(
(N ) 1, n≤N
a = χN · a, (χN )N ∈N =
0, else
1
a(N ) − a(M ) = if M > N, → 0 (N → ∞), that is, a(N ) → a in l∞
∞ N +1

Uniqueness of limits implies ∄b ∈ d : a(N ) → b (because a(N ) → a ̸∈ d). Hence a(N ) not convergent

in d.

“Moral”: d is “too small” or the choice of the norm does not reflect the defining properties of
the vector space that “all but finitely many components are zero”. To get a Banach space, the choice
of the norm must reflect the key properties of the vector space.

The Sequence Spaces lp

We now look at the sequence spaces lp (l ≤ p < ∞).
∞
( )
X p
p
l = (an )n∈N : an ∈ K, |an | < ∞
n=1

This means that an → 0 not to slowly. We define:

∞
!1
X p

(an )n∈N p
= |an |p .
n=1

There remain some questions:

?
(a) Is lp a vector space? I.e. a, v ∈ lp =⇒ a + b ∈ lp

(b) Is ∥·∥p a norm? Positivity ✓, Homogeneity ✓, Triangle-inequality?

(c) Is lp complete?

You should be ablePto show that the

P answer for (a) isP
yes.
p p p
We need to show: n |an | < ∞, n |bn | < ∞ =⇒ n |an + bn | < ∞. We majorize:

|an + bn |p ≤ (|an | + |bn |)p ≤ (2 max {|an | , |bn |})p

= 2p max {|an |p , |bn |p } ≤ 2p (|an |p + |bn |p )
!
X X
≤ 2p |an |p + |bn |p
n n

Theorem 1.1.

1. ∥·∥p is a norm on lp , in particular ∥a + b∥p ≤ ∥a∥p + ∥b∥p (Minkowski’s inequality)

5
 log (λA + (1 − λ)B)

λ log A + (1 − λ) log B

ap = A λA + (1 + λ)B bq = B
x

Figure 1: Illustration of the concavity of the log

 
2. lp , ∥·∥p is a Banach space.

Definition 1.4. A function f : (0, ∞) → R is called concave if

f (λA + (1 + λ)B) ≥ λf (A) + (1 − λ)f (B) ∀ A, B ∈ (0, ∞), λ ∈ [0, 1]. (∗)
We have the criterion f ′′ ≤ 0 ⇐⇒ f concave (if f is smooth).
Lemma 1.1 (Young’s inequality).
1 1 ap bq
a, b, p, q > 0, + = 1 =⇒ ab ≤ + (Y)
p q p q

Proof. First, we rewrite:

ap bq
 
p
1
q
1 1 p 1 q 1 p 1 q
(Y ) ⇐⇒ (a ) · (b ) ≤ p+ ⇐⇒ log (a ) + log (b ) ≤ log
q a + b .
p q log p q p q
′
log is concave, since log′′ = x1 = − x12 < 0. Apply (∗) with λ = p1 , 1 − λ = 1q , A = ap , B = bp .

Notation: a = (an ), b = (bn ) =⇒ ab is the sequence with components (ab)n = an · bn

(component-wise product)
1 1
Lemma 1.2 (Hölder’s inequality). Let 1 ≤ p ≤ ∞, q the solution to p + q = 1 (with the convention
q = ∞ if p = 1, q = 1 if p = ∞). Then
∥ab∥1 ≤ ∥a∥p ∥b∥q ∀ a ∈ lp , ∀ b ∈ lp .
Special case: p = q = 2, a = (a1, . . . , an , 0, . . . ), b = (b1 , . . . , bn , 0, . . . )
 1  1
n n 2 n 2
X X X
2  2
=⇒ |aj bj | ≤  aj bj Cauchy-Schwarz inequality in Kn
j=1 j=1 j=1

Proof. A := ∥a∥p , B := ∥b∥q , without loss of generality A, B > 0.

1 |an | p 1 |bn | q
   
|an | |bn |
· ≤ +
A B (Y ) p A q B
p q
∥ab∥1 1 ∥a∥p 1 ∥b∥q 1 1
=⇒ ≤ + = 1 + 1 = 1 =⇒ ∥ab∥1 ≤ AB
A·B p Ap q Bq p q

6
Proof of triangle inequality (using Hölder).
X
∥a + b∥pp = |an + bn |p
n
| {z }
≤|an +bn |·|an +bn |p−1 ≤|an |+|bn |·|an +bn |p−1
X X
≤ |am | |an + bn |p−1 + |bn | |an + bn |p−1
n
| {z } n
=:(|a+b|p−1 )n

≤ ∥a∥p |a + b|p−q q
+ ∥b∥p |a + b|p−1
Hölder q

Pulling out exponents under the p norm gives us:

!1  ! 1 r
X p X r·

∥|c|r ∥p = |c|r·p = |c|r·p  = ∥c∥r

r·p
n n

Hence:

|a + b|p−1 = ∥a + b∥p−1
(p−1)·q .
q

Now, a miracle occurs:

1 1 1 1 p
=1− =⇒ q = 1 = p−1 = =⇒ (p − 1) · q = p.
q p 1− p p
p−1

We use this and get:

 
|a + b|p−1 = ∥a + b∥p−1
p =⇒ ∥a + b∥ p
p ≤ ∥a∥ p + ∥b∥ p−1
p ∥a + b∥p
q
 
=⇒ ∥a + b∥1p ≤ ∥a∥p + ∥b∥p

Proof that lp is complete. Let a(ν) := a(ν) n a Cauchy sequence of sequences in lp , that is


n∈N

∀ ε > 0 ∃ N ∈ N: a(ν) − a(µ) < ε ∀ µ, ν ≥ N.

p

We need to show ∃ a ∈ lp : a(ν) − a p

→ 0 for ν → ∞.
Idea: 2 steps:
 
(ν)
1. Find candidate for limit by looking at components an
ν∈N

2. Estimate for a(ν) − a p

1.
 1
p
 p 1
p
X (ν) (µ) 
a(ν) (µ)
n − an = a(ν) (µ)
n − an ≤  aj − aj = a(ν) − a(µ) <ε ∀ µ, ν ≥ N
p
j

 
(ν) (ν)
i.e. an Cauchy. R or C complete implies ∃ an := limν→∞ an , so the sequence a with
ν∈N
components an is a candidate for the limit.

7
2. Here the idea(is to split the sequence into “head” and “tail”.
1, M ≤ R
Let (χR )n :=
0, else
χR a head, (1 − χR )a tail
χR a(ν) (1 − χR )a(ν)

χR a(ν) − χR a ≤ χR a(ν) − χR a(M ) + χR a(M ) − χR a

p p p
(ν) (M ) (M )
≤ a −a + χR a − χR a
p p

< ε + χR a(M ) − χR a
p
| {z }
P 1
p p
R (M ) (M )
= n=1 an −an → 0 since an →an
M →∞

So ∃ M (dependent on R) so that χR a(M ) − χR a p

< ε. This implies the following:

(i) χR a(ν) − χR a p
≤ ε + ε = 2ε ∀ ν ≥ N, ∀ R
(ii) ∥χR a∥p ≤ χR a − χR a − χR a(N ) p
+ χR a(N ) p
≤ 2ε + a(N ) p
∀ R.

If we now look at R → ∞, we get:

(ii) ∥a∥p ≤ a(N ) p

+ 2ε < ∞ =⇒ a ∈ lp .
(i) a(ν) − a p
≤ 2ε ∀ ν ≥ N =⇒ a(ν) → a in lp

Function Spaces
Goal: construct Banach spaces of continuous, differentiable or integrable functions

f : Ω ⊆ Rn → R or C.

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

∥f ∥∞ := sup |f (x)| .
x∈[a,b]

This is a normed vector space but not Banach.

2
f (n) (x)

1 f (x)

−3 −2 −1 1 2 3
q
1
Figure 2: Graphs of f (n) = x2 + n and f (x) = |x|

8
 q
Let f (n) = x2 + n1 , f (x) = |x|. Then we have:

f (n) − f → 0, f (n) ∈ C 1 , f ̸∈ C 1
∞

since f is not differentiable at 0. Reason: norm doesn’t capture that C 1 functions are differentiable
(this will be fixed in e)).

b) C ([a, b]) := {f : [a, b] → K : f continuous} with norm

 b
∥f ∥1 := |f (x)| dx.
a

This is again a normed vector space but not Banach:

√
Let [a, b] = [−1, 1], f (n) (x) = min {f (x) , n} , f (x) = √1 . Now we get:
| x|

 1
(n)
n 1 1
f −f = p −q dx
1 1
−n |x| 1
n
 1  1 r
n 1 n 1 √ n1 1 n→∞
≤2 p dx = 2 √ dx = 4 x |0 = 4 − 0 → 0,
0 |x| 0 x n

so f (n) ∈ C ([a, b]) but f ̸∈ C ([a, b]); f (n) is Cauchy but can’t be convergent to any g ∈ C ([a, b]),

f (x)

f (n) (x)

−1 −0.5 0.5 1
√
Figure 3: Graphs of f (n) (x) = min {f (x) , n} and f (x) = √1
| x|

since otherwise ∥g − f ∥1 ≤ g − f (n) + f (n) − f → 0 for n → ∞.

In the limit, continuity goes missing, since continuity is a “pointwise” property, but the norm only
reflects the average behaviour of a function.

c) Ω set, l∞ (Ω) := {f : Ω → K : supx∈Ω |f (x)| < ∞} bounded functions with the norm

∥f ∥∞ := sup |f (x)| .
x∈Ω

Theorem 1.2 shows that this is a Banach space.

d) Ω ⊆ Rn , Cb (Ω) := {f : Ω → K : supx∈Ω |f (x)| < ∞, f continuous} bounded continuous functions

(Note: this is a subspace of l∞ (Ω)) with norm

∥f ∥∞ := sup |f (x)|
x∈Ω

Theorem 1.3 shows that this is a Banach space.

9
e) C 1 ([a, b]) continuously differentiable functions from [a, b] to K
∥f ∥C 1 := ∥f ∥∞ + f ′ ∞
= sup |f (x)| + sup f ′ (x) .
x x

This norm has the extra term that fixes the problem from a).
More generally: Ω ⊆ Rn open bounded, Ω closure of Ω,

n Ω k times continuously differentiable,
o
C k Ω = f : Ω → K : all partial derivatives Dα f, |α|≤k have continuous extension to Ω
X
∥f ∥C k := ∥Dα f ∥∞ .
0≤|α|≤k

Multi-index notation: α = (a1 , . . . , an ) ∈ (N0 )n , |α| = α1 + · · · + αn ,

∂ α1 ∂ αn
Dα f (x1 , . . . , xn ) = ∂xα1 · · ·
∂xαn
f (x1 , . . . , xn ) (differentiate α1 times with respect to x1 , α2 times
1 n

with respect to x2 etc.) with the convention that D0 f = f .

f) Lp spaces (integrable functions), 1 ≤ p < ∞.
Ω measurable, µ Lebesgue measure on Rn
  
p
fp (Ω) := f : Ω → K : f measurable, |f | dµ < ∞
L
Ω
 1
p
∥f ∥p := |f |p dµ fp (Ω) ,
not a norm on L
Ω

the positivity is violated.


∥f ∥p = 0 ⇐⇒ |f |p dµ = 0
Ω
⇐⇒ f = 0 almost everywhere (a.e.) i.e. µ ({x ∈ Ω : f (x) ̸= 0}) = 0
̸ f = 0 everywhere
=⇒
(
1, x ∈ [−1, 1] ∩ Q
Ω = [−1, 1], f (x) =
0, else.

f (x)
1

0.5

−2 −1 1 2

Figure 4: Graph of f (x)

Way out: equivalence classes

n o
[f ] = g ∈ L fp (Ω) : f = g a.e. .
n o
fp (Ω) := [f ] : f ∈ L
Set of equivalence classes: L fp (Ω) with norm

 1
p p
∥[f ]∥p := f˜ = f˜ dµ , f˜ any element of [f ]
p Ω

10
 Independence of choice of f˜, because:
 
˜ p ˜p
f˜, f˜ ∈ [f ] =⇒ f˜ dµ = f˜ dµ.
Ω f˜=f a.e. Ω

Now, positivity is ok, because:


∥[f ]∥p =⇒ |f |p dµ = 0
Ω
=⇒ f = 0 a.e.
=⇒ f ∈ [0]
=⇒ [f ] = [0] = 0Lp (Ω) .

By Theorem 1.3, we will see that Lp (Ω) is Banach.

g) The space L∞ (Ω):

Ω ⊆ Rn measurable.

L
g∞ (Ω) := {f : Ω → K : f measurable, ∃ c > 0 : |f (x)| ≤ c a.e.}
n o
L∞ (Ω) := [f ] : f ∈ Lg∞ (Ω)
n o
[f ] := f˜ ∈ L
g∞ (Ω) : f = f˜ a.e.
n o
∥[f ]∥∞ := inf c > 0 : f˜ (x) ≤ c a.e. , f˜ any element of [f ]

ˆ Independent of the choice of f˜.

ˆ Norm
ˆ Banach

1
p
Ω |f |dµ p
ˆ If µ (Ω) < ∞, limp→∞ = ∥f ∥∞ ∀f ∈ L
g∞ (Ω)
µ (Ω)
| {z }
average value of |f |p

Theorem 1.2. Let Ω be any set. The space l∞ (Ω) = {f : Ω → K : ∥f ∥∞ < ∞} is Banach.

Proof. We need to show: Cauchy sequences in l∞ (Ω) are convergent. Let f (ν) be a Cauchy sequence
in l∞ (Ω). So we have:
∀ ε > 0 ∃ N : f (ν) − f (µ) < ε ∀ µ, ν ≥ N.
∞
Fix any x ∈ Ω. Hence

f (ν) (x) − f (µ) (x) < ε ∀ µ, ν ≥ N (∗)

 
=⇒ f (ν) (x) Cauchy sequence in K
ν∈N
(ν)
=⇒ ∃ lim f (x) =: f (x) .
K complete v→∞

So f : x 7→ f (x) is a candidate for the limit of f (ν) in l∞ (Ω).

Claim 1: f ∈ l∞ (Ω), Claim 2: f (ν) − f ∞ → 0 for ν → ∞.

Proof of Claim 1: Let µ → ∞ in (∗). We then get:

=⇒ | f (ν) (x) − f (x) | ≤ ε ∀ ν ≥ N (∗∗)

 
=⇒ |f (x)| = f (x) − f (ν) (x) + f (ν)

≤ f (x) − f (ν) (x) + f (ν) (x)

11
 ≤ε + f (ν) (x) ∀x ∈ Ω

=⇒ ∥f ∥∞ ≤ε + f (ν) <∞
∞

since f (ν) ∈ l∞ (Ω).

Proof of Claim 2: Take supx in (∗∗)

=⇒ f (ν) − f ≤ε ∀ ν ≥ N,
∞

i.e. f (ν) − f ∞
→ ∞.

Remark. A special case in Theorem 1.2 is Ω = N. Then we have:

 
∞
l (N) = (an )n∈N : sup |an | < ∞ = l∞
n

which is Banach.

Theorem 1.3 (Banach spaces of continuous/differentiable/integrable functions).

a) Let Ω ⊆ Rn . The space Cb (Ω) = {f : Ω → K : supx∈Ω |f (x)| < ∞} of the bounded continuous
functions is Banach.

b) Let Ω ⊆ Rn open and bounded. The space C k Ω of the k times continuous differentiable functions

is Banach.

c) Let Ω ⊆ Rn be measurable, 1 ≤ p ≤ ∞. The space Lp (Ω) of the p-integrable functions is Banach.

Remark. Continuous, differentiable, integrable functions form Banach spaces under suitable choices
of norms.

To deal with a), we introduce an abstract lemma using the notion of closed sets.


Definition 1.5 (Topological notions). Let X, ∥·∥ be a normed vector space. A set A ⊂ X is called

(a) open if ∀ x ∈ A ∃ ε > 0 such that Bε (x) = {y ∈ X : ∥y − x∥ < ε} ⊆ A

(b) closed if X \ A is open, or equivalently, if for every convergent sequence x(n) with x(n) ∈ A


n∈N
we have
lim x(n) ∈ A.
n→∞

(c) dense if its closure A equals X.

The closure, interior and boundary of A are defined as

\
A= B
B⊇A
B closed

or equivalently
n   o
A = x ∈ X : ∃ x(n) with x(n) ∈ A ∀ n, x(n) → x
[
int A = B
b⊆A
b open

δA = A \ int A

12
Lemma 1.3. Let (X, ∥·∥) be a Banach space, U ⊆ X a closed subspace. Then U is again Banach (if
we choose the norm on U to be the restriction of the norm on X to U ).

Remark. In practice, it is much easier to check for closedness than completeness as we only have to
deal with sequences that are already convergent in X.

Proof. Let u(j) be a Cauchy sequence in U . Then it is also a Cauchy sequence in X and because


X is Banach, it is convergent in X, i.e. ∃ x ∈ X : u(j) → x. Because U is closed, x ∈ U hence u(j)



is convergent in U .

Proof of Theorem 1.3 a). Since Cb (Ω) is a subspace of l∞ (Ω) and l∞ (Ω) is Banach, it is sufficient to
show that Cb (Ω) is closed (as a subset of l∞ (Ω)).
Suppose f (n) → f in l∞ (Ω) , f (n) ∈ Cb (Ω). Then we have to prove that f ∈ Cb (Ω). Boundedness is
clear. To continuity: Fix x0 ∈ Ω. We have to show that

∀ ε > 0 ∃ δ > 0 : |f (x) − f (x0 )| < ε ∀ x ∈ Ω with |x − x0 | < δ.

f (x0 ) f (x)

ε ε
3 3

f (n) (x0 ) ε f (n) (x)

3

Figure 5: Idea for the proof

Since f (n) → f in l∞ ,
ε
∃ n : sup f (n) (x) − f (x) < . (∗)
x 3
Since f (n) is continuous,
ε
∃ δ : f (n) (x0 ) − f (n) (x) < ∀ |x − x0 | < δ. (∗∗)
3
We have:

|f (x) − f (x0 )| = f (x) − f (n) (x) + f (n) (x) − f (n) (x0 ) + f (n) (x0 ) − f (x0 )

≤ f (x) − f (n) (x) + f (n) (x) − f (n) (x0 ) + f (n) (x0 ) − f (x0 )
ε ε ε
< + +
(∗),(∗∗),(∗) 3 3 3

Proof of Theorem 1.3 b). We will only show this for C 1 ([a, b]).

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

∥f ∥ = sup |f (x)| + sup f ′ (x) .
x∈[a,b] x∈[a,b]

The norm property is therefore fulfilled. But the non-trivial part is the completeness:
Let f (n) be Cauchy in C 1 ([a, b]). This implies:
′ ′
∀ ε > 0 ∃ N : f (n) − f (m) + f (n) − f (m) < ε ∀ m, n ≥ N
∞ ∞
′
n o n o
=⇒ f (n) , f (n) Cauchy in Cb ([a, b])

13
 ′
=⇒ ∃ f, g ∈ Cb ([a, b]) : f (n) − f → 0, f (n) − g → 0. (∗)
Thm. 1.3 a) ∞ ∞

We must therefore show that f is differentiable and that f ′ = g.

The fundamental theorem of calculus (“Hauptsatz”) gives for f (n) :
 x
′
f (n) = f (n) (x) − f (n) (a),
a | {z }
→f (x)−f (a)
n→∞
x
| {z }
→ a g

since f (n) → f uniformly, integration commuting with uniform conversion and f (n) → f uniformly.
With the fundamental theorem of calculus it follows that f is differentiable and that f ′ = g, i.e.
f ∈ C 1 ([a, b])
We substitute f ′ = g into (∗):
′
f (n) − f + f (n) f ′ →0
∞ ∞

i.e. f (n) → f in C 1 ([a, b]).

′
Corollary. If f (n) →Pf uniformly and f (n) → g uniformly, then f differentiable with f ′ = g.
Special case: fn (x) = nk=0 ak xk partial sum of a power series (“Potenzreihe”).
→ Power series can be differentiated termwise if the termwise differentiated series converges uniformly.


p 1
p
Proof of Theorem 1.3 c). We need to prove that ∥[f ]∥p = Ω f˜ dµ is a norm:

Hölder’s inequality: 
∀ p,q∈[1,∞)∪{∞} such that p1 + 1q =1,
|f · g| dµ ≤ ∥f ∥p ∥g∥q fp (Ω), g∈L
fp (Ω)
.
Ω ∀ f ∈L

Minkowski’s inequality:
∥f + g∥ ≤ ∥f ∥p ∥g∥p fp (Ω) .
∀ p ∈ [1, ∞) ∪ {∞} , ∀ f, g ∈ L
(Proof analogous to lp )
Proof of completeness: see measure theory or [5].

A priori, one might think that [f ] ∈ Lp (Ω) nowhere has a well defined pointwise value, since, for
fp (Ω)
any f ∈ L (
f (x) , x ̸= x0
g (x) := , A ∈ K arbitrary,
A, x = x0
satisfies g ∈ [f ]. /
But:
Theorem (Lebesgue’s

differentiation theorem). Let Ω ⊆ Rn be open, f ∈ L
f1 (Ω) (i.e. f : Ω → K
measurable, Ω |f | dµ < ∞). Then for a.e. x0 ∈ Ω:


Bδ (x0 ) f dµ
lim = f (x0 ) .
δ→0 µ (Bδ )
| {z  }
=: Bδ (x0 ) f dµ average of f over Bδ (x0 )

We use this to construct a canonical representation of [f ] which is unique and has well defined
pointwise values.
Take any f˜ ∈ [f ]. Define
( )
Ω[f ] := x0 ∈ Ω : lim f˜ dµ exists .
δ→0 Bδ (x0 )

14
Independent of choice of f˜
( 
limδ→0 Bδ (x0 ) f˜ dµ, if limit exists, i.e. if x0 ∈ Ω[f ]
f∗ (x0 ) :=
0, else

ˆ Independent of choice of f˜

ˆ f∗ ∈ [f ], since lim = f˜ (x0 ) a.e.

ˆ f continuous, then f∗ = f
fp (Ω). This is ok in the sense of identifying [f ] ∈ Lp (Ω) and its
In literature, people often identify L
canonical representation f∗ ∈ Lfp (Ω) and keeping in mind that [f ] = [g] ⇐⇒ f∗ = g∗ a.e.

2 Basic Mathematics in Banach Spaces

2.1 Continuity of Algebraic Operations
Basic question about notion of normed vector spaces: You have

(a) algebraic operations (addition, multiplication with scalar)

(b) an analytic operation (passing to the limit).

Are (a) and (b) compatible?

Theorem 2.1. Let X be a normed vector space. Then:

(a) if xn → x and yn → y, then xn + yn → x + y

(b) if λn → λ (in K) and xn → x, then λn xn → λx

(c) if xn → x then ∥xn ∥ → ∥x∥.

Proof. Sheet 2, exercise 7.

Example. If X is a normed vector space and U ⊆ X is a subspace, then the closure U is again a
subspace.

Proof (using theorem 2.1). Suppose x, y ∈ U . We need to prove:

λx + µy ∈ U λ, µ ∈ K.

We have:

∃ (xn ), (yn ) ∈ U : xn → x, yn → y
theorem 2.1
=⇒ λxn + µyn → λx + µy
U subspace

=⇒ λx + µy ∈ U

15
2.2 Notion of Basis
What is the “right” generalization of the algebraic concept of basis from finite-dimensional to infinite-
dimensional vector spaces?

Definition 2.1. Let X be a normed vector space. A set S ⊂ X is called basis of X, if

(i) all finite subsets of S are linearly independent

(ii) Span S = X

where (N )
X
Span S = λi ei : ei ∈ S, λi ∈ K, N ∈ N
i=1
is the set of finite linear combinations of the basis vectors.
S is called a Schauder basis if in addition

(iii) S is countable, i.e. S = {ei }∞

i=1 and for all x ∈ X there exists a unique αi ∈ K such that

N
X
x = lim αi ei
N →∞
i=1

Definition 2.2 (Dimension). Let X be a vector space. X is called finite-dimensional if there exists
a finite set {x1 , . . . , xn } such that
 
Xn 
X = Span {x1 , . . . , xn } = λj xj : λj ∈ K .
 
j=1

In this case, the minimal such n is called dimension of X (dim X). Otherwise, X is called infinite-
dimensional (dim X = ∞).

Examples.

1. dim X < ∞. Then we have Span S = XP which is equivalent to Span S = X and for all x ∈ X
there exist unique αi ∈ K such that x = i αi ei , S = {ei }i∈I .

2. X = lp (1 ≤ p < ∞)

a) S = {e1 , e2 , . . . } , ej = (0, 0, . . . , 0, 1, 0, . . . , 0) with a one in the j-th component as the only

non-zero entry. Then S is linearly independent and we have

Span S = d sequences with finitely many non-zero components

Span S = lp (Sheet2, exercise 9).

It follows that S is a basis.

b) S∗ = {e∗ , e1 , e2 , . . . } , e∗ = 1, 212 , 312 , . . . with the entry 1


j2
in the j-th component. This
basis is linearly independent as well and we have:

Span S∗ = lp (since already Span S = lp )

Span S∗ = {α∗ e∗ + a : α∗ ∈ K, a ∈ d} .

But the coefficients of a ∈ lp with respect to S∗ are not unique!

a = e∗
(i) a = 1 · e∗ + 0 · e1 + 0 · e2 + · · · + 0 · en + . . .
(ii) a = 0 · e∗ + 1 · e1 + 212 · e2 + · · · + n12 · en + . . . ,
so S is not a Schauder basis.

16
3. X = C ([0, 1]) continuous functions on [0, 1].
N
( )
X
P ([0, 1]) := f ∈ C ([0, 1]) : f (x) = ai xi for some ai ∈ K, N ∈ N subspace of polynomials
i=0

We have S ([0, 1]) = C ([0, 1]) by sheet 2, exercise 10. Furthermore:

(x)
S = {p0 , p1 , . . . } , pj = xj j = 0, 1, . . . .

Claim: S is linearly independent.

Suppose α0 p0 + · · · + αn pn ≡ 0, without all αi = 0. Without loss of generality, we can assume
that αn ̸= 0. Then we have after n differentiations with respect to x:

0 + · · · + 0 + αn n! · 1 ≡ 1,

which is a contradiction. We have Span S = P ([0, 1]), so Span S = C ([0, 1]) which implies that
S is a basis of P ([0, 1]) and that monomials are a basis of the space of continuous functions on
[0, 1].

Lemma 2.1 (Criterion for a Schauder basis). Let X be a Banach space with basis {ej }j∈J and
ek ̸∈ Span {ej }j̸=k . Then {ej }j∈J is a Schauder basis. for the subspace
 
 n
X 
U = x ∈ X : ∃ (αj )j∈N such that x = lim αj e j
 n→∞ 
j=1

Intuition: This condition strengthens linear independence as linear independence implies ek ̸∈

Span {ej } j ̸= k.
PN (N )
Proof. Suppose x = limN →∞ j=1 αj ej . We will show that the αj s are independent of N .
We fix some component k:
(N )
X (N )
αk ek + lim α ej = x.
| {z } N →∞ j≤N j
∈Span ek j̸=k
| {z }
∈Span{ej }j̸=k

Suppose

v+w =x v, v ′ ∈ Span ek
v ′ + w′ = x w, w′ ∈ Span {ej }j̸=k .

Then if v ̸= v ′ , we have

v′ − v
0 ̸= v ′ − v = x − w′ − (x − w) =⇒ ek = ′


∈ Span {ej }j̸=k ,
| {z } | {z } ∥v − v∥
∈Span ek
∈Span{ej }j̸=k

(ν)
which is a contradiction, hence v = v ′ . Thus αk is independent of ν and since k is arbitrary,
(ν)
all αj are independent of ν.

Remarks. The Assumption that X is Banach can be weakened to X being a normed vector
space, since we didn’t use completeness in the proof.
They do not imply U = X, see Example 3. Thus there can exist elements x ∈ X which can only
(n) (n)
be approximated as x = limn→∞ nj=1 αj ej with n − dependent coefficients αj .
P

17
Returning to example 2: The “good” basis {ej }j∈N satisfies the criterion ek ̸∈ Span {ej }j̸=k .

Proof. We have:
Span {ej }j̸=k = {a = (aj ) ∈ d : ak = 0}
Span {ej }j̸=k ⊆ {a = (aj ) ∈ lp : ak = 0} =: d˜
because:
   
(ν) ˜ lim (ν)
aj ∈ d, aj = (aj ) ∈ lp
ν→∞
(ν)
=⇒ |ak − ak | ≤ a − a(ν) →0
|{z} lp
=0
=⇒ ak = 0
=⇒ ak ̸∈ Span {ej }j̸=k
lemma 2.1
=⇒ {ej } is a Schauder basis.
S∗ = {e∗ , e1 , e2 , . . . } not Schauder because
 
1 1
e∗ = 1, 2 , 2 , . . .
2 3
 
1 1
= lim 1 · e1 + 2 · e2 + · · · + 2 eN ∈ Span (S∗ \ {e∗ }).
N →∞ 2 N

1
x0

x
x2 0.5

−1 −0.5 0.5 1
x3

−0.5

−1

Figure 6: Graph of the first monomials

∞
Coming back to the example, let X = C ([−1, 1]). We have the basis of the monomials xj j=0


which is not a Schauder basis (sheet 3) and not a Schauder basis fo the subspace C 0,1 ([−1, 1]) =
{f : [−1, 1] → K : f Lipschitz continuous}. We consider the Chebyshev polynomials
T0 (x) = 1
T1 (x) = x
T2 (x) = 2x2 − 1
T3 (x) = 4x3 − 3x
..
.
which is a Schauder basis of V (sheet 4). They are widely used in computational maths and are
however, not a Schauder basis of X, since one can constructPcontinuous but irregular functions
f such that there exists no sequence (αj ) with f = limn→∞ nj=1 αj ej

18
 1
T0
T3
T1
T2 0.5

−1 −0.5 0.5 1

−0.5

−1

Figure 7: Graph of the first Chebyshev polynomials

2.3 Failure of compactness

Recall:

Theorem (Bolzano-Weierstrass). Every bounded sequence in Rn or Cn possesses a convergent sub-

sequence. Especially, we have:
If A ⊂ R is a closed and bounded subset and f : A → R is continuous, f attains its maximum and
minimum on A, that is, there exist xmin , xmax ∈ A such that f (xmin ) ≤ f (x) ≤ f (xmax ) for all x ∈ A.

Natural question: Is there an analogue in normed vector spaces or at least in Banach spaces?

Theorem 2.2. Let X be a Banach space. The following are equivalent:

(i) dim X < ∞

(ii) Bolzano-Weierstrass holds, i.e., every bounded sequence has a convergent subsequence.

It would, however, be very useful to have an analogue for dim X = ∞, for instance to show the
existence of maximizers/minimizers for suitable optimization problems or convergence of approximate
solutions to linear equations Ax = b (≈ 41 of the history of functional analysis is about the search for
useful analogues here).

Definition 2.3 (Equivalence of norms). Let X be a vector space over K = R/C. Two norms ∥·∥ , |||·|||
on X are called equivalent if there exist m > 0, M > 0 such that

m ∥x∥ ≤ |||x||| ≤ M ∥x∥ ∀ x ∈ X.

p
Example. Let X = R2 , ∥x∥ = max {|x1 | , |x2 |} , |||x||| = x21 + x22 . Then we have:
r n o
max {|x1 | , |x2 |} =max |x1 |2 , |x2 |2
q
≤ |x1 |2 , |x2 |2
r n o
≤ 2 max |x1 |2 , |x2 |2
√
= 2 max {|x1 | , |x2 |} .

Lemma 2.2. Let X be a finite dimensional vector space. All norms on X are equivalent.

Proof. See e.g. Analysis 2. The proof uses the corollary about attainment of minimum and maximum
of continuous functions on closed sets of the Bolzano-Weierstrass theorem.

19
Corollary 2.1. Let X be a normed vector space and U ⊂ X a finite-dimensional subspace. Then U
is closed.

Proof. We choose a basis {e1 , . . . , en } of U, i.e., for all v ∈ U, there exist unique v1 , . . . , vn ∈ K such
that
v = v1 e 1 + · · · + vn e n .
We define a new norm on U: the euclidean norm of coefficient vectors
v
u n
uX
∥v∥2 = t vi2 .
i=1

With lemma 2.2, we have that there exist m, M > 0 such that

m ∥v∥ ≤ ∥v∥2 ≤ M ∥v∥ . (∗)

Now let v (j) be a sequence in U with limit v ∈ X. We now need to prove that v ∈ U. By (∗), we


get:
(n) (m) min{m,n}→∞
vi − vi ≤ v (n) − v (m) ≤ M v (n) − v (m) → 0,
2

since v (k) is convergent in X and therefore Cauchy in X. This is true for i = 1, . . . , n, so we have:


n o
(k) (k)
vi Cauchy ∀ i = 1, . . . , n =⇒ ∃ vi := lim vi ∀ i = 1, . . . n.
K complete k→∞

We have found or candidate for the limit of v (k) :



n
X
ṽ := vi ei ∈ U.
i=1

Again, by (∗), we have:

v
u n
1 1u X (k) 2 k→∞
v (k) − ṽ ≤ v (k) − ṽ = t vi − v˜i → 0,
m 2 m
i=1

so v (k) → ṽ is in X. With the uniqueness of limits, we get that v = ṽ =⇒ v ∈ U

Lemma 2.3 (Theorem about almost orthogonal elements). Let U be a closed subspace of a Banach
space X and let U ̸= X. Then for all δ ∈ (0, 1), there exists an xδ ∈ X with ∥xδ ∥ = 1 and

∥xδ − u∥ ≥ 1 − δ ∀u ∈ U

Interpretation and intuition in R3 : Choose xδ orthogonal to U. This achieves ∥xδ ∥ ≥ 1 ∀ u ∈ U i.e.

one can even take δ = 0 (see also figure 8). On sheet 3, we will se a subspace of C ([0, 1]) where δ = 0
is not possible.

Proof. Pick x ∈ X \ U. Claim: d := inf {∥x − u∥ : u ∈ U} > 0.

Proof of claim: Otherwise, there exists a sequence u(n) ∈ U with u(n) − x → x which
would imply that u(n) → x =⇒ x ∈ U. Because U is closed, we get x ∈ U which is a
contradiction.

Since inf {∥x − u∥ : u ∈ U} = δ > 0, there exists uδ ∈ U such that

d
∥x − uδ ∥ < .
1−δ

20
 U

0 ·
u

Figure 8: Interpretation in R3

xδ
U 1 x

0 d
>d
1−δ
uδ

Figure 9: Sketch of the argument

x−u
We take xδ := ∥x−u∥ and get ∥xδ ∥ = 1 (see also figure 9). Now let u ∈ U be arbitrary. Then:

x uδ
∥xδ − u∥ = − −u
∥x − uδ ∥ ∥x − uδ ∥

1
= x − (uδ + ∥x − uδ ∥ u)
∥x − uδ ∥ | {z }
∈U
1
≥ ·d
∥x − uδ ∥
1−δ
> · d = 1 − δ.
(∗) d

Proof of 2.2.
(i) =⇒ (ii): X ∼
= Kn , assertion is the Bolzano-Weierstrass theorem.
(ii) =⇒ (i): We prove this by contraposition. Let dim X = ∞. We have to show: There exists a
bounded sequence without convergent subsequence.

1. Choose x1 ∈ X, ∥x1 ∥ = 1 and define U1 := Span X1 . By lemma 2.1, we know that

U1 is closed. Furthermore, we have U ̸= X since dim U < ∞.
2. The theorem about almost orthogonal elements gives us with δ = 12 that there exists
x2 ∈ X such that ∥x2 ∥ = 1, ∥x2 − x1 ∥ ≥ 21 and U2 := Span {x1 , x2 }.
1
3. Again, lemma 2.3 gives us with δ = 2 that there exists x3 ∈ X such that ∥x3 ∥ = 1
and ∥x3 − xi ∥ ≥ 21 for i = 1, 2.
1
We iterate this and get a sequence (xn ) such that ∥n∥ = 1, ∥xn − xm ∥ ≥ 2 for all n ̸= m.
This is bounded but does not have a convergent subsequence.

21
3 Hilbert Spaces
Hilbert spaces are a special class of Banach spaces with nice properties.

Definition 3.1. Let X be a vector space over K = R, C. An inner product on X is a function

B: X × X → K

such that for all x, y, z ∈ X and all λ, µ ∈ K, we have:

Linearity in first variable: B (λx + µy, z) = λB (x, z) + µB (y, z)

Symmetry: B (x, y) = B (y, x)

Positivity: B (x, x) ≥ 0, equality if and only if x = 0.

(i) and (ii) have the obvious consequence of linearity up to conjugation in the second variable:

B (z, λx + µy) = λB (z, x) + µB (z, y) .

Notation: B (x, y) := ⟨x, y⟩.

Remark. In physics, one defines inner products to be linear in the second variable and conjugated
linear in the first variable.


Definition 3.2. A Hilbert space is a Banach space X, ∥·∥ such that there exists an inner product
⟨·, ·⟩ on X with the property p
∥x∥ = ⟨x, x⟩ ∀x ∈ X

Examples.
  
x1 y1 qP
n
Pn  ..   ..  n 2
1. X = R , ⟨x, y⟩ = i=i xi yi , x =  . , y =  . , ∥x∥ = i=1 xi euclidean norm (geomet-
xn yn
ric length in R and R by Pythagoras). See also figure 10, |x|2 = |x|2 + x23 = x21 + x22 + x23 .
2 3

x2

x3
x x

x2
·
x′

·
x1 x1

Figure 10: Illustration of Pythagoras in two and three dimensions

Pn
2. X = Cn , ⟨x, y⟩ = i=1 xi yi . We have symmetry:

n
X n
X
⟨y, x⟩ = yi xi = yi xi = ⟨x, y⟩ .
z=z def
i=1 i=1

22
 n Pn o
2 Pn p
3. X = l2 = (xn )n∈N : xn ∈ K, i=1 |x n | < ∞ , ⟨x, y⟩ = n=1 x i y i , ∥x∥2 = ⟨x, x⟩, so l2 is
Hilbert (we will see: lp , p ̸= 2 is not Hilbert).


4. X = L2 (Ω), Ω ⊆ Rd measurable, ⟨[f ], [g]⟩ = ˜ (x) g̃ (x) dx, f˜ ∈ [f ], g̃ ∈ [g] (i.e. f˜ = f a.e.,
Ωf
g̃ = g a.e.)

ˆ Independent of choice of representative

ˆ Induces the L2 -norm, i.e.:
s
2
f˜ f˜ ∈ [f ].
p
∥[f ]∥2 = ⟨[f ], [g]⟩ =
Ω

We will see that Lp , p ̸= 2 is not Hilbert.

Theorem 3.1. Let X be any vector space over K, ⟨·, ·⟩ : X × X → K any inner product on X. Then
p
∥x∥ := ⟨x, x⟩
p
is a norm on X (i.e. inner product axioms for ⟨·, ·⟩ imply the norm axioms for x 7→ ⟨x, x⟩).

Proof. We have to show for all x, y ∈ X; λ ∈ K:

Homogeneity: ∥λx∥ = |λ| ∥x∥

Positivity: ∥x∥ ≥ 0, equality if and only if x = 0

Triangle Inequality: ∥x + y∥ ≤ ∥x∥ + ∥y∥

We begin with Homogeneity

∥λx∥2 = ⟨λx, λx⟩ = λ ⟨x, λx⟩ = λλ ⟨x, x⟩ = |λ|2 ∥x∥2 .

def lin. in lin. up to conj def
first var. in second var.

Positivity is obvious. For the Triangle Inequality we have:

∥x + y∥ = ⟨x + y, x + y⟩ = ⟨x, x⟩ + ⟨x, y⟩ + ⟨y, x⟩ + ⟨y, y⟩

| {z }
=⟨x,y⟩
p p
≤ ⟨x, x⟩ + 2 ⟨x, x⟩ ⟨y, y⟩ + ⟨y, y⟩
lemma 3.1
= (∥x∥ + ∥y∥)2 .

In this proof, we have used a famous inequality, that we will now restate and prove:
Lemma 3.1 (Cauchy-Schwarz Inequality). Let X be any vector space over K with inner product ⟨·, ·⟩.
Then: p p
|⟨x, y⟩| ≤ ⟨x, x⟩ ⟨y, y⟩.

Remark. For l2 , we get the Hölder Inequality with p = 2:

v v
∞
u ∞ 2 uX
X uX u∞
xj yj ≤ t x +t y2. j j
j=1 j=1 j=1

For L2 (Ω), we get the Hölder Inequality for functions with p = 2:

 s s
2 ∀ f˜,g̃ : Ω→K measurable,
f˜g̃ ≤ f˜ 2 |g̃|2

˜2

2 .
Ω Ω Ω Ω |f | <∞, Ω |g̃| <∞

23
Proof of 3.1. Without loss of generality, we can assume x, y ̸= 0. Then we get for all λ ∈ K:

0 ≤ ⟨x − λy, x − λy⟩
pos.

= ⟨x, x − λy⟩ − λ ⟨y, x − λy⟩

lin. 1. var.
= ⟨x, x⟩ − λ ⟨x, y⟩ − λ ⟨y, x⟩ + |λ|2 ⟨y, y⟩ .
lin. up to conj in 2. var

Now the idea is to pick λ such that right hand side becomes smallest. We assume temporarily K = R
and get:

⟨x, x⟩ − 2λ ⟨x, y⟩ + |λ|2 ⟨y, y⟩ =: f (λ) .

We calculate the minimum:

0 = f ′ (λ) = −2 ⟨x, y⟩ + 2λ ⟨y, y⟩

⟨x, y⟩
⇐⇒ λ = .
⟨y, y⟩
We use this choice for K = C as well and get:
|⟨x, x⟩|2 |⟨x, y⟩|2 |⟨x, y⟩|2
0 ≤ ⟨x, x⟩ − − + ⟨y, y⟩
⟨y, y⟩ ⟨y, y⟩ ⟨y, y⟩2
|⟨x, y⟩|2
= ⟨x, x⟩ −
⟨y, y⟩
⇐⇒ |⟨x, y⟩|2 ≤ ⟨x, x⟩ ⟨y, y⟩ .

Lemma 3.2. Let X be a Hilbert space with inner product ⟨·, ·⟩. The inner product is continuous, that
is if xn → x and yn → y, then
⟨xn , yn ⟩ → ⟨x, y⟩ .

Proof. We have:

|⟨x, y⟩ − ⟨xn , yn ⟩| = |⟨x − xn , y⟩ − ⟨xn , y − yn ⟩|

≤ |⟨x − xn , y⟩| + |⟨xn , y − yn ⟩|
triangle ineq.
→∥x∥ (cont. of norm)
z }| { n→∞
≤ ∥x − xn ∥ ∥y∥ + ∥xn ∥ y − yn → 0.
Cauchy-Schwarz | {z } | {z }
→0 →0

Theorem 3.2 (Characterization of Hilbert Spaces). LetpX be a Banach space with norm ∥·∥. X
is Hilbert (i.e. there exists an inner product with ∥x∥ = ⟨x, x⟩ ∀ x ∈ X) if and only if the norm
satisfies the parallelogram identity
 
∥x + y∥2 + ∥x − y∥2 = 2 ∥x∥2 + ∥y∥2 ∀ x, y ∈ X.

Proof.
“ =⇒ ”: We have:

∥x + y∥2 + ∥x − y∥2 = ⟨x + y, x + y⟩ + ⟨x − y, x − y⟩
= ⟨x, x⟩ + ⟨x, y⟩ + ⟨y, x⟩ + ⟨y, y⟩ + ⟨x, x⟩ − ⟨x, y⟩ − ⟨y, x⟩ + ⟨y, y⟩
= 2 ∥x∥2 + 2 ∥y∥2 .

24
 x−y
y
x+y

Figure 11: Parallelogram Equation in R2

“ ⇐= ”: We will give this proof only for K = R. Define:

1 
∥x + y∥2 − ∥x − y∥2 .
B (x, y) :=
4
 
Then we have B (x, x) = 14 ∥2x∥2 − ∥0∥2 = ∥x∥2 , so we now just need to show that B is in
fact an inner product. Positivity is obvious. We show Additivity in the first component:
For u, v, w ∈ X, we have, by the parallelogram identity:
1 
B (u + w, v) + B (u − w, v) = ∥u + w + v∥2 − ∥u + w − v∥2 + ∥u − w + v∥2 − ∥u − w − v∥2
4
1 2 2 2 2

= 2 ∥u + v∥ + 2 ∥w∥ − 2 ∥u − v∥ − 2 ∥w∥
4
= 2B (u, v) .
 
Since B (0, y) = 1
4 ∥0 + y∥2 − ∥0 − y∥2 = 0 ∀ y, we have, letting w = u:

B (2u, v) = 2B (u, v)

and more generally

B (u + w, v) + B (u − w, v) = B (2u, v) .

Let x, x′ , y ∈ X be given. Choose u, w such that x = u + y, x′ = u − w, y = v (i.e. take

′ ′
u := x+x
2 , w
:= x−x
2 ). Then

B (x, y) + B x′ , y = B x + x′ , y ,

which proves additivity in the first component.

Symmetry can be shown very easily from the definition:
1  1 
B (y, x) = ∥y + x∥2 − ∥y − x∥2 = ∥x + y∥2 − ∥x − y∥2 .
4 4
Now to linearity in the first component. With λ = 0, −1 we have:

B (0 · x, y) = 0 = 0 · B (x, y) ✓
1  1 
B (−x, y) = ∥−x + y∥2 − ∥−x − y∥2 = ∥x − y∥2 − ∥x + y∥2 = −B (x, y) .
4 4
We now generalize this by taking λ ∈ Z:

B (2x, y) = B (x + x, y) = B (x, y) + B (x, y) = 2B (x, y)

B (3x, y) = B (x + x + x, y) = B (x, y) + B (x, y) + B (x, y) = 3B (x, y)
..
.

25
 p
so with the rule for λ = −1, we have shown it for λ ∈ Z. We now go on to λ = q ∈ Q:

qB (λx, y) = B (qλx, y) = B (px, y) = pB (x) = qλB (x, y) .

If we divide by q on both sides, we get the desired result.

Now the only thing that is missing, is the generalization for λ ∈ R. Since Q is dense in R,
we can find a sequence (λn )n∈N in Q with λn → λ. We then get:
  
B (λx, y) = B lim λn x, y
n→∞
 
= B lim (λn · x), y
n→∞
= lim B (λn x, y) = lim (λn · B (x, y)) = λ · B (x, y) .
∥·∥ cts. n→∞ n→∞

For K = C, define
1 
B (x, y) := ∥x + y∥2 − ∥x − y∥2 + i ∥x + iy∥2 − i ∥x − iy∥2 .
4

Example. C ([0, 1]) = {f : [a, b] → K : f continuous} with the supremum norm ∥f ∥ = supx∈[a,b] |f (x)|
is not Hilbert.

1
f g

−g
-1

We take f and g mirrored to each other on [a, b] and get ∥f ∥ = ∥g∥ = ∥f + g∥ = ∥f − g∥ = 1 and
therefore:  
∥f + g∥2 + ∥f − g∥2 = 1 · ∥f ∥2 + ∥g∥2 .

Definition 3.3. Let X be a Hilbert space with inner product ⟨·, ·⟩. Two elements x, y ∈ X are called
orthogonal if ⟨x, y⟩ = 0. The orthogonal complement of a set M ⫅ X is defined as

{x ∈ X : ⟨x, v⟩ = 0 ∀ v ∈ M } =: M ⊥

Remarks (Obvious properties).

(i) Pythagorean theorem: x, y ∈ X orthogonal =⇒ ∥x∥2 + ∥y∥2 = ∥x + y∥2

Proof.
∥x + y∥2 = ⟨x + y, x + y⟩ = ∥x∥2 + ⟨x, y⟩ + ⟨y, x⟩ + ∥y∥2 .
| {z } | {z }
=0 =0

(ii) For all sets M ∈ X, M ⊥ is a subspace of X.

26
 Proof. We take x, y ∈ M ⊥ and get

⟨λx + µy, v⟩ = λ ⟨x, y⟩ + µ ⟨y, v⟩ = 0 ∀ v ∈ M,

so λx + µy ∈ M ⊥ .

Remark (On the definition of orthogonality). In R2 with the standard inner product, we have
⟨x, y⟩ = x1 y1 + x2 y2 . We define:
 
x x
cos θ := , ̸ 0.
x, y =
∥x∥ ∥x∥
Hence we get:

⟨x, y⟩ = 0 ⇐⇒ cos θ = 0
⇐⇒ θ = 90° or 270° for θ ∈ [0, 2π]

Theorem 3.3 (Projection Theorem). Let X be a Hilbert space and V ∈ X a closed subspace. For
any x ∈ X, we have:
(i) There exists a unique best approximation in V , i.e. a unique v ∈ V such that

∥x − v∥ ≤ x − v ′ v′ ∈ V

(ii) There exist unique elements v ∈ V, v ⊥ = V ⊥ such that

x = v + v⊥.

Moreover, this v is the unique best solution from (i).

x
V
0 v′
v ·

Figure 12: Intuition in R3

Remark. (i) means existence of a minimizer of an infinite-dimensional problem. In finite dimension

(i.e. dim V < ∞), it follows from the continuity of v 7→ f (v) = ∥x − v∥ and Bolzano-Weierstrass.
But in infinite dimensions (i.e. dim V = ∞), Bolzano-Weierstrass fails (see Section 2), so the result is
highly non-trivial.

Proof.
(i) Let d := inf ṽ∈V ∥x − ṽ∥. We choose a minimizing sequence (vn )n∈N , i.e. vn ∈ V , ∥x − vn ∥ → d
for n → ∞.
Claim: (vn )n∈N is convergent.
Since X is Hilbert, it suffices to show that (vn )n∈N is Cauchy, i.e. ∥vm − vn ∥ → 0 for min {m, n} →
∞.
We use the Parallelogram Equality with a = x − vm +v 2
n
, b = vm −v
2
n
:

vm + vn 2 vm − vn
2 ∥a∥2 + 2 ∥b∥2 = 2 x − +2 = ∥x − vn ∥2 + ∥x + vm ∥
2 2
| {z }
vm +vn
≥2d2 since 2
∈V

27
 min{m,n}→∞
= ∥a + b∥2 + ∥a − b∥2 → ∞
vm − vn
=⇒ 2 →0 min {m, n} → ∞,
2
i.e. (vn )n∈N Cauchy. Since X is Hilbert and V is closed, there exists a v ∈ X such that
vn → v ∈ V . We then have:

∥x − v∥ = x − lim vn
n→∞
= lim ∥x − vn ∥ = d,
n→∞

since we only have continuous algebraic operations. So v is a minimizer. The proof of uniqueness
is omitted (and is left as an exercise to the reader ). ,
(ii) We first prove existence. Define v ⊥ := x − v, where v is any minimizer from (i). Then we have
x = v + v ⊥ and v ∈ V ✓. We now need to show that

x − v ∈ V ⊥. (∗)

To prove (∗), we use small variations of v:

ṽ := v + εw w ∈ V, ε ∈ R.

Then, f (ε) := ∥x − (v + εw)∥2 is minimal for ε = 0 and we get:

0 = f ′ (ε) |ε=0
d
= (∥x − v∥2 − −ε ⟨x − v, w⟩ − ε ⟨w, x − v⟩ +ε2 ∥w∥2 )|ε=0
dε | {z }
=−2 Re (ε⟨x−v,w⟩)
 
2
= 0 − 2 Re (⟨x − v, w⟩) + 2ε ∥w∥ |ε=0
= −2 Re ⟨x − v, w⟩ ,

so Re ⟨x − v, w⟩ = 0 ∀ w ∈ V .
If K = R, we just get x − v ∈ V ⊥ . If K = C, we use above property with iw instead of w:

0 = Re ⟨x − v, iw⟩
 
1
= Re ⟨x − v, w⟩
i
= Im ⟨x − v, w⟩ ∀ w ∈ V,

where we use that Re 1i (a + ib) = Re (b − ia) = b = Im (a + bi). Now with this, we get

0 = ⟨x − v, w⟩ ∀ w ∈ V,

i.e. x − v ∈ V ⊥ . Now we come to uniqueness. Suppose

⊥ ⊥
x = v + v⊥ = v′ + v′ v, v ′ ∈ V, v ⊥ , v ′ ∈ V ⊥ .

Then it follows that:

   
⊥
0 = v + v⊥ − v′ + v′

 ⊥

= v − v′ + v⊥ − v′ v − v′, · (∗∗)
| {z } | {z }
∈V
∈V ⊥
* +
′ ′ ′ ⊥ ′⊥
=⇒ 0 = v − v , v − v + v − v , v − v
| {z } | {z } | {z }
⊥ ⊥ ∈V ∈V
=∥v−v ′ ∥2

28
 2
=⇒ 0 = v − v ′ +0
′
=⇒ v = v .

If we now substitute that into (∗∗), we get v ⊥ = v ′ ⊥ .

Definition 3.4. A normed vector space X is called separable, if there is a countable set D that is
dense in X (i.e. D = X).
Examples.
ˆ lp (1 ≤ p < ∞) is separable, D is all elements of d with rational coefficients.

ˆ l∞ is not separable, as seen in exercise 11.

ˆ C ([0, 1]) is separable, as seen in exercise 10, D is all polynomials with rational coefficients.

ˆ Lp ([0, 1]) 1 ≤ p < ∞ is separable, C ([0, 1]) (with Lp norm) is dense in Lp and so is it’s D.

ˆ L∞ ([0, 1]) is not separable, we consider A = 1[0,a] a∈[0,1] which is uncountable but 1[0,a]−1[0,a′ ]

=
L∞
1 for a ̸= a′ .
Definition 3.5. Let X be a Hilbert space and B = {en }n∈I a finite (I = {1, . . . , n}) or countable
(I = N) set. B is called Orthonormal Basis of X if
(
1, i = j
(i) B is orthonormal, i.e. ⟨ei , ej ⟩ = .
0, i ̸= j

(ii) Span B = X
Theorem 3.4. Any countable orthonormal basis {e1 , e2 , . . . } of a Hilbert space X is a Schauder basis,
i.e. for all x ∈ X there exist αi ∈ K such that:
n
X
x = lim αi ei . (∗)
n→∞
i=1

Moreover, the αi are unique and given by the formula

αi = ⟨x, ei ⟩ i = N. (∗∗)

Proof. See exercise 16.

Proof of (∗∗) using (∗). From (∗), we get:

n
* +
X
⟨x, ek ⟩ = lim αi ei , ek
n→∞
i=1
* n +
X
= lim αi ei , ek
⟨·,·⟩ cts. n→∞
i=1
n
X
= lim αi ⟨ei , ek ⟩
⟨·,·⟩ lin. n→∞ | {z }
i=1
1, i = k

= , since B orthonormal
0, else

(
0, n<k
= lim = αk
n→∞ αk , n ≥ k

29
Example (Fourier Series, important). Let X = L2 ([−π, π]) and f ∈ L2 ([−π, π]) the
o equivalence class
n
π 2
of functions f˜ ∈ L
f2 ([−π, π]) = g : [−π, π] → C : gmeasurable, i.e.
−π |g| < ∞ . Let furthermore
eiξx
ek be the equivalence class of ẽk (x) := √
2π
for ξ ∈ Z.

Re eiξx
1
ξ=0

ξ=1
x
−π π

ξ=2
−1

Figure 13: Real parts of the ek

Now, B = {ek }k∈Z is orthonormal, since

 π
⟨ek , el ⟩ = ẽk (x) ẽl (x) dx
−π
 π
1
= eikx{zeikx} dx
2π −π
|
ei(k−l)x
(
1, k=l
= 1 1 i(k−l)x |π
.
2π i(k−l) e −π = 0, k ̸= l

Claim: B is an orthonormal basis of L2 ([−π, π]).

Proof. Proof of the claim We need to show that Span B = L2 ([−π, π]). This is not trivial
but follows from classical results about Fourier series as follows:
Classical fact: Let f˜: [−π, π] → C be Lipschitz continuous with f˜ (−π) = f˜ (π). The
periodical Fourier sum of f˜
Xn
fˆk eikx ,
k=−n

where  π
1
fˆk = f˜e−ikx dx
2π −π

are the Fourier coefficients, converges uniformly to f˜ as n → ∞.

Hence Span B is dense in Lipschitz continuous functions with f˜ (−π) = f˜ (π) with respect
to supremum norm and it follows that the same is true with respect to th L2 norm.
From measure theory we know that Lipschitz functions with f˜ (−π) = f˜ (π) or even
C0∞ ((−π, π)) = {g : (−π, π) → C : g infinitely differentiable with compact support in (−π, π)}
are dense in L2 ([−π, π]), so Span B is dense in L2 ([−π, π]).

Corollary 3.1 (of theorem 3.4, L2 theory of Fourier series). For all f ∈ L2 ([−π, π]), we have:
n
X
f = lim ⟨f, ek ⟩ ek ,
n→∞
k=−n

with the lim being in the L2 sense

30
Remark (Interpretation as Fourier series).
 π −ikx
e √
⟨f, ek ⟩ = f˜ (x) √ dx = 2π · fˆk f˜ ∈ f a representative
−π 2π
1
ek = [ẽk ], ẽk (x) = √ eikx
2π
n
X Xn
=⇒ ⟨f, ek ⟩ ek = [sn ], sn (x) = fˆk eikx partial Fourier sum.
k=−n k=−n

For f˜ ∈ L
f2 ([−π, π]), the partial Fourier sum sn (x) satisfies
 π 2
f˜ − sn →0 n → inf ty.
−π

In sloppy notation, we write: f˜ = ∞ ˆ ikx .

P
k=−∞ fk e

But when does there exist an orthonormal basis?

We have the obvious necessary condition that X is separable (since finite linear combinations of the
basis vectors with rational coefficients are dense)

Theorem 3.5 (Existence of orthonormal basis). Let X be a separable Hilbert space. Then there exists
an orthonormal basis X.

Proof. Use Schmidt orthonormalization.

Since X is separable, there exists U = {u1 , u2 , . . . } with {u1 , u2 , . . . } = X. Without loss of generality,
u1 , u2 , . . . ̸= 0.
We normalize e1 := ∥uu11 ∥ . If u2 ̸∈ Span {u1 }, by theorem 3.3, we can find v2 ∈ Span {u1 } , v2 ⊥ ∈
v2 ⊥
Span {u1 }⊥ such that u2 = v2 + v2 ⊥ and we take e2 := .
∥ 2⊥∥
v
We then get Span {u1 , u2 } = Span {e1 , e2 }.
If u2 ∈ Span {u1 }, consider U \ {u2 }.
We iterate this: if un ̸∈ Span {u1 , . . . , en−1 }, we take un = vn + vn ⊥ with vn ∈ Span {u1 , . . . , un−1 }
⊥
and vn ⊥ ∈ Span {u1 , . . . , un−1 }⊥ . Then we take en := vvn ⊥ .
∥n ∥
Then {e1 , e2 , .S . . } are a orthonormal basis and Span {e1 , . . . , eSn } = Span {u1 , . . . , un } ∀ n.
Furthermore, n∈N {u1 , . . . , un } is dense in X and therefore n∈N Span {u1 , . . . , un } as well.
So {e1 , e2 , . . . } is a orthonormal basis of X.

Definition 3.6. Two Hilbert spaces X, X ′ are called isomorphic (notation X ∼

= X), if there exists
i : X → X ′ such that

(i) i linear bijection

(ii) ⟨x, y⟩X = ⟨i (x) , i (y)⟩X ∀ x, y ∈ X

Remark (Classification of separable Hilbert spaces).

ˆ dim X < ∞. Then there exists an Orthonormal Basis {e1 , . . . , en } for some n ∈ N and X ∼
= Kn
with  
⟨x, e1 ⟩
 ⟨x, e2 ⟩ 
i (x) =  .  ∈ Kn .
 
 . .
⟨x, en ⟩

ˆ dim X = ∞. Then there exists an Orthonormal Basis {ek }k∈N and X ∼

= l2 with i (x) =
2
(⟨x, e1 ⟩ , ⟨x, e2 ⟩ , . . . ) ∈ l .

31
Example. We consider L2 [−π, π] ∼
= l2 and have

I (f ) = (⟨f, e0 ⟩ , ⟨f, e1 ⟩ , ⟨f, e−1 ⟩ , ⟨f, e2 ⟩ , . . . ) a vector of Fourier coefficients

√  
= 2π fˆ0 , fˆ1 , fˆ−1 , fˆ2 , . . .
X X 2
∥f ∥2L2 = |⟨f, ek ⟩|2 = 2π fˆk ,
k∈Z k∈Z

which is called Perseval’s equation.

4 Examples of Linear Functionals and Linear Operators

Definition 4.1. A map A : X → Y between K-vector spaces X and Y is called linear if
A λx + µx′ = λA (x) + µA x′ ∀ x, x′ ∈ X, λ, µ ∈ K.

In Functional Analysis, linear maps are called linear operators. If the target space Y = K, they are
called linear functionals.
(Somewhat misleadingly, the latter convention underlies the name “Functional Analysis”.)

Convention: For linear maps, we usually write A· instead of A (·).

Recall the following notations from Linear Algebra for linear maps A : X → Y between vector spaces:
ˆ injective (Engl.: one-to-one), i.e. x ̸= x′ =⇒ Ax ̸= Ax′
ˆ surjective (Engl.: onto)
ˆ bijective or invertible, i.e. injective and surjective
ˆ kernel
ˆ range
Recall furthermore the origin of notations of injective and surjective from solving linear equations:
ˆ A is injective if and only if ∀ b ∈ Y the equation Ax = b has at most one solution

ˆ A is surjective if and only if ∀ b ∈ Y the equation Ax = b has at least one solution

In infinite dimensions, we have a new effect: For linear operators A : X → X, injective and surjective
aren’t equivalent any more (for A : Kn → Kn linear, we had A injective ⇐⇒ A surjective ⇐⇒
det A ̸= 0).
Example 4.1. X = l2
a) Right shift S + (a1 , a2 , a3 , . . . ) = (0, a1 , a2 , a3 , . . . ). This is injective, but not surjective, since
(1, 0, 0, . . . ) ̸∈ Ran S + .
b) Left shift S − (a1 , a2 , a3 , . . . ) = (a2 , a3 , . . . ). This is surjective since ∀ x ∈ l2 , S − (0, x1 , x2 , x3 , . . . ) =
x, but not injective, since S − (1, 0, 0, . . . ) = S − (2, 0, 0, . . . ).


Definition 4.2. A map A : X → Y between normed vector spaces X, ∥·∥ and (Y, ∥·∥∗ ) is called
continuous if:
xn → x =⇒ A (xn ) → A (x) “sequential definition of continuity”
(i.e. ∥xn − x∥ → 0) (i.e. ∥A (xn ) − A (x)∥∗ → 0)

or equivalently
V ⊆Y =⇒ A−1 (V ) “topological definition of continuity”
open (in Y ) open (in X)

The new effect in infinite dimensions is that linear maps can be discontinuous (for A : Kn → Kn , we
had linear =⇒ continuous).

32
Example 4.2 (Derivative Operator, important).
∞ = {f : R → R : f 2π-periodic, f infinitely often differentiable} , A = d i.e. (Au) (x) =
X = Cper dx
′
u (x). Then A is linear.
We consider un (x) := n1 sin (nx) with ∥u∥ := sup[0,2π] |u (x)| as a norm on X.
We then have ∥un − 0∥ = n1 → 0, so un → 0.
We apply A: (Aun ) (x) = u′n (x) = cos (nx) and get ∥Aun − A0∥ = 1 ̸→ 0, so Aun ̸→ A0. Hence A is
not continuous.

Example 4.3 (for illustration). X = d = {(a1 , a2 , . . . ) : ai ∈ R, only finitely many ai are non zero}
with lp norm. We define A(a1 , a2 , a3 , . . . ) = (1 · a1 , 2 · a2 , 3 · a3 ) and have:
1
A(0, 0, . . . , 0, , 0, . . . ) = 0, 0, . . . , 0, |{z}1 , 0, . . . .
n
|{z} n-th comp.
n-th comp. | {z }
| {z } ̸→0
→0

As a matrix with respect to the standard basis {(1, 0, 0, . . . ), (0, 1, 0, . . . ), (0, 0, 1, . . . ), . . . }, we have:
 
1
 2 
 
A=
 3 
 “∞ × ∞-matrix”

 4 

..
.


Lemma 4.1. Let X, ∥·∥ and (Y, ∥·∥∗ ) be normed vector spaces and A : X → Y linear. The following
are equivalent:

(i) A is continuous.

(ii) There exists an M ≥ 0 such that ∥Ax∥∗ ≤ M ∥x∥ for all x ∈ X.

Proof.

“(ii) =⇒ (i)”: We have:

xn → x ⇐⇒ ∥xn − x∥ → 0
=⇒ ∥Axn − Ax∥∗ = ∥A(xn − x)∥∗ ≤ M ∥xn − x∥ → 0,

i.e. Axn → Ax.

“(i) =⇒ (ii)”: We prove this indirectly. Suppose, there is no M such that ∥Ax∥∗ ≤ M ∥x∥ ∀ x ∈ X.
Hence ∀ n ∈ N there exists an xn ∈ X such that

∥Axn ∥∗ ≥ n ∥xn ∥ . (∗)

Now the idea is to use this to find an abstract analogon to 0, . . . , 0, n1 , 0, . . . , 0 in

previous example.
We consider x′n := n∥x
xn
n∥
. Then we have ∥x′n ∥ = n1 → 0 and therefore x′n → 0.
But
xn 1
Axn′ ∗ = A = ∥Axn ∥ ≥ 1,
n ∥xn ∥ ∗ n ∥xn ∥ | {z }∗
≥n∥xn ∥

so Ax′n ̸→ A0.

33
Example 4.4 (Antiderivative (Ger.: Stammfunktion)). X = C ([a, b]) with standard norm ∥f ∥ =
supx∈[a,b] |f (x)|. We define
 x
(Af ) (x) := f (s) ds.
0
We then have A : X → X since the antiderivative is differentiable and in particular continuous and A
linear. We check continuity via lemma 4.1:
 x
∥Af ∥ = sup |(Ax) (x)| = sup f (s) ds ≤ (b − a) ∥f ∥ ,
x x a
| {z }
≤(x−a) sups∈[a,b] |f (s)|

so A is indeed continuous.

Example 4.5 (Derivative, viewed as a linear operator between different spaces). X = C 1 ([a, b]) the
space of continuously differentiable functions with norm ∥f ∥C 1 = supx∈[a,b] |f (x)| + supx∈[a,b]|f ′ (x)| ,
d
Y = C ([a, b]) the space of continuous functions with norm ∥f ∥C = supx∈[a,b]|f (x)| . We define A = dx ,
′
i.e. (Af ) (x) = f (x) and get A : X → Y , since the derivative of a continuously differentiable function
is by definition continuous, and A linear. We again use lemma 4.1:

∥Af ∥C = sup |(Af ) (x)| = sup ≤ ∥f ∥C 1

c∈[a,b] x∈[a,b]|f ′ (x)|

and conclude that A is continuous.

Definition 4.3. Let (X, ∥·∥X ), (Y, ∥·∥Y ) be normed vector spaces over K = R or C. We define the
following spaces:

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

X ∗ := {l : X → K : l linear and continuous} .

The elements of L (X, Y ) are called continuous linear operators. The elements of X ∗ are called
continuous linear functionals. X is called the dual space of X.

Obvious properties of L (X, Y ) and X ∗

(i) They are vector spaces over K under pointwise addition and pointwise multiplication with scalars:

(A + B) (x) := Ax + Bx x∈X
(λA) (x) := λAx λ ∈ K.

(ii) L (X, X) is an algebra under composition:

(AB) (x) := A (B (x)) x ∈ X.

 
x1
More generally, if B ∈ L (X, Y ) and A ∈ L (Y, Z), then AB ∈ L (X, Z). (If X = Kn , x =  ... ,
 

xn
n×n
P
A∈K
P acting as (Ax)i = j Aij xj , then AB is the usual matrix multiplication (AB)ik =
A B
j ij jk .)

Theorem 4.1 (Natural Norm and Banach Space Property for L (X, Y )). Let (X, ∥·∥X ) and (Y, ∥·∥Y )
be normed vector spaces.
∥Ax∥Y
a) ∥L (X) , Y ∥L(X,Y ) := supx̸=0 ∥x∥X is a norm on L (X, Y ) (called operator norm).

34
b) If Y is a Banach space, then, regardless of whether X is Banach, L (X, Y ) with the operator norm
is Banach. In particular, the dual space X ∗ of X with the norm
|L (x)|
∥l∥X ∗ = sup
x̸=0 ∥x∥X
is Banach.

Proof.
a) First of all, we need to show that the sup is less than infinity. By lemma 4.1, we get:

∃ M ≥ 0 : ∥Ax∥Y ≤ M ∥x∥X x ∈ X \ {0}

∥Ax∥Y
=⇒ ≤M
∥x∥X
∥Ax∥Y
=⇒ sup ≤ M < ∞.
x̸=0 ∥x∥ X

We show the norm axioms:

∥Ax∥Y
Positivity: supx̸=0 ∥x∥X = 0 =⇒ Ax = 0 ∀ x ̸= 0 =⇒ A = 0.
∥λAx∥Y ∥Ax∥Y
Homogeneity: supx̸=0 ∥x∥X = |λ| supx̸=0 ∥x∥X ✓
Triangle Inequality:
∥Ax + Bx∥Y
∥A + B∥L(X,Y ) = sup
x̸=0 ∥x∥X
 
∥Ax∥Y ∥Bx∥Y
≤ sup +
x̸=0 ∥x∥X ∥x∥X
∥Bx′ ∥Y
 
∥Ax∥Y
≤ sup +
x̸=0,x′ ̸=0 ∥x∥X ∥x′ ∥X
= ∥A∥L(X,Y ) + ∥B∥L(X,Y ) ,

where in the penultimate transformation we just increased the size of the

set for the sup.
b) Completeness is analogous to the proof that (Cb (Ω) , ∥·∥∞ ) is complete (Theorem 1.3) only with
the different factor ∥x∥1 .
X

5 Open Mapping Theorem

5.1 Motivation
Definition 5.1 (Inverse). Let A : X → Y be a linear map between normed vector spaces. If A is
invertible (i.e. injective and surjective), the map

Y → X, b 7→ unique solution of Ax = b

is called the inverse of A, notation: A−1 .

Elementary fact: A−1 : Y → X is again linear.

Proof. Let x1 = A−1 b1 , x2 = A−1 b2 , so Ax1 = b1 , Ax2 = b2 . We consider:

A(λx1 + µx2 ) = λAx1 + µAx2 = λb1 + µb2 .

Now we apply A−1 from the left and get linearity.

35
Natural question (pure mathematician): Is the property of continuity also preserved under tak-
ing the inverse?
Natural question (applied mathematician): Does the solution of the equation Ax = b (“predic-
tion”) depend continuously on the “data” b?

5.2 The Theorem

Theorem 5.1. Let X, Y be Banach spaces and T : X → Y an invertible and continuous linear map.
Then T −1 is continuous.

Example. d = a = (ai )i∈N : ai ∈ R, only finitely many ai ̸= 0 with norm ∥a∥= supi |ai |. We con-
sider the function  
1
 1 
2
 : d → d,
 
 1
 3 
..
.
which is continuous and linear. Its inverse
 
1
 2 
:d→d
 

 3 
..
.

however is not continuous (use lemma 4.1).

ˆ This does not contradict the theorem, since d is not Banach.

ˆ It shows that the assumption of X, Y being Banach cannot be omitted

ˆ Shows that the theorem is very remarkable, since the extra assumption is on the space and not
on the operator.

Theorem 5.2 (Open Mapping Theorem). Let t : X → Y be a continuous linear onto map between
Banach spaces. Then T is open, i.e. the images of open sets are open.

Proof of theorem 5.1 using 5.2. Suppose in addition T is invertible, so T −1 : Y → X exists. Then:

−1
T −1 continuous ⇐⇒ T −1 (U) open ∀ U open ⇐= T continuous, linear, onto
theorem 5.2

5.3 The Proof

Lemma 5.1 (Baire Category Theorem). A complete metric space (e.g. a Banach space) is not the
countable union of nowhere dense sets.

Definition 5.2. Let X be a metric space, A ⊂ X. A is called nowhere dense if A does not contain
an open set.

Examples.

1. A = Z, X = R

2. A a one-dimensional subspace of R3

3. Q ⊂ R is dense, Q2 ⊂ R2 is dense.

36
 T : X → Y onto T : X → Y continuous

Baire Category Theorem Y complete

1. T (Bε (0)) contains an open set

translation argument

2. T (Bε (0)) contains a ball around zero

X complete

3. T (B2ε (0)) contains a ball around zero

translation argument

4. ∀ Ω open, T (Ω) open

Figure 14: Outline of the proof

Terminology: X is called
ˆ of first category if it is a countable union of nowhere dense sets (“small”)

ˆ of second category, otherwise. (“big”)

The Baire Category Theorem says that Banach spaces are of second category.
Example. There exists no norm on P ([0, 1]) = {polynomials p : [0, 1] → R} such that P ([0, 1]) is
Banach. (Generalizes find from Exercise 10 that (P ([0, 1]) , ∥·∥∞ ) not Banach; you showed that limits
of Cauchy sequences in P ([0, 1]) give any continuous
S function.)
Mn := {polynomials with deg ≤ n} , P ([0, 1]) = n∈N Mn . Claim: Mn is nowhere dense
Proof. Mn is finite dimensional and therefore closed. Suppose Mn = Mn ⊇ Bδ (pn ) for
some δ > 0 and some pn ∈ Mn . Then we have:
(
δ en+1 ∈ Bδ (en )
en+1 (x) := xn+1 ̸∈ Mn pn + .
2 ∥en+1 ∥ ̸∈ Mn

This is a contradiction, so with Baire follows: P ([0, 1]) is not Banach.

is a complete metric space, X = ∞
S
Proof of Baire, indirect. Suppose x S n=1 An , An nowhere dense.
Now the idea is to construct an x ̸∈ ∞ n=1 A n for a contradiction.
Mn := An closed, nowhere dense, X = ∞
S
n=1 Mn .
Claim: X \ M1 is open and dense.

37
It is open since M1 is closed. It is dense, since otherwise there exists an x such that dist (x, X \ M1 ) > 0
which implies that there exists ε > 0 such that Bε (x) ∩ X \ M1 = ∅, i.e. Bε (x) ⊂ M1 which is a
contradiction to M1 is nowhere dense.

Step 1 Since X \ M1 is open and non-empty, there exists a closed ball S1 ⊂ X \ M1 , S1 = Br1 =
{x ∈ X : d (x1 , x) ≤ r1 }, r1 ≤ 21

Step 2 X \ (M1 ∪ M2 ) is open and dense, therefore there exists a closed ball S2 ⊂ S1 , S2 ⊂ X \

2
(M1 ∪ M2 ), S2 = Br2 (x2 ) , r2 ≤ 21
..
.

n
This implies that there exists a sequence {Sn } of nested closed balls with radii 0 < rn ≤ 21 and
Sn ⊂ X \ (M1 ∪ · · · ∪ Mn ).
The sequence of centres {xn } is Cauchy which implies, since X is complete, that limn→∞ xn = x∞ ∈ X.
Claim: x∞ ∈ Sn ∀ n
S
Proof. xm ∈ Sn ∀ m > n, xm → x∞ , Sn closed. Hence x∞ ̸∈ Mn ∀ n. But n∈N Mn = X
which is a contradiction.

5.3.1 Proof of Open Mapping Theorem

We have to show that for all balls Bε (0), T (Bε (0)) ⊇ V0 for some open V0 .
S
Proof. Let Bε (0) := {x ∈ X : ∥x∥ < ε} , ε > 0 be arbitrary. We have X = n∈N nBε (0)
(since x ∈ nBε (0) = Bnε (0) as soon as n > ∥x∥
ε ).
This implies that
!
[ [
Y = T (X) = T nBε (0) = T (nBε (0)) .
T onto
n∈N n∈N

With Y being complete and Baire it follows that not all T (nBε (0)) are nowhere dense,
i.e. there exists an n such that:

T (nBε (0)) ⊇ Ve0 for some open Ve0

1 1
=⇒ T (Bε (0)) = T (nBε (0)) ⊇ Ve0 =: V0 open
T lin. n n

5.3.2 Translation Argument

We need to prove: ∀ ε > 0 we have T (B2ε (0)) ⊇ some ball around 0.

Proof. We pick y0 ∈ V0 ∩ T (Bε (0)) and x0 ∈ Bε (0) such that T (x0 ) = y0 and get:

T (B2ε (0)) ⊇ T (Bε (0) − x0 ) = T (Bε ) (0) − T (x0 ) ⊇ V0 − y0

T lin.
⊇ Bδ (y0 ) − y0 = Bδ (0)
V0 open

for some δ > 0.

38
5.3.3 Cauchy Sequence Argument
We need to show: T (B2ε (0)) contains a ball around 0.

Proof. Let Bε be a ball around 0 in X of radius ε and Bη′ be a ball in Y of radius η.

Consider Bεi , εi := 2−i ε0 , i = 0, 1, 2, . . .
From 5.3.2, it follows that there exist ηi > 0 with ηi → 0 such that T (Bεi ) ⊇ Bη′ i
Let y ∈ Bη′ 0 . We claim: There exists x ∈ B2ε0 such that T x = y.

y ∈ T (Bε0 ) =⇒ ∃ x0 ∈ Bε0 : ∥y − T x0 ∥ < η1 =⇒ y − T x0 ∈ Bη′ 1

y − T x0 ∈ T (Bε1 ) =⇒ ∃ x1 ∈ Bε1 : ∥y − (T x0 + T x1 )∥ < η2
..
.
=⇒ ∃ xn ∈ Bεn : ∥y − T (x0 + · · · + xn )∥ < ηn+1 n = 0, 1, 2, . . .
(∗)
Pn
Now \ i=0 xi is Cauchy, since
n
X m
X n
X n
X
xi − xi = xi ≤ ∥xi ∥ = ε0 · 2−m → 0 m → ∞.
m>n |{z}
i=0 i=0 i=m+1 i=m+1
≤ε0 ·2−i

Since X is complete, there exists an x such that x = limn→∞ (x1 + · · · + xn ). Because T

is continuous, this holds in Y as well and we get T x = limn→∞ (T x1 , . . . , T xn )
Now, by (∗) we know that this converges to y and we get:
∞
X ∞
X
∥x∥ ≤ ∥xi ∥ ≤ ε0 · 2−i = 2ε0
i=0 i=0

which concludes the proof.

5.3.4 Translation Argument

We need to prove: Ω open =⇒ T (Ω) open.
Let y = T x be an arbitrary point in T (Ω). Then we get:

T (Ω) T (Bε (x)) = T (Bε (0) + x) = T (Bε (0)) + T x ⊇ Bδ′ (0) + T x = Bδ′ (T x) ,
Ω open, i.e. T lin. 5.3.3
Ω⊇Bε (x) for some ε>0

so T (X) is open.

Example (Sensitivity Analysis with respect to “Data”). We consider a linear equation Ax = b. We

can find the exact solution for x as a function of b by inverting A:

x = A−1 b.

Combine theorem 5.1 and the continuity criterion from lemma 4.1. Due to uncertainties in b, we
measured b′ instead of b. Then we get x′ as a solution with respect to perturbed data. We can find
an abstract “sensitivity estimate” with respect to the data:

x − x′ = A−1 b − b′

=⇒ x − x′ ≤ M b − b′

for some constant M , since the prediction error x − x′ depends continuously on the measurement error
b − b′ . Of course, in practice, one would want to know M explicitly.

39
5.4 Related Results
a) Pointwise vs. bounded families of maps

Definition 5.3. Let X, Y be normed vector spaces, I an index set and Ti ∈ L (X, Y ) ∀ i ∈ I. The
Ti are called

ˆ pointwise bounded if for all x, supi∈I ∥Ti x∥ < ∞

ˆ uniformly bounded if
∥Ti x∥
sup sup <∞
i∈I x̸=0 ∥x∥
| {z }
=∥Ti ∥L(X,Y )

Example. X = Y = d with norm ∥·∥∞ . Ti x = (0, . . . , 0, ixi , 0, . . . , 0), where Ti has zero entries
everywhere besides (Ti )i,i = i.

ˆ Ti pointwise bounded since for all x ∈ d, there exists N (x) such that

xj = 0 ∀ j > N (x)
=⇒ Ti x = 0 ∀ i > N (x)
=⇒ sup ∥Ti x∥ = max {|ixi | : 1 ≤ i ≤ N (x)} < ∞
i

ˆ Ti is not uniformly bounded since ∥Ti ∥L(d,d) = i, so supi ∥Ti ∥L(d,d) = ∞

Theorem 5.3 (Banach-Steinhaus Theorem or Principle of Uniform Boundedness). Let X be

Banach, Y a normed vector space, I an index set and Ti ∈ L (X, Y ) ∀ i ∈ I. Then

Ti pointwise bounded =⇒ Ti uniformly bounded.

Proof. Uses Baire, loosely similar to proof of Open Mapping Theorem, details omitted.

Definition 5.4. A linear operator A : X → Y between normed vector spaces X and Y is called
closed if its graph
graph A = {(x, y) ∈ X × Y : y = Ax}
is closed, or equivalently, if
xj →x
o
Axj →y =⇒ Ax = y

Example 5.1. A continuous =⇒ A closed

n
Axj →y for some y
Proof. xj → x =⇒ Axj → Ax i.e. y=Ax

Example 5.2 (A non-closed Operator). Let X = Y = d and ei = (0, . . . , 0, 1, 0, . . . ) the sequence

with only a 1 as i-th entry. We consider Aei = ie1 for
 
1 2 3 ...
0 0 0 . . .
A = 0 0 0 . . . .
 
 
.. .. ..
. . .
n
∈graph A
We define x(i) = 1i e1 → 0 and get Ax(i) = i 1i e1 → e1 ̸= A0, so (0, e1 ) ∈graph
/ A
.

Theorem 5.4 (Closed graph theorem). Let X, Y be Banach and A : X → Y a closed linear
operator. Then A is continuous.

40
 Remark. This gives us a simplified strategy to prove that a linear operator is continuous:
Before, we needed to show

xj → x for some x =⇒ Axj → y for some y ∧ Ax = y.

Now we can prove a simpler statement:

xj → x for some x ∧ Axj → y for some y =⇒ Ax = y.

Proof. See class.

Example 5.3 (important).

Let X = C ∞ ([0, 1]) = {f : [0, 1] → R : f is n times continuously differentiable ∀ n ∈ N}. We con-
sider Af = f ′ , i.e. A = dx
d
. No matter which norm we choose on C ∞ ([0, 1]), A is not continuous
(→ class).
Take e.g. ∥·∥∞ . Claim: A is closed.
′
Proof. Suppose f = f (n) → f uniformly and Af (n) = f (n) → g uniformly, then f is
differentiable and f ′ = g. This is because:
 x  x
(n) (n) (n) (n)
f (x) = f (0) + f →f (0) + g = f (x)
0 0

With the Fundamental Theorem of Calculus follows: f is differentiable and f ′ = g.

6 Invertibility, Notion of Spectrum

Questions
1. How do I investigate whether a linear operator is invertible, i.e. whether the equation Ax = b
has a unique solution for all b.

2. What is the right generalization of the notion of Eigenvalues of matrices? (Related to whether
A − λI is invertible)

We basically have three approaches:

ˆ Perturbation Theory (simplest, already useful)

ˆ Index Theory
)
more general but more complicated
ˆ Duality

6.1 Invertibility
Theorem 6.1 (Invertibility by Perturbation Theory). Let X, Y be Banach, T ∈ L (X, Y ) invertible,
S ∈ L (X, Y ) and ∥T − S∥L(X,Y ) < ∥T −1 ∥1 .
L(X,Y )
Then S is invertible.

This has the following abstract consequence: If T is invertible, there exists a δ > 0 such that S is
invertible for all ∥T − S∥ < δ, i.e. the set of invertible operators in L (X, Y ) is open.

Lemma 6.1 (Neumann Series).  

1
Let X be Banach, A : X → X continuous and linear and ∥I − A∥L(X,X) < 1 = ∥I −1 ∥
.
Then A is invertible with the explicit inverse
 
−1 2 n
A = lim I + A + A + · · · + A , A
e e e e := I − A. (∗)
n→∞

41
 In the special case X = R we get A ∈ R, A
e = 1 − A ∈ R and:
1 1 1 e2 + . . . .
= = = 1+A
e+A (Geometric Series)
A 1 − (1 − A) 1−Ae |Ae|<1

So (∗) is a “Geometric Series” for operators instead of numbers. Multiplication of numbers corresponds
to composition of operators, addition corresponds to addition of operators.
Lemma 6.2 (Submultiplicativity of the Operator Norm). Let X, Y, Z be normed vector spaces,
B A
A ∈ L (Y, Z) and B ∈ L (X, Y ), resulting in X → Y → Z. We then have
∥AB∥L(X,Z) ≤ ∥A∥L(Y,Z) ∥B∥L(X,Y ) .
If X = R, C, we even have equality.

Proof. Recall the Operator Norm:

∥Bx∥Y
∥B∥L(X,Y ) = sup .
x̸=0 ∥x∥X
So we have:
∥ABx∥ ∥ABx∥ ∥Bx∥
∥AB∥L(X,Z) = sup = sup · ≤ ∥A∥ · ∥B∥ .
x̸=0 ∥x∥ x̸=0 ∥Bx∥ ∥x∥
Bx̸=0 | {z } | {z }
∥Ay∥ ≤∥B∥
≤supy̸=0 ∥y∥
=∥A∥

Proof that the Series (∗) is Convergent. We look at the partial sums
Sn := I + A e2 + · · · + A
e+A en .

With m > n, we have:

m
X
Sm − Sn = ek
A
k=n+1
m m
X X k n+1  
=⇒ ∥Sm − Sn ∥ ≤ ek
A ≤ A
e = Ae 1 + ∥A∥ + A em−(n+1) ,
tri. lemma 6.1
k=n+1 k=n+1 | {z } | {z }
e∥<1)
→0 (∥A ≤ 1 e (geo. series)
1−∥A∥

so {Sn }n∈N is Cauchy.

 so is L (X, X) which implies that {Sn } is convergent, i.e. the limit exists.
Since X isBanach,
Claim: S I − A = I
e

Proof. We use a telescope sum idea:

    
Sn I − A e = I +A e+ ··· + A en I − A
e
   
= I +A e + · · · + An
e − A e+A e2 + · · · + A
en + A
en+1 = I − A
en+1
 
=⇒ S I − A e = I,

n+1
en+1 ≤ A
since A e → 0.


Analogously, we show I − A
e S = I, so we get:
 −1
S = I −A
e = (I − (I − A))−1 = A−1 .

42
Proof of Theorem 6.1. The idea is to find an expression for S to which we can apply a Neumann
Series:

S = T − (T − S)
T − T −1 (T − S)


= T invertible .
| {z }
invertible if ∥T −1 (T −S)∥<1 by lemma 6.1

But we have
T −1 (T − S) ≤ T −1 ∥T − S∥ < 1,
lemma 6.2
1
since ∥T − S∥L(X,Y ) < ∥T −1 ∥L(X,Y )
.

2
Remark. The proof shows that S −1 = I + T −1 (T − S) + T −1 (T − S) + . . .

6.2 Analogue of Eigenvalues

Definition 6.1. Let X be a Banach space and A ∈ L (X, X). The spectrum of A is defined as

spec A = {λ ∈ C : A − λI not invertible} .

C \ spec A = ρ (A) is called the resolvent set of A.

For λ ∈ ρ (A), the operator (A − λI)−1 is called the resolvent of A. The subset

pt spec A = {λ ∈ C : A − λI not one-on-one}

is called Point Spectrum and its elements are called Eigenvalues.

We have λ ∈ pt spec A =⇒ ∃ x ̸= 0 : Ax = λx.

Examples.

1. X = l2 , A = S− the left shift, i.e. S− (x1 , x2 , x3 , . . . ) = (x2 , x3 , . . . ).

λ = 21 is an Eigenvalue of S− , since
! ! !
1 1 2 1 1 2 1 3 1 1 2
       
1
S− 1, , ,... = , , ,... = 1, , ,... .
2 2 2 2 2 2 2 2
| {z }
∈l2

/ spec (S− ), since 2I − S− = 2 I − 12 S− , 1 1 1

λ=2∈ 2 S− = 2 ∥S− ∥ = 2 and we can apply lemma
6.1.

2. Example where λ ∈ spec but λ is not an Eigenvalue. X = l2 , S+ (X, X) the right shift
S+ (x1 , x2 , x3 , . . . ) = (0, x1 , x2 , x3 , . . . ).
λ = 0 ∈ spec+ but not an Eigenvalue, since S+ − 0 · I is not onto but one-on-one, since
(1, 0, 0, . . . ) ∈
/ Ran S+ .

7 More about Invertibility: Index Theory

Goal: Introduce a quantity which “measures” the amount of non-invertibility, called index.

43
7.1 Definition of the Index, Main Theorem
Revision from Linear Algebra
Definition. Let V be a vector space, U a subspace, and v ∈ V .

[V ] := {x + u : u ∈ U }

is the equivalence class of v ∈ V with respect to U . We then get the set of equivalence classes
{[v] : v ∈ V } =: V /U .
V /U endowed with addition
[v] + v ′ := v + v ′ v, v ′ ∈ V
   

and scalar multiplication

λ[v] := [λv] λ ∈ K, v ∈ V.
of representatives is again a vector space and is called Quotient Space of V and U .

Definition. A subspace W ⊆ V is called complement of U ⊆ V if for all v ∈ V there exist unique

u ∈ U, w ∈ W such that
v = u + w.
If W is a complement of U , then

V /U ∼
= W. (isomorphism of vector spaces)

[v]
v u

w U
0

W ∼
= V /U

Figure 15: Illustration of a quotient space in R3

Remark. Complements, if they exist, are not unique.

Definition 7.1. Let T : X → Y be a linear operator between vector spaces.

{x ∈ X : T x = 0} =: ker T

is a well known subspace. Another “analogous” space is

Y /Ran T =: co ker T,

i.e. the quotient of Y and Ran T , called co-kernel of T .

ˆ = 0 ⇐⇒ T is injective



dim ker T ˆ measures amount of non-


injectivity
ˆ = 0 ⇐⇒ T is surjective



dim co ker T ˆ measures amount of non-


surjectivity

44
Definition 7.2. Let X, Y be Banach spaces and T ∈ L (X, Y ). T is called Fredholm operator if
(i) dim ker T < ∞
(ii) dim co ker T < ∞.
The number
dim ker T − dim co ker T =: ind (T ) ∈ Z
is called the index (or Fredholm index) of T . The set of Fredholm operators is denoted F (X, Y ).
Theorem 7.1 (Index Theory). Let X, Y be Banach spaces.
(i) For all T ∈ F (X, Y ), there exists a δ > 0 such that
T ∈ F (X, Y ) , S ∈ L (X, Y ) , ∥T − S∥ < δ =⇒ S ∈ F (X, Y ) .
In other words, F (X, Y ) is an open subset of L (X, Y ).
(ii) The index ind : F (X, Y ) → Z is locally constant, i.e. for all T ∈ F (X, Y ), there exists a δ > 0
such that
ind (T ) = ind (S) ∀ S ∈ L (X, Y ) with ∥T − S∥ < δ.

(iii) The index ind : (F (X, X) , ·) → (Z, +) is a semi-group homomorphism (where the semi-group
operator in F (X, X) is composition and the semi-group operator in Z is addition), i.e.
ind (T ◦ S) = ind T + ind S.

Example 7.1. X = Y = l2 , At (a1 , a2 , a3 , . . . ) = (ta1 , a2 .a3 , . . . ) for some t ∈ R.

(a) If t ̸= 0, At is invertible, A−1 1


t (b1 , b2 , b3 , . . . ) = t b1 , b2 , . . . , so dim ker At = 0, dim co ker At = 0,
ind (At ) = dim ker At − dim co ker At = 0 − 0 = 0.
(b) If t = 0, A0 is not invertible. We get:
ker A0 = Span {(1, 0, 0, . . . )}
dim ker A0 = 0
∞
( )
X 2
Ran A0 = (0, b2 , b3 , . . . ) : |bi | < ∞ = {(b1 , b2 , b3 , . . . ) : b1 = 0}
i=2
co ker A0 = l2 /Ran A0 ∼
= Span {(1, 0, 0, 0, . . . )}
dim co ker A0 = 1
ind (A0 ) = dim ker A0 − dim co ker A1 = 1 − 1 = 0

Example 7.2 (Right Shift). X = Y = l2 , S + (a1 , a2 , . . . ) = (0, a1 , a2 , . . . ). Here we get:

dim ker S + = 0
Ran S + = (b1 , b2 , b3 , . . . ) ∈ l2 : b1 = 0


co ker S + = l2 /Ran S + ∼= Span {1, 0, 0, 0, . . . }

dim co ker S + = 1
ind S + = dim ker S + − dim co ker S + = 0 − 1 = −1

Example 7.3 (Left Shift). S − (a1 , a2 , a3 , . . . ) = (a2 , a3 , . . . ). Then we get:

ker S − = Span {(1, 0, 0, . . . )}
dim ker S − = 1
Ran S − = l2
co ker S − = l2 /l2 = {0}
dim co ker S − = 0
ind S − = dim ker S − − dim co ker S − = 1 − 0 = 1

45
 n
Pn
Example 7.4 (n-fold Right Shift). ind (S + ) = i=1 ind (S
+) = −n
(iii)

n
Elementary proof. (S + ) (a1 , a2 , . . . ) = (0, . . . , 0, a1 , a2 , . . . ). We calculate:
| {z }
n times

+ n



ker S = {0}
+ n



dim ker S =0
+ n



Ran S = {(b1 , . . . , bn , bn+1 , bn+2 , . . . ) : b1 = · · · = bn = 0}
n  o
+ n ∼



co ker S = Span (1, 0, . . . ), (0, 1, 0, . . . ), . . . , 0, . . . , 0, 1, 0, . . .
n
+ n



dim co ker S =n
+ n



=⇒ ind S = 0 − n = −n

n

Analogously, ind (S − ) = n.

Remark. In particular, any integer a ∈ Z can occur as index of some Fredholm operator.

7.2 Proof of the Main Theorem

We need some topological properties of quotient spaces and range.

Lemma 7.1 (Completeness of Quotient spaces). Let X be a Banach space and U a closed subspace.
Then

a) ∥[x]∥ := inf u∈U ∥x + u∥ is a norm on X/U .

b) X/U with the above norm is a Banach space.

Proof.

a) Homogeneity: Let λ ∈ K, x ∈ X, λ ̸= 0. Then we have:

∥λ[x]∥X/U = ∥[λx]∥X/U = inf ∥ λx + u ∥X

u∈U | {z }
λ(x+ u
λ)
=:λ(x+u′ )

= |λ| inf x + u′ X
homog. of ∥·∥X u∈U

= |λ| ∥[x]∥X/U .

If we now have λ = 0, we get:

∥0[x]∥X/U = ∥[0x]∥X/U = ∥[0]∥X/U = 0 = 0 ∥x∥X/U .

Triangle Inequality: Let x, y ∈ X. Then we get:

∥[x] + [y]∥ = ∥[x + y]∥

= inf ∥x + y + u∥
u∈U
= inf x + y + u + u′ u′ ∈ U
u∈U
≤ inf ∥x + y∥ + y + u′
u∈U
= ∥[x]∥ + y + u′
=⇒ ∥[x] + [y]∥ = ∥[x]∥ + inf
′
y + u′ = ∥[x]∥ + ∥[y]∥ .
u ∈U

46
 Positivity: By definition, U is closed.

0 = ∥[x]∥ = inf ∥x + u∥ = inf x − u′ .

u∈U u′ :=−u u′ ∈U

We therefore can find a sequence of uk with uk → x. Since U is closed, we

know that x ∈ U and therefore

[x] = [0] = 0X/U .

b) Let {[xn ]}n∈N be a Cauchy Sequence in X/U . We now have to show that {[xn ]} has a convergent
sub-sequence (since every Cauchy Sequence with a convergent sub-sequence is itself convergent).
We take {[xnk ]}k∈N such that
   1
xnk − xnk+1 < k ∀k
2
and inductively choose yk ∈ [xnk ] such that
 
∥yk+1 − yk ∥ ≤ 2 [xnk ] − xnk+1 .

This is possible since for all x ∈ X there exists a u ∈ U with ∥x − u∥ ≤ 2 ∥[x]∥. Then we get
1
∥yk+1 − yk ∥ < ∀ k.
2k−1
Hence {yk }k∈N is Cauchy in X. But the map x 7→ [x] is continuous (since ∥[x]∥ ≤ ∥x∥) and we
therefore get:
[xnk ] = [xk ] → [y].

Corollary 7.1. In L (X, Y ) with X = Y = l2 , there exists an open set of non-invertible operators.

In contrast, in L (Rn , Rn ) there does not exist an open set of non-invertible operators, since T is not
invertible if and only if det T = 0.

Proof. Take, e.g. a small ball around a shift operator.

Corollary 7.2. In L (X, Y ) with X = Y = l2 , the set of invertible operators is not dense in L (X, Y ).

In contrast, in L (Rn , Rn ) the set of invertible operators is dense.

Proof. Clear from corollary 7.1

Proof of Theorem 7.1. Our idea is to “make T invertible” and use that invertibility is preserved under
small perturbations.
But first we need to introduce some notation: We write X = U ⊕ V if V is the complement of U , i.e.
if for all x ∈ X we find unique u ∈ U, v ∈ V such that x = u + v.
For the proof of the theorem, we still need two lemmas that we will not prove here:

Lemma 7.2. Let X, Y be Banach spaces, T ∈ L (X, Y ) and dim co ker T < ∞. Then Ran T is closed.

Lemma 7.3. Let X be a Banach space and U a finite-dimensional subspace of X. Then there exists
a closed subspace V such that X = U ⊕ V

Remark. The assumption for U to be finite-dimensional is essential. E.g. U = c0 , X = l∞ but there

exists no V as in the lemma.

47
 By lemma 7.3 we now have:

dim ker T < ∞ =⇒ ∃ closed subspace V of X such that X = V ⊕ ker T.

Let dim co ker T < ∞. We define W := Span {y1 , . . . , yn } with [y1 ], . . . , [yn ] the basis of co ker T =
Y /Ran T and claim: Y = W ⊕ Ran T .
P
Proof of Claim. Let y ∈ Y . We then can write [y] = i αi [yi ] since {[yi ]} is a basis. We
first show the existence of a decomposition:
!
X X
y= αi yi + y− α i yi .
i i
| {z } {z }|
∈W
∈Ran T since [y− i αi yi ]=0
P

For uniqueness, we suppose that y = i αi yi + b = i αi yi + b′ for b, b′ ∈ Ran T . We then

P P
have: " #
X X  X ′
αi yi = αi′ yi =

αi [yi ] = αi [yi ]
i i i

and since {[yi ]} is a basis, αi = αi′ for all i.

We now define an auxiliary operator Se : V × W → Y, Se (v, w)?Sv + w. Obviously Te (v, w) = T v + w

is bijective, since we have the two bijections V ∋ v 7→ T v ∈ Ran t and w 7→ w ∈ W .
Since bijectivity is preserved under small perturbations (theorem 6.1), Se is bijective if Se − Te is
small.
But we have
 
Se − Te (V, W ) = (SV + W ) − (T V + W ) = (S − T )V

=⇒ Se − Te = ∥S − T ∥ .

So Se is bijective if ∥S − T ∥ is small.
Moreover we claim that by construction of S,
e S is Fredholm and more precisely:

ker S ⊆ ker T,

implying particularly dim ker S ≤ dim ker T .

Proof of the Claim. We know that S V = Se V ×{0} is injective since Se is invertible. This
implies that ker S ∩ V = {0} and therefore ker S ⊆ ker T . Furthermore, we get that
ker T ⊕ V which we can rewrite as X = V ⊕ ker S ⊕ Z since ker T = ker S ⊕ Z for some
complement Z. In particular, dim co ker S ≤ dim co ker T .

We now claim that Y = W ⊕ Ran S V

.

Proof of the Claim. This follows from Se being bijective. Let y ∈ Y . We then write:
y = Sv + w for some v ∈ V, w ∈ W since Se is onto. To see uniqueness, we suppose there
is also y = Sv ′ + w′ . Then we get:

0 = (Sv − w) − Sv ′ + w′

= S v − v ′ + w − w′

= Se v − v ′ , w − w′

=⇒ v − v ′ = w − w′ = 0,

since Se is injective.

48
On the other hand, since S is Fredholm, we have Y = W ′ ⊕ Ran S and dim W ′ = dim co ker S for
W ′ = ker S (I think). With the fact that X = V ⊕ ker S ⊕ Z, we can rewrite the first impression to
W ′ ⊕ S Z⊕V = W ′ ⊕ Ran S Z ⊕ Ran Z ⊕ Ran S V with the injectivity of S Z⊕V . We get the following
consequences for the dimensions:

dim ker T = dim ker S + dim Z

′
dim
| {zW} = dim
| {zW} + |dim Ran S Z.
=dim co ker T =dim co ker S
{z }
=dim Z

This implies
dim ker T − dim co ker T = dim ker S − dim co ker S.
This concludes the proof of (i) and (ii). We omit the proof of (iii) and refer to [3].

We now consider an interesting class of operators: Fredholm Operators of index 0.

An elementary fact from the definition of the index is that if T ∈ F (X, Y ) with ind T = 0, then T is
injective if and only if T is surjective.
For an equation T x = b we thus get that uniqueness implies existence, which is particularly interesting.
This motivates to look for criteria for when an operator is Fredholm of index 0.
One criterion is theorem 7.1 that gives us that an operator is close to a Fredholm Operator of index
0. We now give a more global criterion, but first some definitions:

Definition 7.3. Let X be a Banach Space. A subset A ⊆ X is called compact if one of the following
equivalent conditions is met:

ˆ Every sequence {xj }j∈N in A has an accumulation point in A.

ˆ Every open cover of A has a finite sub-cover.

The subset is called relatively compact if A is compact.

Examples.

1. dim X < ∞. Then B1 (0) is compact and therefore B1 (0) is relatively compact.

2. dim X = ∞ Here, B1 (0) is not compact and therefore B1 (0) is not relatively compact (Theorem
about Almost Orthogonal Elements).

3. Prototypical compact set in function space: f ∈ C 1 ([a, b]) : ∥f ∥∞ + ∥f ′ ∥∞ ≤ 1 is a compact



subset of {f ∈ C ([a, b]) : ∥f ∥∞ ≤ 1}.

Definition 7.4. Let X, Y be Banach Spaces. A linear map T ∈ L (X, Y ) is called compact if one of
the following equivalent conditions is met:

ˆ It maps bounded sets to relatively compact sets.

ˆ For every bounded sequence {xn } in X there exists a convergent subsequence of {T xM }.

Examples.

1. X = Y , dim X = ∞. Then I : x 7→ x is not compact.

2. X = Y = l2 , T = D (t1 , t2 , t3 , . . . ) a diagonal matrix with limj→∞ tj = 0 is compact.

Theorem 7.2. Let X, Y be Banach spaces.

(i) If T ∈ L (X, Y ) is Fredholm and K ∈ L (X, Y ) compact, then T + K is Fredholm with the same
index as T .

(ii) In particular, if T ∈ L (X, Y ) is invertible and K ∈ L (X, Y ) compact, then T + K is Fredholm

with index 0.

49
Remark. As compared to theorem 7.1, this is a “global” perturbation result as the perturbation K
can be arbitrarily large. On the other hand, the class of admissible operators K is smaller (as K must
be compact).

Proof of Theorem 7.2 (ii).

Step 0: W.l.o.g. we can say that Y = X, T = I, since otherwise we consider T −1 (T − K) = I −T −1 K,
with T −1 K being compact since it is a composition of a compat and a continuous operator
and compact sets under continuous maps are compact.
Step 1: Here we want to show that
dim ker (I − K) < ∞.
| {z }
=:S
Suppose dim ker S = ∞. Then ker S is the preimage of {o} under the continuous map S,
hence closed. Since ker S is a closed subspace of a Banach space, it is itself Banach and we
get:
x ∈ ker S =⇒ 0 = Sx = x − Kx
⇐⇒ x = Kx
=⇒ B1 (0) = KB1 (0) ,
where B1 (0) is the unit ball in the kernel, which is not relatively compact but KB1 (0) is
since K is compact. This is a contradiction.
Step 2: We now want to show that dim co ker S < ∞. We assume that X is Hilbert to simplify the
proof (General proof: see [5, Satz VI.2.1(b)]).
Then we have:
co ker S = X/Ran S ∼ ⊥
= (Ran S) = {y ∈ X : ⟨y, S⟩ = 0 ∀ s ∈ Ran S} ,
if Ran S is closed, which we will later show. We furthermore have the isomorphism
X/Ran S ∋ [x] 7→ P ⊥ x,
which is the projection onto Ran S ⊥ (check with the projection theorem that this is indeed
an isomorphism).
a) Now we want to show that dim co ker S < ∞, provided that Ran S is closed.
The intuition behind this fact is that if dim (Ran S)⊥ = ∞, then Ran S is “too small”,
i.e. S would map too many vectors onto too small a space.
To formalize this argument, we suppose that dim (Ran S)⊥ = ∞ and choose an orthonor-
mal sequence {e1 , e2 , . . . } ⊂ (Ran S)⊥ . Then:
√ √
ˆ ∥ei − ej ∥ = ⟨ei − ej , ei − ej ⟩ = 1 − 0 − 0 + 1 = 2 if i ̸= j
p

ˆ ei − ej ∈ (Ran S)⊥
From the projection theorem, we know that for all z ∈ (Ran S)⊥ , ∥z∥ = dist (z, Ran S).
Hence for i ̸= j
√
2 = ∥ei − ej ∥ = dist (ei − ej , Ran S) ≤ ∥(ei − ej ) − S(ei − ej )∥
= ∥Kei − Kej ∥ ,
since S = I − K. So the set {Kei }i∈N is not relatively compact, even though it is the
image of the bounded set {ei }i∈N , which is a contadiction.
b) Now, we show that Ran S is closed. Suppose that we have a sequence yj ∈ Ran S =
Ran (I − K) with yj → y ∈ X. We then need to prove that y ∈ Ran S. We have:
yj ∈ Ran S =⇒ ∃ xj ∈ X : Sxj = xj − Kxj = yj → y. (∗)
Now the difficulty is that we don’t know whether the xj are convergent.

50
 Case 1: The {xj } are bounded. Since K is compact, there exists a subsequence {xj }
such that Kxj → z ∈ X. Then:
xj = Kxj + yj → z + yj =: x.
(∗)

But
Sx = lim Sxj = lim yj = y,
j→∞ j→∞
where we have used the continuity of S, so y ∈ Ran S.
Case 2: The {xj } are unbounded. We define αj := dist (xj , ker S) and choose wj ∈ ker S
as the almost-optimizer of distance to xj in ker S, i.e.:
 
1
αj ≤ ∥xj − wj ∥ ≤ 1 + αj .
j
Then we have
S(xn − wn ) = Sxn = yn ,
so xn − wn is still the inverse image of yn under S.
If {xn − wn } are bounded, we apply Case 1 to x̃n = xn − wn .
If {xn − wn } are unbounded we can, by passing to a subsequence, assume that
−wn
αn → ∞. Then with zn := ∥xxnn −w n∥
we have:
convex,
hence bounded
z}|{
Sxn xn
Szn = = → 0.
∥xn − wn ∥ ∥xn − wm ∥
| {z }
→∞
Since ∥zn ∥ = 1 and K compact, there exists a subsequence such that
Kzn → w0 =⇒ zn → w0 .
Hence (I − K)w0 = 0 knowing that Szn = (I − K)zn → 0, i.e. w0 ∈ ker S. But
x n − wn 1
∥zn − w0 ∥ = − w0 = ∥xn − wn − ∥xn − wn ∥ w0 ∥
∥xn − wn ∥ ∥xn − wn ∥
αn
≥ →1
∥xn − wn ∥
with wn ∈ ker S, which is a contradiction to zn → w0 .
Step 3: Now, we need to show that ind (S) = 0. For this, we use a “continuous deformation argu-
ment”.
Consider the map
S : [0, 1] → L (X, Y ) , t 7→ I − tK.
By step 1 and 2, we know that S (t) is Fredholm for all t and by theorem 7.1 it follows
that for all t, there exists a δ (t) > 0 such that Bδ(t) (S (t)) ⊂ F (X, X) and the index on
Bδ(t) (S (t)) is constant.
We know that A = {S (t) : t ∈ [0, 1]} is compact since S is continuous and [0, 1] is compact.
Then, the covering Bδ(t) (S (t)) t∈[0,1] of A has a finite sub-cover
Bδ1 (S (t1 )) , . . . , BδN (S (tN )) .
W.l.o.g. t1 < t2 < · · · < tN with none of the Bδi (S (ti )) being redundant. We get:
Bδi (S (ti )) ∩ Bδi+1 (S (ti+1 )) ̸= ∅ ∀ i.
So the index on Bδi (S (ti )) is the same as the index on Bδi+1 (S (ti+1 )). This implies that
the index of S (t) is constant and independend of t, so
0 = ind I = ind (S (0)) = ind (I − K) = ind (S (1))
and the claim follows.

51
8 Spectral Theory of Compact Operators
Theorem 8.1 (Spectrum of a Compact Operator). Let X be a Banach space, K ∈ L (X, X) compact
and K = C. Then:

(a) Exactly one of the following alternatives holds:

(i) spec K = {λ1 , . . . , λn } for distinct λj ∈ C, n ∈ N

(ii) spec K = {λj }∞
j=1 ∪ {0} for distinct λj ∈ C \ {0} , λj → 0 (j → ∞)

4 4

2 2

C C
−4 −2 2 4 −4 −2 2 4

−2 −2

−4 −4

(a) First case (b) Second case

(b) dim X = ∞ =⇒ 0 ∈ spec K

(c) λ ∈ spec K \ {0} =⇒ λ eigenvalue of K

Remarks.

1. The spectrum is “quantized”, i.e. only a discrete set.

2. λ = 0 has a special role; it is the only value for which K −λI is not Fredholm (Fact: An invertible
and compact operator has Fredholm index 0).

One possible example of this are eigenvalues of a boundary value problem for an ordinary differ-
ential equation:
u′ = λu u 2π-periodic, u : R → C.
We will look at a different one in more detail:

Example (Function spaces). Consider

  2π 
1 1 ′ ′
Cper,0 := u ∈ C ([0, 2π]) : u (0) = u (2π) , u (0) = u (2π) , u=0
0

with norm ∥u∥C 1 = ∥u∥∞ + ∥u′ ∥∞ and

per,0

  2π 
Cper,0 := u ∈ C ([0, 2π]) : u (0) = u (2π) , u=0
0

with norm ∥u∥Cper,0 = ∥u∥∞ .

2π

d
We define the operator L := dx , i.e. Lu = u′ with L : Cper,0
1 1
→ Cper,0 (note: U ∈ Cper,0 =⇒ 0 Lu =

2π ′ 2π ′
0 u = u |0 = 0, i.e. the boundary condition on u implies the integral condition on u ).
We now have some problems: The domain space is not the target space and the operator is not

52
compact. We solve them by considering the inverse operator instead.
This inverse operator is the Integral Operator. We get:

Lu = f ⇐⇒ u′ = f
⇐⇒ u = the antiderivative of f with zero mean
 x  2π  x 
1
⇐⇒ u (x) = f (x) dx − f (s) ds dx,
0 2π 0 0
| {z }
mean of the first term

1
where we used that u ∈ Cper,0 . We want to write the right hand side in a nicer way. For that, we
define: (
1, s ∈ [0, x]
χ[0,x] (s) = .
0, else
We then write:
 2π 2π 2π  
1
u (x) = χ[0,x] (s) f (s) ds − χ[0,x] (s) f (s) ds dx
0 2π 0 0
 2π  2π  2π 
1
= χ[0,x] (s) f (s) ds − χ[0,x] (s) dx f (s) ds
Fubini 0 2π 0
| 0 {z }
=2π−s
 2π  
2π − s
= χ[0,x] (s) − f (s) ds.
0 2π
| {z }
=:k(x,s)

We set  2π
(Kf ) (x) := k (x, s) f (s) ds (Integral Operator)
0

and get L−1 = K with

1
K : Cper,0 → Cper,0 .
The inverse of the Differential Operator is an Integral Operator.
1
Since Cper,0 ⊆ Cper,0 , we can consider K as an operator K : Cper,0 → Cper,0 , so its spectrum is well-
defined.
Integral operators like the K above are compact (see class). This also follows from the following
lemma:

Lemma 8.1. Let a < b, a, b ∈ R, R > 0. Then the set

A = u ∈ C 1 ([a, b]) : ∥u∥∞ + u′


∞
≤R

is relatively compact in C ([a, b]).

Principle: Bounds on higher derivatives yield compact sets in function spaces.

1
We know that K : Cper,0 1
→ Cper,0 is continuous (since it is the inverse of the continuous operator L
and using the Open Mapping Theorem). This implies that KB1 (0) with B1 (0) ⊆ Cper,0 is a bounded
1
set in Cper,0 and a relatively compact set in Cper,0 , so K : Cper,0 → Cper,0 is compact.
We now want to take a look at the spectrum of K. We know:

λ ∈ spec K \ {0} ⇐⇒ λ eigenvalue of K

theorem 8.1
⇐⇒ ∃ f ̸= 0 : Kf = λf
⇐⇒ f = λLf
1
⇐⇒ f = Lf
λ

53
 1
⇐⇒ µf = Lf, µ :=
λ
⇐⇒ f ′ = µf
⇐⇒ f (x) = Ceµx ,

since f ∈ Cper,0 . This gives us:

f (0) = f (2π)
=⇒ eµ·2π = eµ·0
⇐⇒ eRe µ·2π · ei(Im µ)·2π = 1.

Since eRe µ·2π = 1, we know that Re µ = 0 and get:

ei(Im µ)·2π = 1
=⇒ µ = ik k ∈ Z.

With this result, we get:

 2π
f = 0 =⇒ k ∈ Z \ {0}
0
 
1
=⇒ spec K = : k ∈ Z \ {0} ∪ {0} .
ik |{z}
because of theorem 8.1

For the eigenvectors, we get:

1 n 1 o
λ= =⇒ ker (K − λI) = Span e λ x .
ik
Proof of Theorem 8.1.

(a) This is the most difficult to prove out of the three statements; it will require three lemmas and
one proposition:

Lemma 8.2. Let X be a vector space, T : X → X a linear map and v1 , . . . , vr eigenvectors with
different eigenvalues λ1 , . . . , λr . Then v1 , . . . , vn are linearly independent.

Proof. We use induction. The case r = 1 is obviously true. We assume the hypothesis for an
arbitrary, but fix r − 1 and take linearly independent v1 , . . . , vr , meaning that

a1 v1 + · · · + ar vr = 0 (8.1)

for not all ai = 0. We multiply with T and get:

a1 λ1 v1 + . . . ar λr vr = 0. (8.2)

Now we multiply (8.1) with λr and subtract (8.2), resulting in:

(λr − λ1 )a1 v1 + · · · + (λr − λr−1 )ar−1 vr−1 = 0.

The induction hypothesis gives us that (λr − λ1 )a1 , . . . , (λr − λr−1 )ar−1 = 0. Since the λi are
pairwise distinct, we know that a1 , . . . , ar−1 = 0 and with (8.1), this gives us that ar = 0, since
vr ̸= 0.

Lemma 8.3. Let X be Banach, T = I − K for K ∈ L (X, X), M and L closed subspaces with
M ⊊ L and T (L) ⊆ M , i.e. the image of larger the larger subspace lies in the smaller subspace.
Then there exists a ∈ L \ M such that ∥a∥ = 1 and ∥Ka − Kx∥ ≥ 12 for all x ∈ M .

54
Proof. By the theorem about almost orthogonal elements, there exists an a ∈ L \ M such that
∥a∥ = 1 and dist (a, M ) ≥ 12 .
Then the conclusion follows from T = I − K and T (L) ⊆ M as follows:

T a = a − Ka =: x̃ ∈ M
=⇒ Ka = a − x̃
=⇒ Ka − Kx = a − x̃ − Kx
|{z} x∈M
∈M since K(M )⊆M
1
=⇒ ∥Ks − Kx∥ ≥ dist (a, M ) ≥ .
2

Proposition 8.1. Let X be a Banach space and K ∈ L (X, X) a compact operator. The spec K
̸ 0.
has no accumulation point λ∗ =

Remark. It is essential that K is compact, recall spec S− on l2 is {λ ∈ C : |λ| ≤ 1}.

Proof. We use a distance estimate and the linear independence lemma from above to prove this
indirectly.
Suppose spec K has an accumulation point λ∗ ̸= 0. Let ε > 0 and choose a sequence {λj } ∈
spec K \ {0} such that
1 1
− <ε ∀ j.
λ∗ λj
From theorem 8.1 c), we know that λj are eigenvalues, i.e. there exist corresponding eigenvectors
vj .
Without loss of generality, let all λj be distinct. Let Un := Span {v1 , . . . , vn }. From lemma 8.2,
we know
U1 ⊊ U2 ⊊ U3 ⊊ . . . .
Now, we would like to apply lemma 8.3 and note that (K − λn I)Un ⊆ Un−1 and moreover all Un
are finite dimensional and hence closed. We therefore know that there exists xn ∈ Un \ Un−1 and
∥xn ∥ = 1 such that for λ1n K − I we have:

1 1 1
Kxn − Kx ≥ ∀ x ∈ Un−1 .
λn λn 2

We take x = xm with m < n and get:

1 1 1
≤ Kxn − Kxm
2 λn λn
! !
1 1 1 1
≤ Kxn − Kxm + 1 Kxn + 1 Kxm
λ∗ λ∗ λn − λ∗ λn − λ∗
| {z } | {z }
≤ε∥K∥·1 ≤ε∥K∥·1
1 1
=⇒ − 2ε ∥K∥ ≤ ∥Kxn − Kxm ∥ ∀ m < n,
2 λ∗
so {Kxn } has no convergent subsequence, even though {xn } is bounded, which is a contradiction
to K being compact.

Lemma 8.4. Let X be a Banach space and T ∈ L (X, X). Then

spec T ⊆ {λ ∈ C : |λ| ≤ ∥T ∥} .

In particular, the spectrum of T is bounded.

55
 Proof. Suppose |λ| > ∥T ∥. Then we have:
1 1
|λ| = 1 = =⇒ λI − T invertible.
λI (λI)−1 Neumann series

Now, we can finally conduct the proof of our initial statement. By b), we know that spec K ̸= ∅
and can therefore assume |spec K| = ∞. We then have to show that the spectrum is countable
with the only accumulation point 0.
By Linear Algebra, we must have dim X = ∞. Lemma 8.4 gives us that the spectrum of K is
bounded and therefore we get with proposition 8.1 that it has no accumulation point λ∗ ̸= 0.
We still need to show that spec K is countable and 0 is an accumulation point.
Our idea is to use a decomposition of C into concentric rings:

A1 := {λ ∈ C : |λ| ≥ 1}
 
1
A2 := λ ∈ C : 1 > |λ| ≥
2
 
1 1
A3 := λ ∈ C : > |λ| ≥
2 3
..
.

By the boundedness and lack of non-zero accumulation point of spec K, we have that spec K ∩ Aj
is finite for all j, since otherwise there exists an accumulation point by Bolzano-Weierstrass. We
set:

spec K ∩ A1 =: {λ1 , . . . , λr1 }

spec K ∩ A2 =: {λ1 , . . . , λr2 }
..
.

Since C \ {0} = ∞
S
j=1 Aj , every λ ∈ spec K \ {0} lies in exactly one Aj . Hence spec K \ {0} = {Aj }.
By definition of λj it is clear that λj → 0 for j → ∞.

(b) We will show this via contraposition. We have:

0∈
/ spec K ⇐⇒ K − oI invertible
  
−1
⇐⇒ B1 (0) = K
|{z} K B1 (0)
continuous by OMT | {z }
compact, since K compact

=⇒ B1 (0) compact (since the image of compact sets under cts. functions is cpt.)
⇐⇒ dim X < ∞
theorem 2.2

(c) Suppose λ ∈ spec K with λ ̸= 0. Then we need to show that λ is an eigenvalue. We know:

λ ̸= 0 =⇒ λI − K is invertible and compact

=⇒ λI − K Fredholm with index 0, (∗)
theorem 7.2 (ii)

so λ ∈ spec K ⇐⇒ λI − K not invertible ⇐⇒ λI − K not injective ⇐⇒ λ is an eigenvalue.

(∗)

56
8.1 Comparison with Jordan Normal Form

dim X < ∞, T ∈ L (X, X) automatically com- dim X = ∞, T ∈ L (X, X) compact

pact
Jordan normal form: there exists a basis  such For all non-zero eigenvalues λi , there exist Ui , Vi
λ 1 such that T − λi I U is nilpotent, Ui , Vi are in-
that T = D (J1 , . . . , Jr ) and Ji = λi or i i
0 λi variant under T , i.e. T Ui ⊆ Ui and X = Ui ⊕ Vi .
 
λi 1 0
or  0 λi 1  etc. with Ji − λi I nilpotent.
0 0 λi
There exists a basis of eigenvectors and general- In general, there does not exist a basis of eigen-
ized eigenvectors (i.e. (T − λI)n v = 0 for some vectors and generalized eigenvectors.
n).

Example (Compact Operator without Eigenvectors and Generalized Eigenvectors). We consider the
operator  
0
1 0 
 1

K=

2 0 

1

 3 0 

.. ..
. .

on l2 . We then have K = S+ · D 1, 21 , 31 , . . . , with S+ being continuous and D 1, 12 , 13 , . . . being a

compact operator. K has no eigenvalues, as we will shortly prove and spec K = {0}, since by theorem
8.1, spec K \ {0} is all non-zero eigenvalues.
We have  a a a 
0 1 2 !
K (a0 , a1 , a2 , . . . ) = 0, , , , . . . = (λa0 , λa1 , λa2 , . . . ),
1 2 3
so if λ ̸= 0, we know by the first component, that a0 = 0, by the second component that a1 = 0 etc.
If λ = 0, the second component gives us a0 = 0, the third a1 = 0 etc.

We will now consider an important class of Compact Operators with very nice theory.

Definition 8.1 (Charles Hermite, 1822-1901). Let X be a Hilbert space. An operator T ∈ L (X, X)
is called self-adjoint (or sometimes symmetric or hermitian) if

⟨T x, y⟩ = ⟨x, T y⟩ x, y ∈ X.

Example. X = Rn with standard inner product. We consider a matrix (Tij )ij and get:

⟨T ei , ej ⟩ = ⟨Ti∗ , ej ⟩ = Tji
⟨T ej , ei ⟩ = ⟨Tj∗ , ei ⟩ = Tij ,

so T is self-adjoined if and only if:

⟨T ei , ej ⟩ = ⟨T ej , ei ⟩ ∀ i, j
⇐⇒ Tji = Tij ∀ i, j
⊤
⇐⇒ T = T ,

meaning that T has to be symmetric.

Self-adjoined operators are the generalization of symmetric matrices in infinite dimen-

sions.

57
Proposition 8.2. Let X be a Hilbert space and T ∈ L (X, X) self-adjoined. Then

(a) All eigenvalues of T are real.

(b) Eigenvectors corresponding to different eigenvalues are orthogonal.

Proof. (a) Let T x = λx for ∥x∥ = 1. We can write:

λ = λ ⟨x, x⟩ = ⟨λx, x⟩ = ⟨T x, x⟩
=
λ = λ ⟨x, x⟩ = ⟨x, λx⟩ = ⟨x, T x⟩

(b) T x = λx, T y = µy with λ ̸= µ. Then we have:

(λ − µ) ⟨x, y⟩ = (λ − µ) ⟨x, y⟩ = ⟨λx, y⟩ − ⟨x, µy⟩ = ⟨T x, y⟩ − ⟨x, T y⟩ = 0,

(a)

since T is self-adjoint. This gives us that ⟨x, y⟩ = 0, since λ ̸= µ.

This does not yet rule out Jordan blocks, but the following theorem does:

Theorem 8.2 (Spectral Theorem for Compact Self-Adjoint Operators). Let X be a Hilbert space
and T ∈ L (X, X) compact and self-adjoint. Then we have:

(1) There exist finitely or countably many eigenvalues λj ∈ R \ {0} with |λ1 | ≥ |λ2 | ≥ · · · > 0 and
corresponding orthonormal eigenvectors e1 , e2 , . . . such that

(i) X = ker T ⊕ Span {e1 , e2 , . . . }

P
(ii) T x = j≥1 λj ⟨x, ek ⟩ ej ∀ x ∈ X.

(2) If X is separable, there exists an orthonormal basis of X consisting of eigenvectors of T .

Lemma 8.5. Let X be a Hilbert space, A ∈ L (X, X) self-adjoint and U ⊂ X an invariant subspace
of A (i.e. AU ⊆ U ), then so is U ⊥ (i.e. AU ⊥ ⊆ U ⊥ ).

Example. If U is an eigenspace of A, it is invariant.

Proof. Let y ∈ U ⊥ . Then we need to prove that Ay ∈ U ⊥ . Let furthermore u ∈ U . Then we can
write: * +
⟨Ay, u⟩ = y, |{z}
Au = 0,
A self-adj.
∈U

since y ∈ U ⊥ . Hence Ay ∈ U ⊥ .

Example. X = R with standard inner product. We consider the matrix

 
0 1
A= .
0 0

We then have Ae1 = 0, so U = Span {e1 } is invariant. Then U ⊥ = Span {e2 }, but Ae2 = e1 ∈
/ U ⊥.

The lemma gives us a heuristic (but non-rigorous) proof of the spectral theorem: We first look
for some eigenvector and restrict T to the orthogonal complement of the eigenvector. Then by the
lemma, T maps the orthogonal complement to itself. We look for another eigenvector in the orthogonal
complement and restrict T to the orthogonal complement of both eigenvectors etc.
Gaps: existence, need to show that we exhaust X entirely.
The rigorous proof proceeds analogously, but addresses the gaps.

58
Lemma 8.6 (Characterization of the Operator Norm via the Inner Product). Let X be Hilbert and
A ∈ L (X, X) self-adjoint. Then
∥A∥ = sup |⟨Ax, x⟩| .
∥x∥=1

This is interesting, even in Rn .

Proof. We have:
 
Ax
∥A∥ = sup ∥Ax∥ = sup Ax,
∥x∥=1 ∥x∥=1 ∥Ax∥
= sup ⟨Ax, y⟩
Ax
eq. in CSI with y= ∥Ax∥ ∥x∥=∥y∥=1

1
= (⟨A(x + y), x + y⟩ − ⟨A(x − y), x − y⟩)
sup
4
identity ∥x∥=∥y∥=1

1  
≤ sup sup |⟨Az, z⟩| ∥x + y∥2 + ∥x − y∥2
∥x∥=∥y∥=1 4 ∥z∥=1
1  
= sup sup |⟨Az, z⟩| 2 ∥x∥2 + 2 ∥y∥2
∥x∥=∥y∥=1 4 ∥z∥=1 | {z }
=4
≤ sup |⟨Az, z⟩| .
∥z∥=1

The other inequality is trivial:

sup |⟨Ax, x⟩| ≤ sup (∥Az∥ · ∥z∥) = ∥A∥ .

∥x∥=1 CSI ∥z∥=1

Lemma 8.7. Let X be a Hilbert space, B ∈ L (X, X) self-adjoined and B ≥ 0, i.e. ⟨Bx, x⟩ ≥ 0 for
all x ∈ X. Then
∥Bx∥2 ≤ ∥B∥ · ⟨Bx, x⟩ ∀ x ∈ X.
In particular, ⟨Bx, x⟩ = 0 implies Bx = 0.

Proof. The proof is easy with lemma 8.6. Let b := ∥B∥.

1. We have:

b ⟨x, x⟩ = (2b − b) ⟨x, x⟩ ≥ ⟨(2B − b)x, x⟩ ≥ −b ⟨x, x⟩ ∀ x

l. 8.6 B≥0
=⇒ |⟨(2B − b)x, x⟩| ≤ b ⟨x, x⟩ ∀x
=⇒ ∥2B − b∥ ≤ b.
l. 8.6

2. The last inequality above is, by definition of the Operator Norm, equivalent to

0 ≥ ∥(2B − b)x∥2 − ∥bx∥2 = ∥Bx + (B − b)x∥2 − ∥Bx − (B − b)x∥2

 
= 4 ⟨Bx, (B − b)x⟩ = 4 ∥Bx∥2 − b ⟨Bx, x⟩ .

Theorem 8.3 (Rayleigh-Ritz Variational Principle). Let X be a Hilbert space and A ∈ L (X, X)
self-adjoined and compact. Then

sup ⟨Ax, x⟩ and inf ⟨Ax, x⟩

∥x∥=1 ∥x∥=1

are, if they are non-zero, attained and eigenvalues of A.

59
Proof. Let λ := sup∥x∥=1 ⟨Ax, x⟩ > 0. Then B := λI − A ≥ 0, since ⟨Bx, x⟩ = λ ⟨x, x⟩ − ⟨Ax, x⟩ ≥ 0.
From lemma 8.7, we know:

∥(A − λ)x∥2 ≤ ∥A − λI∥ (λ − ⟨Ax, x⟩) ∀ x ∈ X, ∥x∥ = 1. (∗)

In particular, if x is a maximizer, the right hand side is zero and therefore the left hand side as well.
Then x is an eigenvalue of A with eigenvalue λ.
In Rn , this corollary could be proved via the Lagrange multiplier rule.
We choose a maximizing sequence for the problem of maximizing ⟨Ax, x⟩ on {∥x∥ = 1}, i.e. a sequence
{xn }n∈N such that ∥xn ∥ = 1, ⟨Axn , xn ⟩ → λ. By (∗),

(A − λ)xn → 0. (8.3)

On the other hand, since A is compact, there exist subsequences of {xn }, again denoted by {xn }, such
that
Axn → y ∈ X. (8.4)
Claim: y ̸= 0.

Proof of the Claim. Otherwise we get:

|⟨Axn , xn ⟩| ≤ ∥Axn ∥ ∥xn ∥ → 0,

| {z } | {z }
→|λ| =1

which would be a contradiction to λ ̸= 0.

Claim: y is an eigenvector of A.

Proof of the Claim. We take (8.4) − (8.3). This gives us:

λxn → y =⇒ λy ← λAxn = Aλxn → Ay

Hence Ay = λy, i.e. y is an eigenvector of A with eigenvalue λ.

Proof of Theorem 8.2. By the projection theorem, we have

X = ker T ⊕ (ker T )⊥ .

By general properties of (not necessarily self-adjoint) compact operators, there exist at most countably
many non-zero eigenvalues (→ theorem 8.1) and the corresponding eigenspaces are finite-dimensional
(by compactness). Let {e1 , e2 , . . . } be the union of orthonormal bases of all these eigenspaces. By
proposition 8.2, {e1 , e2 , . . . } is an orthonormal basis. Hence, we need to show: Span {e1 , e2 , . . . } =
(ker T )⊥ .
By lemma 8.5, we know that V := (ker T )⊥ is an invariant subspace of T and therefore T V ∈ L (V, V ) is
compact and self-adjoint. Suppose Span {e1 , e2 , . . . } ⫋ (ker T )⊥ . Let W be the orthogonal complement
of Span {e1 , e2 , . . . } in V , that is

{x ∈ V : ⟨x, ej ⟩ = 0 ∀ j = 1, 2, . . . } .

Again, by lemma 8.5, W is invariant under T and therefore T W

∈ L (W, W ) is compact and self-
adjoint. We furthermore have:

W ⊥ ker T =⇒ T W
̸= 0 =⇒ T W L(W,W )
̸= 0
=⇒ sup |⟨T w, w⟩| =
̸ 0.
l. 8.6 w∈W, ∥w∥=1

60
By Rayleigh-Ritz, this implies that T W has a non-zero eigenvalue with an eigenvector in W . This is
a contradiction to the fact that W ⊥ Span {e1 , e2 , . . . }, which proves (1)(i).
For the proof of (1)(ii), we have by (i):
N
X
x = x0 + lim ⟨x, ej ⟩ ej ∀x ∈ X
|{z} N →∞
∈ker T j=1
N
X N
X
=⇒ T x = T x0 + lim ⟨x, ej ⟩ T ej = lim ⟨x, ej ⟩ λj ej .
T cont. & lin. N →∞ N →∞
j=1 j=1

For (2), we note that X is separable and therefore ker T as well. This implies that there exists
an orthonormal basis {f1 , f2 , . . . } of ker T . Hence {e1 , f1 , e2 , f2 , . . . } is an orthonormal basis of X
consisting of eigenvectors of T .

Functional-Analytic Derivation & Viewpoint of the Fourier Series

Representation of Functions
This is a spectacular application of theorem 8.2.

The Fourier Series Representation

We take the space X = L2 ([0, 2π]) with basis ek (x) = √1 eikk .
2π
For a plot of the real part of the basis,
1
see figure 13. The 2π is a normalization factor so that ∥ek ∥L2 = 1. Note that
 2π 2
1
∥ek ∥2L2 = √ eikx dx = 1.
0 2π
| {z }
1
= 2π ·1

Theorem 8.4 (Fourier Series Representation). The {ek }k∈Z are an orthonormal basis of L2 ([0, 2π]).
That is, for f ∈ L2 ([0, 2π]) we have
N
X
f = lim ⟨f, ek ⟩ ek (∗)
N →∞
k=−N

with the limit denoting convergence in L2 ([0, 2π]), i.e.

N
X
f− ⟨f, ek ⟩ →0 N → ∞.
k=−N L2

Explicitely, this gives us:

 2π  2π
1
⟨f, ek ⟩ = f (x) ek (x) dx = f (x) e−ikx dx
0 2π 0
| {z }
=:fˆk , k-th Fourier coefficient

with
N
X 1 ˆ ikx
⟨f, ek ⟩ ek = fk e .
2π
k=−N

In classical notation, we write:

∞
1 X ˆ ikx
f (x) = fk e ,
2π
k=−∞

which is the Fourier series representation of f .

61
Functional-Analytic View Point
Theorem 8.4 is a special case of theorem 8.2 with the idea to find a compact self-adjoined operator
whose eigenvectors are the ek , k ∈ Z.
But we know that there are eigenvectors of the derivative operator:
d ikx d
e = ikeikx =⇒ ek = ikek .
dx dx
d
This poses the problems that dx is not defined on L2 , and it is neither compact nor self-adjoined.

Considering the Inverse

d −1
and note that if L is invertible, L and L−1 have the


As a first step, we look at the inverse dx
same eigenvectors. This will fix the first two problems.
We start with the inverse on a smaller space and then extend that. From an example from lecture
d 1
16, we already know that L := dx with L : Cper,0 → Cper,0 is invertible with inverse function K, which
gives the antiderivative of a function with zero mean. We then have:
 x  2π  x
1
(Kf ) (x) = f (s) ds − f (s) ds dx. (∗)
0 2π 0 0

But we cannot work an Cper,0 , since it is not Hilbert.

Hilbert Space on which K is Well Defined

We take the space 
 2π 
X= L20 ([0, 2π]) 2
= f ∈ L ([0, 2π]) : f =0 .
0
2
Then (∗) is well defined for f ∈ L for each x ∈ [0, 2π].
Claim: Kf ∈ L20 ∀ f ∈ L20 , i.e. K : X → X.

Proof of the Claim. Kf is even continuous. Namely

 x  y
|Kf (x) − Kf (y)| = f (s) ds − f (s) ds
0 0
 y
= − 1 · f (s) ds
w.l.o.g. x<y x
s s
y y
≤ 12 ds · f (s)2 ds ,
CSI x x
√
| {z } q

| {z }
= x−y 2π 2
≤ 0 f (s) ds=∥f ∥L2

i.e.
p
|Kf (x) − Kf (y)| ≤ |x − y| ∥f ∥L2 (∗∗)

Same Eigenvectors despite Bigger Space

We use a regularity argument. In lecture 16, we showed that the eigenvectors of K : Cper,0 → Cper,0 are
precisely the eikx k∈Z\{0} . But on the bigger space L20 ⫌ Cper,0 , there might be further eigenvectors.


Claim: This does not happen.

Proof of the Claim. Suppose u ∈ L20 with Ku = λu. Since λ ̸= 0 w.l.o.g., we can write
u = λ1 Ku with Ku ∈ Cper,0 (see above). Therefore u ∈ Cper,0 .

62
Remark.
1 ∞
u ∈ Cper,0 =⇒ u ∈ Cper,0 =⇒ u ∈ Cper,0 .
Fund. theo. of calc. iterate

This is an instance of a typical phenomenon for differential or integral operators: Eigenvectors have
much “higher” regularity than general functions in the Hilbert space under consideration.

Compactness
We know from lecture 16 that K is compact on Cper,0 , but this is not enough.

Definition 8.2 (Hölder-Continuous Functions). Let α ∈ (0, 1] and A ⊂ Rn compact. We define:

 
 
α
 |u (x) − u (y)| 
C (A) := u ∈ C (A) : sup α <∞

 x,y∈A |x − y| 

x̸=y

with norm
|u (x) − u (y)|
∥u∥α := ∥u∥∞ + sup .
x,y∈A |x − y|α
x̸=y

For α = 1 this is Lipschitz continuity.

Lemma 8.8. Let a ∈ (0, 1]. Bounded sets in C α ([a, b]) are relatively compact subsets of C ([a, b]).

Claim: K : L20 → L20 is compact.

Proof of the Claim. We need to show: If {fj } is bounded in L20 , then {Kfj } has a conver-
gent subsequence in L20 .
Since fj is bounded in L20 , we know by (∗∗), that Kfj is bounded in C α ([0, 2π]) with α = 21 .
By lemma 8.8, we know that Kfj has a convergent subsequence in C ([0, 2π]) and since
 2π
s
q √
∥fj − f ∥L2 = |fj − f |2 ≤ 2π · ∥fj − f ∥2∞ = 2π ∥fj − f ∥∞ ,
0

it even has a convergent subsequence in L20 .

Self-Adjointness
We first look at K −1 = d
dx
1
and suppose u, v ∈ Cper,0 . We have:
   2π  2π  2π  
d d d d d
u, v = u · v dx = [u · v]02π − u· v dx = u· v dx = − u, v .
dx 0 dx 0 dx 0 dx dx

So the derivative operator is self-adjoined up to a minus sign. This is called a skew-symmetric operator.
But there is a very simple but clever way out: We use 1i dx d
, which in physics is called the momentum
operator. This gives us:
       
1 d 1 d 1 d 1 d
u, v = u, v = − u, v = u, − v .
i dx i dx i dx i dx

d −1
We now look at the inverse: 1i dx 1


= iK : Cper,0 → Cper,0 .
2
Claim: iK is self-adjoined on L0 .

63
 0 . Let u := Kf , v := Kg, i.e. f = u′ ,
Proof of the Claim. First, suppose that f, g ∈ Cper
g = v ′ . We then have:
   
d d
⟨f, Kg⟩ = u, v = − u, v = − ⟨Kf, g⟩
dx dx
=⇒ ⟨f, iKg⟩ = ⟨iKf, g⟩ .

So Cper,0 is dense in L20 and the inner product and K are continuous. By theorem 8.2 for
iK on L20 , we know that {ek }k∈Z\{0} is an orthonormal basis of L20 . We now only need to
abolish the zero mean: Let f ∈ L2 ([0, 2π]). Then:

f = (f − ⟨f ⟩) + ⟨f ⟩
| {z } |{z}
∈L20 ∈Span{e0 }

1

2π
with ⟨f ⟩ := 2π 0 f (s) ds and therefore {ek }k∈Z is an orthonormal basis of L2 .
As a summary, we can say the following:
ˆ From the point of view of Functional Analysis, Fourier Series are “nothing special”. Functions
can be represented as a series with respect to the eigenfunctions of any compact selfadjoined
d −1
operator. To get the Fourier series, we e.g. take 12 dx


.

ˆ Functional Analysts (unlike most textbooks) don’t need to check that ⟨ek , el ⟩ = 0 for k ̸= l.
They believe in proposition 8.2.
Theorem 8.5 (Compactness). Let A ⊂ Rn be compact and α ∈ (0, 1]. Then bounded sets in C α (A)
are relatively compact in C (A).

Lemma 8.8 is the special case of this with A = [a, b].

Proof.
1. First we need to show that a subsequence converges on a dense subset.
Therefore we take a bounded set B in C α (A) and dense subset D ⊂ A, e.g. A = {a, b} and
D = A∩Q and get D = {q1 , q2 , . . . }. With a more general A, we have: {q˜1 , q˜2 , . . . } = Qn ∩[−r, r]n
and D = {q1 , q2 , . . . } where qi is any nearest point to q˜i in A.
Let u(ν) be a bounded sequence in C α (A). If we apply the Bolzano-Weierstrass theorem in
K = R or C to the sequence u(ν) (q1 ), which is a bounded sequence in K, we get that there exist
the following subsequences:
(1) (1) (1) (1)
uν1 , uν2 , uν3 , . . . : uν1 (q1 ) → u∗ (q1 ) ∈ K
n→∞
(2) (2) (2) (2)
uν1 ν2
,u ν3
,u , . . .: u ν1
(q2 ) → u∗ (q2 ) ∈ K
n→∞
(3) (3) (3) (3)
uν1 ν2
,u ν3
,u , . . .: u ν1
(q3 ) → u∗ (q3 ) ∈ K
n→∞
..
.
(1) (2) (3) (n)
We take the diagonal sequence uν1 , uν2 , uν3 , . . . which satisfies uνn (q) → u∗ (q) for all q ∈ D.

2. Now, we’ll show the continuity of the limit u∗ : D → K.

(n) (n) |u (x) − u (y)|

uvn (x) − uvn (y) ≤ R |x − y|α R = sup <∞
x̸=y |x − y|α
u∈B
n → ∞: |u∗ (x) − u∗ (y)| ≤ R |x − y|α .

So u∗ is continuous. Hence u∗ has a unique continuous extensions to A.

64
 (n)
3. Now we want to show that uvn − u∗ → 0.
∞
ε
We do this by an “ 3 -argument”. Let ε > 0.

ε ε

1
(a) Choose δ > 0 such that δ α R ≤ 3 (i.e. δ ≤ 3R
α
)
(b) Choose L so large that {q1 , . . . , qL } such that dist ({q1 , . . . , qL } , x) ≤ δ ∀ x ∈ A
(n)
(c) Choose N so large that uvN (qi ) − u∗ (qi ) ≤ ε
3 ∀ i = 1, . . . , L, n ≥ N .

For any x ∈ A, choose q ∈ {q1 , . . . , qL } such that |q − x| ≤ δ. This gives us:

(n) (n) (n) (n)
uvn − u∗ (x) = uvn (x) − uvn (q) + ubn (q) − u∗ (q) + u∗ (q) − u∗ (x)
(n) (n) (n)
≤ uvn (x) − uvn (q) + ubn (q) − u∗ (q) + |u∗ (q) − u∗ (x)|
ε
≤ R |x − q| + + R |x − q|α ,
3
(n)
Where in the first absolute value, we used that uvn ∈ B, the second inequality holds for all
n ≥ N for N big enough and the third follows from step 2. With the right choice of {q1 , . . . , qL }
as in (b) and the right choice of δ, according to (a), we get:
ε
≤ Rδ α + + |R |x − q||α
3
ε ε
≤ + + |R |x − q||α
3 3
ε ε ε
≤ + + ∀ n ≥ N.
3 3 3

Remark (Arzela-Ascoli Theorem). Let A ⊂ Rn be compact. Then: B ⊂ C (A) relatively compact if

and only if the following two conditions are met:
(i) B is a bounded subset of C (A)

(ii) Elements of B are “equicontinuous”, that is:

∀ x0 ∈ A, ε > 0 ∃ δ (x0 , ε) : |f (x) − f (x0 )| < ε ∀ |x − x0 | < δ, ∀ f ∈ B.

δ is allowed to depend on x0 , but not on the function f . In our case, (ii) holds with δ =
 1
ε α
sup α ∥f ∥
, i.e. δ is independent of x0 .
f ∈C α

9 Duality in Hilbert Spaces

Example. We give an example of an element of X ∗ :

l (x) = ⟨x, η⟩

for some fixed vector η ∈ X.

But all elements of the Dual Space are of this form:

Theorem 9.1 (Riesz Representation Theorem). Let X be a Hilbert space and l ∈ X ∗ (i.e. l continuous
and linear map from X to K). Then there exists a unique η ∈ x such that

l (x) = ⟨x, η⟩ .

Abstractly, this means that i : X ∗ → X, l 7→ η is an isomorphism and in particular X ∗ ∼

= X. This
is not the case in general Banach spaces!

65
Definition 9.1. Let X be a Hilbert space and A ∈ L (X, X). For every y ∈ X, the map

ly (x) := ⟨Ax, y⟩ x∈X

is obviously a continuous linear map from X to K. By Riesz, there exists a unique element of X,
denoted by A∗ (y), such that
⟨Ax, y⟩ = ⟨x, A∗ (y)⟩ ∀ x ∈ X.
The map A∗ : X → X is called the adjoint operator of A.

At the moment, it is not clear if A∗ is linear or continuous.

Proposition 9.1. The operator A∗ : X → X is linear and continuous.

Proof. We first show linearity. Let λ, λ′ ∈ K, y, y ′ ∈ X. By the definition of A∗ (λy + λ′ y ′ ), we get:

x, A∗ λy + λ′ y ′ = Ax, λy + λ′ y ′ = λ ⟨Ax, y⟩ + λ′ Ax, y ′ = λ ⟨x, A∗ (y)⟩ + λ′ x, A∗ y ′

= x, λA∗ (y) + λ′ A∗ y ′ .

With η := A∗ (λy + λ′ y ′ ) and η ′ := λA∗ (y) + λ′ A∗ (y ′ ), we get ⟨x, η⟩ = ⟨x, η ′ ⟩ for all x ∈ X.
To conclude linearity, we still need a lemma:
Lemma 9.1. Let X be a Hilbert space and

⟨x, η⟩ = x, η ′ ∀ x ∈ X.

Then η = η ′ .

Proof. We have for all x:

0 = x, η − η ′
2
=⇒ 0 = η − η ′ , η − η ′ = η − η ′
=⇒ η − η ′ = 0.

In our case, it follows that

A∗ λy + λ′ y ′ = λA∗ (y) + λ′ A∗ y ′ .

Now to continuity. We have:

∥A∗ y∥ = sup ⟨x, A∗ y⟩

∥x∥=1

= sup ⟨Ax, y⟩ ≤ sup ∥Ax∥ ∥y∥

∥x∥=1 CSI ∥x∥=1

= ∥A∥ ∥y∥ ,
∥A∗ ∥y
so ∥A∗ ∥ = supy̸=0 y ≤ ∥A∥ < ∞ and in particular A∗ ∈ L (X, X).

Examples.
1. X = Rn . We take the matrix (Aij )ij ∈ Rn×n and get:
X X X   D E
⟨Ax, y⟩ = (Ax)i yi = Aij xj yi = xj A⊤ yi = x, A⊤ y ,
ji
i i,j j,i | {z }
=(A y )j
⊤

so we have A∗ = A⊤ . It’s adjoint is the transposed matrix and transposition is generalized by

the concept of adjoint.

66
 2. X = Cn with inner product ⟨x, y⟩ = Then we have A∗ = A⊤
P
i xi yi .

3. X = l2 with the right-shift A = S+ . We get:

⟨S+ x, y⟩ = ⟨(0, x1 , x2 , . . . ), (y1 , y2 , y3 , . . . )⟩
= x 1 y2 + x 2 y3 + . . .
= ⟨(x1 , x2 , . . . ), (y2 , y3 , . . . )⟩
= ⟨x, S− y⟩ ,
so the adjoint of the right-shift is the left shift and vice versa.
The main motivation for introducing the adjoint operator is to get a solubility criterion for linear
equations Ax = b.
Theorem 9.2 (Solvability criterion for Ax = b). Let X be a Hilbert space, A ∈ L (X, X) and b ∈ X.
Then:
(a) Ran A ⊆ (ker A∗ )⊥
(b) Ran A = (ker A∗ )⊥ .
In particular, if Ran A is closed, Ran A = (ker A∗ )⊤ .

Part (a) means that if Ax = b has a solution, that implies b⊥ ker A∗ . We therefore get a necessary
condition for solvability. With the “in particular”-part, this even becomes an equivalence.

Proof.
(a) Let y ∈ ker A∗ and b ∈ Ran A, i.e. there exists a solution x to Ax = b. Then:
⟨b, y⟩ = ⟨Ax, y⟩ = ⟨x, A∗ y⟩ = 0,
since y ∈ ker A∗ .
(b) See class.

Perturbation Theory (theorem 6.1) Duality (theorem 9.2)

Sufficient condition for solubility for all b (invert- Complete picture of solubility
iblity of A)
explicit solution (as series in inverse of nearby non-constructive
operator and perturbation) A−1 = A−1 0 +
(A − A0 )A−10 (A − A 0 ) + . . .
Proof of Theorem 9.1.
1. Analogously to the proof of the projection theorem, one shows that the function
1
F : X → R, F (x) := ∥x∥2 − Re l (x)
2
has a minimizer η (details: see class).
2. Show that the stationarity condition is precisely the Riesz representation equation.
Suppose that η ∈ X is a minimizer of F . Then ε = 0 is a minimizer of f : R → R, f (ε) =
F (η + εx′ ) and
d
0 = f ′ (0) = F η + εx′


ε=0
dε
ε2

d 1 2 ′ 2 ′


(9.1)
= ∥η∥ + ε Re x , η + ∥η∥ − Re l (η) − ε Re l x
dε ε=0 2 2
= Re x′ , η − Re l x′ ∀ x′ ∈ X.

67
 This is not quite 0 = ⟨x′ , η⟩ − l (x′ ). What should we do?
We use 1i x′ in place of x′ :
   
1 ′ 1 ′
0 = Re x , η − Re l x
i i
   
1 ′ 1 ′

(9.2)
= Re x , η − Re l x
i i
= Im x′ , η − Im l x′ ∀ x′ ∈ X,

1


where we used that Re i (a + bi) = Re (b − ai) = b = Im (a + bi). With (9.1) and (9.2) follows:

0 = x′ , η − l x′ ∀ x′ ∈ X.

3. For the uniqueness of η, we do the following: Suppose, ⟨x, η⟩ = l (x) = ⟨x, η ′ ⟩ for all x ∈ X.
Then by lemma 9.1, η = η ′ .

We will give two examples on the solubility criterion; the first to illustrate the role of ker A∗ , the
second to illustrate the subtle but in applications (e.g. differential equations) important phenomenon
that the condition of Ran A being closed cannot be ommited.

Examples.

1. X = l2 . We continue example 3 from before and consider the right-shift S+ with its adjoint S− .
We want to solve the equation S+ x = b.
We have:
( )
X 2
Ran S+ = (0, b2 , b3 , b4 , . . . ) : bi ∈ K, |bi | < ∞
i
∗


ker S+ = ker S− = {(b1 , 0, 0, 0, . . . ) : b1 ∈ K} .

By inspection,
∗

⊥
= b ∈ l2 : ⟨b1 = (1, 0, 0, 0, . . . ), b⟩ = 0 = Ran S+ .

ker S+
∗.
Hence, the equation is soluble if and only if b⊥ ker S+
2
2. We

x consider the solution operator to a first order differential equation: X = L ([0, T ]), (Kf ) (x) =
0 f (s) ds, which is the solution to the equation
(
u′ = f, in [0, T ]
.
u (0) = 0, else

We are interested in K ∗ .
First some heuristics: If u, v ∈ C 1 ([0, T ]), u (0) = u (T ) = 0 and v (0) = v (T ) = 0. Furthermore,
we have:  T      T
′ d d
u ·v = ,v = u, − v = − u · v′.
0 dx parts dx 0
d
So, ignoring that dx is not well defined on all of L2 ⫌ C 1 :
 ∗
d d
=− .
dx dx

68
 Hence, we should have  ∗

d −1
   −1
  d
= − .
 
dx dx
 
 
| {z }
=K, if we ignore bound. c.

So we expect K∗= −K. But what about the boundary condition u (0) = 0?
Now rigorously: Suppose in the beginning that f, g ∈ C ([0, T ]). We have:

u := K = antiderivative of f with u (0) = 0

e = antiderivative of g with v (T ) = 0, so that u (T ) ∨ v (T ) term below is zero
v := Kg
 T
⟨Kf, g⟩ = u, v ′ = u · v′
0
 T
= u (T ) v (T ) − u (0) v (0) − u′ · v
parts 0
D E
= f, −Kge (∗)

Since C ([0, T ]) is dense in L2 ([0, T ]) and K, K

e and the inner product are continuous, we have:
D E
⟨Kf, g⟩ = f, −Kg e ∀ f, g ∈ L2 ([0, T ])
=⇒ K ∗ = −K.
e

The adjoint of the solution operator and the differential equation

(
d
dx u = f, in [0, T ]
u (0) = 0, else

is the solution operator of the differential equation

(
d
− dx u = f, in [0, T ]
.
u (T ) = 0, else

Now, we consider the kernel of the adjoint: Suppose u =: K ∗ f = 0. Then we have

d
− u=f =0
dx
=⇒ u ≡ constant
u (T ) = 0 =⇒ u = 0,

hence ker K ∗ = 0.
If we knew that Ran K is closed, then theorem 9.2 would imply Ran K = X. But Ran K ⫋ X!
1
(e.g. since Ran K ⊆ C 2 ([0, T ]))
This is because the elements of Ran K, being antiderivatives of L2 functions, have “higher
regularity” than general L2 functions. This is a typical phenomenon for solution operators for
ordinary or partial differential equations.
But theorem 9.2 implies Ran K = X. This could be shown here directly, by noting that

Ran K ⊇ u ∈ C 1 ([0, T ]) : u (0) = 0 ,



with the observation that C 1 ([0, T ]) is dense in L2 .

We will now disucaa basic properties of adjoints.
Proposition 9.2. Let X be a Hilbert space over K, A, B ∈ L (X, X), and λ ∈ K.

69
 (i) (A∗ )∗ = A

(ii) (A + B)∗ = A∗ + B ∗

(iii) (λA)∗ = λA∗

(iv) (AB)∗ = B ∗ A∗

(v) ∥A∗ ∥ = ∥A∥.

Proof. See class. E.g. for (iv):

⟨x, (AB)∗ y⟩ = ⟨ABx, y⟩

= ⟨Bx, A∗ y⟩
= ⟨x, B ∗ A∗ y⟩ ∀ x, y.

The notion of adjoint leads to interestion subsets of L (X, X) in Hilbert spaces.

Definition 9.2 (Self-Adjoint, Skew-Symmetric, Unitary, Normal Operators). Let X be a Hilbert

space and A ∈ L (X, X). A is called:

ˆ self-adjoint (or hermitean or symmetric) if A∗ = A

ˆ skeq-symmetric if A∗ = −A

ˆ unitary if A∗ A = I = AA∗

ˆ normal if A∗ A = AA∗ .

Examples.

ˆ self-adjoint, skeq-symmetric, unitary all imply normal

ˆ S+ is not normal! S+ ∗ S+ = S− S+ = I but S+ S+ ∗ (x1 , x2 , . . . ) = S+ S− (x1 , x2 , . . . ) = (0, x2 , x3 , . . . ).

Proposition 9.3. Let X be a Hilbert space and A ∈ L (X, X). We have:

ˆ A self-adjoint =⇒ spec A ⊆ R

ˆ A skew-symmetric =⇒ spec A ⊆ iR

ˆ A unitary =⇒ spec A ⊆ S 1 = {λ ∈ C : |λ| = 1}.

This is trivial for eigenvalues; for the proof, see class.

10 Dual Spaces
Dual spaces are important in many problems, e.g. “weak convergence” (see section 11). By section 2,
(X ∗ , ∥·∥X ∗ ) is again Banach and its elements are called linear functionals.

b
A typical example is integration, where X = C ([a, b]) and Lf := a f . This operator is continuous,
since  b
f ≤ (b − a) ∥f ∥∞ = (b − a) ∥f ∥X ,
a

so ∥L∥X ≤ (b − a).
Our first goal is to determine the dual spaces of important Banach spaces.

Definition 10.1. Let (X, ∥·∥X ) and (Y, ∥·∥Y ) be two Banach spaces. They are called:

70
 1. isomorphic, with notation X ∼
= Y , if there exists a continuous isomorphism T : X → Y . (In this
case, by the Open Mapping Theorem, T −1 is also continuous.)

2. isometric, if in addition ∥x∥X = ∥T x∥X for all x ∈ X.

Theorem 10.1 (Dual Space of a Hilbert Space). Let X be Hilbert. Then

X∼
= X∗

and the map

T : X → X ∗, y 7→ ⟨·, y⟩
is an ismetric isomorphism.

Proof. We need to show that T is onto, injective and isometric.

Onto: This is clear by the Risz Representation Theorem.

Injective: Suppose T (y) = ⟨·, y⟩ = 0X ∗ . Then:

⟨x, y⟩ = 0 ∀x ∈ X
2
=⇒ ∥y∥ = ⟨y, y⟩ = 0
=⇒ y = 0.

Isometric: We have:
|⟨x, y⟩|
∥T (y)∥X ∗ = ∥⟨·, y⟩∥X ∗ = sup
x∈X\{0} ∥x∥
 
x
= sup ,y
x∈X\{0} ∥x∥
∥y∥2
 
±y
= ,y = = ∥y∥ .
∥y∥ ∥y∥

∗
In particular: l2 ∼
= l2 and L2 (Ω)

ˆ The only continuous functions L : l2 → R are

X
a 7→ ai bi
i

for some b ∈ l2

ˆ The only continuous linear functionals L : L2 (Ω) → R are


f 7→ f (x) g (x) dx
Ω

for some g ∈ L2 (Ω)

These two statements have interesting generalizations to lp and Lp (Ω).

1 1
Theorem 10.2. Let 1 ≤ p < ∞ with p + q = 1 and K = R. then:

(a) (lp )∗ ∼
= lq

(b) (Lp (Ω))∗ ∼

= Lq (Ω).

71
Moreover:
(a’) The map
∞
T : lq → (lp )∗ , (·)j yj ∈ (lp )∗ ,
X
y 7→ T (y) =
j=1
P∞
i.e. T (y) x = j=1 xj yj for all x ∈ lp , is an isometric isomorphism.

(b’) The map 

q p ∗
T : L → (L ) , g 7→ T (g) = (·) (x) g (x) dx,
Ω


i.e. T (g) f = Ωf (x) g (x) dx for all f ∈ Lp , is an isometric isomorphism.

Remarks.
P

(a) K = C is possible as well with T (y) = j (·)j yj and T (y) = Ω (·) (x) g (x) dx

(b) When p = 1, q = ∞ and the theorem applies.

(c) When p = ∞, q = 1 and the theorem fails (see next week).

∗ 3
Example. l3 ∼ = l 2 , since 13 + 23 = 1.

Proof.
(a) Here we need to prove the following:

1. T (y) = j (·)j yj ∈ (lp )∗ , when y ∈ lq , i.e. T : lq → (lq )∗ .

P

2. T is injective
3. T is onto
4. T is an isometry
P
1. The proof of this relies on Hölders Inequality: j xj yj ≤ ∥x∥p ∥y∥q for p, q as in the require-
ments.
Let x = (xj )j∈N ∈ lp and y = (yj )j∈N ∈ lq . Then:

X
|T (y) x| = x j yj ≤ ∥x∥p ∥y∥q
Hölder
j

|T (y) x|
=⇒ ≤ ∥y∥q ∀ x ∈ lp , y ∈ lq , x ̸= 0
∥x∥p
T (y) x
=⇒ sup ≤ ∥y∥q . (∗)
x∈lp \{0} ∥x∥p
| {z }
=∥T (y)∥(lp )∗

In particular, T (y) ∈ (lp )∗ .

2. Let T (y) = 0. We then need to show that y = 0. We have:
X
T (y) = ⇐⇒ xj yj = T (y) x = 0 ∀ x ∈ lp
j
X X
=⇒ yn = δnj yj = (en )j yj = T (y) en = 0 ∀n
j j

=⇒ y = 0

with en the n-th canonical unit vector.

72
 3. Let L ∈ (lp )∗ . We then need to show that there exists a y ∈ lq such that Lx =
P
j xj yj for all
x ∈ lp . Our idea for P
ths is to directly “guess” y, by using the basis vectors en .
Suppose that Lx = j xj yj for all x ∈ lp . Then:
X
Len = (ej )j yj = yn .
j

Hence define yn := L (en ). Claim: y = (yn )n∈N ∈ lq , Lx = T (y) x ∀ x ∈ lp , ∥y∥q ≤ ∥L∥(lp )∗ .

Let (
|yn |q−1 sgn yn , yn =
̸ 0
ỹn := .
0, else
We get:
N
X N
X N
X
|ỹn |p = |yn |p·(q−1) = |yn |q (10.1)
p·(q−1)=q
n=1 n=1 n=1
N N N N N
!
X X X X X
|ỹn |p = |yn |q = yn · y˜n = ỹn L (en ) = L ỹn en (10.2)
L lin.
n=1 n=1 n=1 n=1 n=1
 1
N p
X q
= L ((ỹ1 , . . . , ỹN , 0, . . . )) ≤ ∥L∥(lp )∗ · ∥(ỹ1 , . . . , ỹN )∥p = ∥L∥(lp )∗ ·  |yj |
j=1
1
N
! q N
! 1− p1
X q
X q
=⇒ |yn | = |yn | ≤ ∥L∥(lp )∗
n=1 n=1
N
! 1q
X q
=⇒ y ∈ lq , ∥y∥q = |yn | ≤ ∥L∥(lp )∗ . (∗∗)
n=1

By construction of yn , by taking finite linear combinations of L (en ) = yn , we know that

X
L = T (y) = (·)j yj
j

on Span {e1 me2 , . . . } = d, the space of sequences with only finitely many non-zero components.
From this follows with the continuity on lp of T (y) and L that L = T (y) on d, with respect to
∥·∥p . We have two cases: If 1 ≤ p < ∞, d is dense in lp ; if p = ∞, d is not dense in lp . Hence
for a ≤ p < ∞, L = T (y) on lp .
4. We have ∥T (y)∥(lp )∗ = ∥a∥q by T (y) and (∗) and (∗∗).

Remark. For p = 2, the above proof gives an alternative derivation of theorem 10.1 in the case
of l2 , by using the standard orthonormal basis of l2 instead of variational methods.

(b) “⊇” follows directly from Hölder, even for p = ∞:


g ∈ Lq =⇒ L : f 7→ fg
Ω


is linear and continuous on Lp , since Ω f g ≤ ∥f ∥p ∥g∥q .
For “ ⊆”, there is a measure theoretical proof using the Radon-Nikodym Theorem on differentia-
tion of measures [5, §II.2].
Here, we will give a functional-analytic proof, similar to the proof of the Riesz Representation
Theorem. Our idea is to find the representing g by minimizing some functional.
For p = 2, we can take I (f ) := 21 ∥f ∥2L2 − L (f ), for p ̸= 2, we take I (f ) := p1 ∥f ∥pLp − L (f ).

73
Claim: There exists a minimizing f . We will show this later.
Let f ∈ Lp be a minimizer. Then:

I (f + εφ) ≥ I (f + 0 · φ) ∀ ε ∈ R, φ ∈ Lp ,

which gives us

d d 1
0= ε=0
I (f + εφ) = ε=0
|f + εφ|p − L (f + εφ)
dε dε Ω p | {z }
L(f )+εL(φ) since L lin.

= |f + εφ|p−1 sgn (f · εφ) · φ − L (φ) ε=0
Ω
= |f |p−1 sgn f ·φ − L (φ) ∀ φ ∈ Lp .
Ω| {z }
=:g

Now we still need to check g ∈ Lq . We compute:

  
q (p−1)·q
|g| = |f | = |f |p < ∞,
Ω Ω Ω
p
where we used that f ∈ L and
1 1 1 1 p
+ = 1 ⇐⇒ q = 1 = p−1 = =⇒ q · (p − 1) = p.
p q 1− p 1
p−1

We still need to show the existence of a minimizer: In case of a Hilbert space, we used the
parallelogram equation, which is both elegant and obscure. But in fact, we used the “strict
convexity” of ∥·∥2 , which transfers to the case 1 < p < ∞.
To see this, we revisit the parallelogram equation:
 
∥f + g∥2 + ∥f − g∥2 = 2 ∥f ∥2 + ∥g∥2
2 2
f +g f −g ∥f ∥2 + ∥g∥2
⇐⇒ + =
2 2 2
2 2
∥f ∥2 + ∥g∥2 f +g f −g
⇐⇒ − = , (10.3)
2 2 2
which in particular is non-negative. We recall: F : V → R is convex if it satisfies

F ((1 − λ)a + λb) ≤ (1 − λ)F (a) + λF (b) ∀ λ ∈ [0, 1], a, b ∈ V.

For λ = 12 , we get in particular F a+b ≤ F (a)+F (b)

2 2 . Since result in (10.3) is non-negative, we
2
know that it is convex and the convexity is quantified by f −g
2 . To generalize the existence proof

2 to p ̸= 2, we just need some inequality that quantifies the amount of strict convexity
from p =
of f 7→ Ω |f |p .
Lemma 10.1 (Clarkson’s inequalities). Let f, g ∈ Lp (Ω). Then we have:
p p ∥f ∥pp +∥g∥pp
f +g f −g
(a) 2 ≤ p < ∞ =⇒ 2 + 2 ≤ 2
p p
q q  ∥f ∥p +∥g∥p q−1
1 1 f +g f −g p p
(b) 1 < p < 2, p + q = 1 =⇒ 2 + 2 ≤ 2
p p

Proof. This is elementary to check, but not to come up with. See [3].

Lemma 10.2 (Clarkson type inequality). Let 2 ≤ p < ∞. Then we have:

∥f ∥pp + ∥g∥pp f +g p
p f −g p
− ≥ .
2 2 p 2p−1 2 p

74
Remark. The constant here is non-optimal, compared to Clarkson, which for p = 2 is optimal.
But this is enough for our existence proof and easy to show.

Proof. It suffices to show:

p p
|a|p + |b|p f +g p a−b
− ≥ ∀ a, b ∈ R,
2 2 2p−1 2
since we just epply this with f (x) = a and g (x) = b and integrate over x.
W.l.o.g a < b. We introduce the mean m = a+b b−a
2 and d = 2 , so b = m + d and a = m − d. Let
p
F (t) = |t| . Then we have:
 
F (a) + F (b) a+b F (b) − F (m) F (m) − F (a)
−F = −
2 2 2 2
 m ′ ′
F (s + d) − F (s)
= ds
F.T. of C. a 2
 
1 m s+d ′′
= F (t) dt ds.
2 a s
In our case, F ′′ (t) = p(p − 1) |t|p−2 , so we get that
F ′′ (t) = p(p − 1) |t|p−2
 m d
1 2
≥ F ′′ (t) dt ds
2 a − d2
 m
1 d
= p |t|p−1 sgn t |−2 d ds
2 a 2
 m p−1
1 d
= 2p ds
2 a 2
p−1 p
d p p a−b
=d·p = |d|p = .
2 2p−1 2p−1 2

Proof of the Claim. We now use this for the proof of existence of minimizers of I (f ) =
∥f ∥p p
p − L (f ) for 2 ≤ p < ∞. We chose a minimizing sequence {fj }j∈N , fj ∈ L , i.e.

I (fj ) → inf I (f ) =: d.
j→∞ f ∈Lp

We now compare I (fj ) and fk :

1
d ← (I (fj ) + I (fk ))
2
min{j,k}→∞
   
1 1 p 1 p fj + fk
= ∥fj ∥p + ∥fk ∥p − L
2 p p 2
p
fj − fk p
 
1 fj + fk fj + fk p
≥ −L +
≥ p 2 p 2 2p−1 2 p
| {z  } | {z }
fj +fk quantifies strict convexity

=I 2
p
p fj − fk
≥d+ ,
2p−1 2 p
p
f −f ′
which gives us that 2 → 0 as min {j, k} → ∞, i.e. {fj } is Cauchy.
p
Since Lp is complete, the limit f of fj exists. Furthermore, I is continuous, since ∥·∥
is continuous and L is continuous by assumption. Therefore I (f ) = limj→∞ I (fj ) = d
and f is a minimizer of I.

75
 Our next goal is to formulate general statements about the dual of a Banach space, e.g. at the
moment it is not clear, that it contains any non-zero elements.
Theorem 10.3 (Hahn-Banach Theorem). Let X be a real (or complex) vector space, U a subspace
of X, p : X → R sublinear (i.e. linear scalar multiplication but sublinear addition), and l : U → R (or
C) linear and suppose Re l ≤ p on U .
Then there exists L : X → R (or C) such that

L/U = l, Re L ≤ p

on X.

Corollary 10.1 (Hahn-Banach for norms). Let X be a normed K-vector space, U a subspace and
l : U → K a continuous linear functional. Then there exists a continuous linear functional L : X → K
with ∥L∥ = ∥l∥ and L/U = l.

This means that every continuous linear functional on a subspace can be extended with the same
norm onto the whole space.

Proof. We take p (x) := ∥l∥ ∥x∥, which is homogeneous and sublinear. Recall that on U , we have
|l (x)|
∥l∥ = sup =⇒ |l (x)| ≤ ∥l∥ · ∥x∥ = p (x) .
x∈U \{0} ∥x∥

By theorem 10.3 we have that there exists L : X → R linear with l = L on U , such that

L (x) ≤ p (x) = ∥l∥ ∥x∥

on X. We apply this to −x in place of x and get

−L (x) = L (−x) ≤ p (−x) = ∥l∥ ∥−x∥ = ∥l∥ ∥x∥

=⇒ |L (x)| ≤ ∥l∥ ∥x∥ on X
=⇒ ∥L∥ ≤ ∥l∥ .

Since l = L on U , “≥” is trivial.

Proof of Hahn-Banach. We only state the proof for K = R. Our idea is to set co dim U = 1 and use
induction (but note that in general, X has uncountably many further dimensions than U ).
Step 1 Let co dim U = dim X/U = 1 and choose x0 ∈ X/U . We then get that for all x ∈ X, we can
write x = u + λx0 with x ∈ U and x0 ∈ R.
Our Ansatz will be to write L (x) = l (u) + r · λ where we can freely choose r ∈ R. Obviously,
L : X → R is linear and L = l on U .
We need to prove that L ≤ p on X. We get:

l (u) + rλ ≤ p(u + λx0 ) ∀ u ∈ U, λ ∈ R.

We assume λ > 0 and get:

u  u
⇐⇒ r ≤ p + x0 − l
λ λ
⇐⇒ r ≤ inf (p (v + x0 ) − l (f )). (10.4)
v∈U

Now we assume λ < 0:

   
u u
⇐⇒ −r ≤ p − x0 − l ∀u ∈ U
−λ −λ

76
    
u u
⇐⇒ r ≥ l −p − x0
−λ −λ
⇐⇒ r ≥ sup (l (w) − p (w − x0 )). (10.5)
w∈U

We thus have an upper and lower bound on r. If the upper bound is greater than the lower
bound, there actually exists a solution, i.e.

l (w) − p (w − x0 ) ≤ p (v + x0 ) − l (v) ∀ v, w ∈ U
⇐⇒ l (v) + l (w) ≤ p (v + x0 ) p (w − x0 ) .

But

l (v) + l (w) = l (v + w) = p (v + w) = p ((v + x0 ) + (w − x0 ))

| {z } l≤p on U
∈U
≤ p (v + x0 ) + p (w − x0 ) .
p sublinear

Step 2 Here, we use a result from set theory:

Lemma 10.3 (Zorn’s Lemma). Let (A, ≤) be a partially ordered set, in which each totally
ordered subset has an upper bound. Then A has a maximal element, i.e. there exists m ∈ A
such that there is no m′ with m′ > m.

We apply this to:

V subspace of X with V ⊇ U ,
 
A= (V, LV ) : LV p-bounded extension of l to V , i.e. LV : V → R linear, ,
LV U = l, LV ≤ p V

which is the set of solutions of the extension problem onto subspaces V with U ⊆ V ⊆ X.
We take the partial ordering on A:

(V1 , LV1 ) ≤ (V2 , LV2 ) ⇐⇒ V1 ⊆ V2 , LV2 V1

= LV1 .

This is not a total ordering, since there can exist V1 , V2 with V1 ̸⊆ V2 and V2 ̸⊆ V1 .
Furthermore, A is non-empty, since (U, l) ∈ A.
Let {(Vi , LVi )}i∈I be totally ordered with index set I.
Claim: There exists an upper bound.
S
Proof of the Claim. Take V := i∈I Vi and LV := LVi for all ∈ Vi . Then (Vi , LVi ) ≤
(V, LV ) for all i. By Zorn, there exists a maximal element (X0 , LX0 ).

Claim: X0 = X
 
Proof of the Claim. Suppose X0 ⫋ X. By Step 1, there exists X e0 , L e such
X0
 
that X0 , LXe0 > (X0 , LX0 ) with dim X0 /X0 = 1, which is a contradiction to
e e
(Xo , LX0 ) being maximal.

Corollary 10.2. Let X be a normed vector space. Then, for all x ∈ X \ {0}, there exists L ∈ X ∗
such that
∥L∥ = 1, L (x) = ∥x∥ ,
i.e. L(y) ∗
∥x∥ is maximized at x. In particular, for all x1 ̸= x2 , there exists L ∈ X such that L (x1 ) ̸=
L (x2 ).

77
 D E
x
In Hilbert spaces, this is trivial, since for L = ·, ∥x∥ we get Lx = ∥x∥.

Proof. Let l : Span {x} → K and l (λx) := λ ∥x∥. Then l (x) = ∥x∥ and ∥l∥ = 1. L is an extension of
l to X with the same norm. In particular, L with Lx1 − Lx2 = L (x1 − x2 ) = ∥x1 − x2 ∥.

Corollary 10.3 (Dual Characterization of the Norm). Let X be a normed vector space. Then

|Lx|
∥x∥ = sup .
L∈X ∗ \{0} ∥L∥

|Lx|
Compare: ∥L∥ = supx∈X\{0} ∥x∥ .

Proof. We show the two inequalities:

|Lx| |Ly|
≤ sup = ∥L∥ ∀ L.
∥x∥ y∈X\{0} ∥y∥

For the other inequality, we choose the L from corollary 10.2 and get

|Lx| ∥x∥
= .
∥L∥ 1

Corollary 10.4. Let X be Banach, U a closed subspace and x0 ∈ X \ U . Then there exists L ∈ X ∗
with L U = 0 and L (x0 ) ̸= 0.

Remark. This corollary means that X ∗ separates points from closed subspaces.
For Hilbert spaces, this is easy to show, since we take L = ·, x0 ⊥ where x0 ⊥ is the part of x0 in U ⊥ .

Proof. Let
ω : X → X/U, x 7→ {x + u : u ∈ U }
be the quotient map. (In the Hilbert case, X/U ∼
= U ⊥ .) We have:

ω (0) = 0X/U ∀u ∈ U
ω (x0 ) ̸= 0X/U .

We choose l : Span {ω (x0 )} → K with l (ω (x0 )) ̸= 0. By Hahn-Banach there exists a continuous linear
extension L
e of l onto X/U . Define

e (ω (x)) , L : X → K
L (x) := L
e (ω (x0 )) = l (ω (x0 )) ̸= 0
L (x0 ) = L by construction


L (u) = Le (ω (u)) = Le 0X/U = 0 ∀u ∈ U
L
e lin.

This result is useful in convex analysis.

Corollary 10.5 (Existence of closed complements). Let X be Banach and U a finite-dimensional

subspace. Then there exists a closed complement of U (i.e. a closed subspace V such that X =
U ⊕ V, U ∩ V = {0}).

Remark. One cannot replace “finite-dimensional” with “closed”

Outline of the Proof (details: see class).

78
 1. Choose basis {e1 , . . . , en } of U . Consider “dual basis” {L1 , . . . , Ln } of U ∗ , i.e. Li : U →
K, Li (ej ) = δij .

2. By Hahn-Banach, there exists a continuous extension Li : X → K.

3. Define Px := ni=1 Li (x) ei . Then P : X → X is continuous and linear and we get Ran P = U
P
and P U = I, since P (ei ) = ei for all i (check this; write x = P x + (I − P )x).

4. Verify that P 2 = P , X = U ⊕ ker P and ker P is closed.

So ker P is the sought complement of U .

11 Weak Convergence
Remark (Mathematician’s Motivation). An important result in Rn , which fails in every infinite-
dimensional Banach space is the Bolzano-Weierstrass theorem (bounded sequences have convergent
subsequences, see section 1). √ We give an elementary example in l2 : If we take the unit vectors
e1 , e2 , . . . , we have ∥ei − ej ∥ = 2 for i ̸= j, but obviously there is no convergent subsequences.
Those, however, are often useful. Our way out will be a weaker notion of convergence, to restore this
theorem.

Remark (Physicist’s/Engineer’s/Statistician’s Motivation).

Definition 11.1. Let X be a Banach space and X ∗ the dual space of X.

1. A sequence {xj }j∈N in X converges weakly to x ∈ X with notation xj ⇀ x, if

L (xj ) → L (x) ∀ L ∈ X ∗.

∗
2. A sequence {Lj }j∈N in X ∗ converges weak∗ to L ∈ X ∗ , with notation Lj ⇀ L if

Lj (x) → L (x) ∀x ∈ X

Remark. Weak∗ convergence is pointwise convergence of linear functionals.

Proposition 11.1 (Justification of the Name “weak”). Under the assumptions of the above definition,
we have:

(i) xj → x (i.e. ∥xj − x∥ → 0) =⇒ xj ⇀ x.

∗
(ii) Lj → L (i.e. ∥Lj − L∥X ∗ ) =⇒ Lj ⇀.

Proof.

(i) xj → x, L ∈ X ∗ (i.e. L continuous) and therefore L (xj ) → L

(ii) We have:

|Lj x − Lx|
Lj → L ⇐⇒ ∥Lj − L∥X ∗ = sup →0
def. x̸=0 ∥x∥
=⇒ |Lj x − Lx| → 0 ∀ x,

i.e. Lj x → Lx.

The converse is false:

79
Example 11.1. X = l2 . We again take unit vectors

e1 = (1, 0, 0, 0, 0, . . . )
e2 = (0, 1, 0, 0, 0, . . . )
e3 = (0, 0, 1, 0, 0, . . . )
..
.

Viewed as functions on N, obviously the en converge pointwise to zero, i.e. (en )j → 0 for n → ∞ and
fixed j. But the map en 7→ (en )j is a continuous linear functional L : X → R. Thus, certain linear
functionals applied to the sequence converge to zero.
Claim: For all L ∈ X ∗ , L (en ) → 0 for n → ∞. In other words, en ⇀ 0.
Proof of the Claim. Fix L ∈ X ∗ . By Riesz, there exists a = (a1 , a2 ,P . . . ) ∈ l2 such that
L = ⟨·, a⟩. Hence L(en ) = ⟨en , e⟩ = an → 0 for n → ∞, since ∥a∥ = ∞ 2 2
j=1 |aj | < ∞ and
the elements of a convergent series have to be a null sequence.
√
But en ̸→ 0, since ∥en ∥ = 1, it does in fact not converge at all, since ∥en − em ∥ = 2 for m ̸= n.
Example 11.2 (Fundamental). X = Lp ([0, 1]) with 1 ≤ p < ∞. un
ˆ takes only two values, A and B

ˆ is periodic with period 1

n, hence oscillates fast for large n

ˆ is bounded in supremum norm, hence it is bounded in Lp (0, 1) for 1 ≤ p < ∞

ˆ is not convergent in Lp (we will prove this later)

ˆ λ ∈ (0, 1) is the volume of the set on which un takes the value A.

Claim: un ⇀ u with u (x) = λA + (1 − λ)B

Proof of the Claim. We need to show:

 1  1
1 1
un (x) g (x) dx → (λA + (1 − λ)B)g (x) dx ∀ Lq ([0, 1]) , + = 1.
0 0 q p
Our strategy will be to show this for increasingly
 general g.

1 1
Claim 1: 0 un 1[0,b] → (λA + (1 − λ)B) 1[0,b]
| 0 {z }
=:b

Proof of Claim 1. We have:

 1  b
un 1[0,b] = un (x) dx
0 0
 bn  b
= un + un ,
0 bn

k
where in the last equality, we choose bn to be the largest element such that n = bn ≤ b ≤
b + n1 with k ∈ N0 . We then get for the first term:
 bn  1
n
un = k ·
0
0 
λ 1−λ
=k· A+
n n
k →
= (λA + (1 − λ)B)n → ∞b(λA + (1 − λ)B)
n

80
 and for the second term:
 b
un ≤ ∥un ∥∞ (b − bn )
bn
1
≤ max {|A| , |B|} → 0.
n n→∞

1
1
Claim 2: 0 un 1[a,b] → (λA + (1 − λ)B) 0 1[a,b] = (λA + (1 − λ)B)(b − a).
Proof of Claim 2. We have:
 1  1  1
un 1[a,b] = un 1[0,b] − un 1[0,a] → b(λA + (1 − λ)B) − a(λA + (1 − λ)B)
0 0 0 Claim 1

1
1
Claim 3: 0 un ni=1 αi 1[ai ,bi ] → ni=1 αi (λA + (1 − λ)B)(bi − ai ) = 0 u · ni=1 αi 1[ai ,bi ] .
P P P
This follows directly from Claim 2.

1
1
Claim 4: 0 un g → 0 ug for all g ∈ Lq ([0, 1]).

Proof of Claim 4. By measure theory, we know that step functions are dense in Lq for
1 ≤ q < ∞. That is for all g ∈ Lq , ε > 0 there exists a step function gε ∈ Lq such that
∥g − gε ∥q < ε. Thus
 1  1  1  1
un g − ug = (un − u)gε + (un − u)(u − gε ).
0 0 0 0
| {z }
→0 by Step 3

For the second term, we get:

 1
(un − u)(u − gε ) ≤ ∥un − u∥p ∥g − gε ∥q ≤ max {|A| , |B|} · ε
0 Hölder

and therefore  
1 1
lim sup un g − ug ≤ ε max {|A| , |B|} .
n→∞ 0 0

1
1
Since ε > 0 was arbitrary, lim supn→∞ 0 un g − 0 ug → 0 for n → ∞. Hence un ⇀ u in
Lp .

Remark (Interpretation). The weak limit is the local spatial average of the function values.

Claim 5: un does not converge in Lp , in particular un ̸→ u in Lp .

Proof of Claim 5. We suppose that un → ũ for some ũ in Lp . Then by proposition 11.1

un ⇀ ũ, hence by proposition 11.2 ũ = u. But we have:
 1  p1
p
∥un − u∥p = |un (x) − (λA + (1 − λ)B)| dx
0
 1  p1
p
≥ min {|A − (λA + (1 − λ)B)| , |B − (λ − A + (1 − λ)B)|}
0
= min {|A − (λA + (1 − λ)B)| , |B − (λ − A + (1 − λ)B)|}

81
Proposition 11.2 (Properties of Weak Convergence). Let X be a Banach space, xj ∈ X and
x, x′ ∈ X.

a) If xj ⇀ x and xj ⇀ x′ , then x = x′ (uniqueness of weak limits)

b) If xj ⇀ x, then {xj }j∈N is bounded

c) If xj ⇀ x, then ∥x∥ ≤ lim inf j→∞ ∥xj ∥ (weak lower semi-continuity of the norm)

d) If xj ⇀ x and xj ∈ U with U ⊆ X a closed subspace, then x ∈ U .

Example (Regarding c)). X = l2 . We know that en ⇀ 0 but ∥en ∥ = 1 > 0 = ∥0∥, so the limit has
smaller norm.

Example (Regarding d)). Here you see that “closed subspace” cannot be replaced by “closed set”:
X = l2 , U = {x ∈ X : ∥x∥ = 1}. Then en ∈ U but 0 ∈
/ U.

Proof.

a) Let x ↼ xj ⇀ x′ . Then L (x − x′ ) = limj→∞ L (xj − xj ) = 0 for all L ∈ X ∗ . If X is Hilbert, we

pick L = ⟨·, x − x′ ⟩ and get L (x − x′ ) = ∥x − x′ ∥2 = 0 and thus x = x′ .
If X is Banach, we have by corollary 10.2 L ∈ X ∗ with L (x − x′ ) = ∥x − x′ ∥, so again x = x′ .

b) Here, our idea is to use the Banach-Steinhaus theorem.

Consider the evaluation functionals (xn )ev : X ∗ → K, (xn )ev (L) := L (xn ). Then xn is weakly
convergent if and only if (xn )ev (L) is convergent for all L ∈ X ∗ which implies that (xn )ev is
bounded for all L ∈ X ∗ .
Hence (xn )ev is a pointwise bounded sequence of linear functionals on X ∗ . By Banach-Steinhaus
we now get
|(xn )ev (L)|
∥(xn )ev ∥L(X ∗ ,K) = sup
L∈X ∗ \{0} ∥L∥
bounded. But
|(xn )ev (L)| |L (xn )|
sup = sup = ∥xn ∥ .
∗
L∈X \{0} ∥L∥ L∈X \{0} ∥L∥
∗ cor. 10.3

c) If X is Hilbert, this is elementary: Suppose xj ⇀ x. Then we have:

0 ≤ ∥x − xj ∥2 = ∥x∥2 − 2 Re ⟨xj , x⟩ + ∥xj ∥2

⇐⇒ 2 Re ⟨xj , x⟩ − ∥x∥2 ≤ ∥xj ∥2
=⇒ 2 Re ⟨x, x⟩ − ∥x∥2 ≤ lim inf ∥xj ∥2 .
| {z } j→∞
=∥x∥2

If X is Banach, the proof becomes less elementary, but shorter, using Hahn-Banach.
Suppose xn ⇀ x and L ∈ X ∗ . We then have:

|L (x)| = lim |L (xn )| ≤ ∥L∥ lim inf ∥xn ∥ . (∗)

n→∞ | {z } n→∞
≤∥L∥∥xn ∥

∗
By Hahn-BanachDwe get E that there exists L ∈ X such that L (x) = ∥x∥ , ∥L∥ = 1 (if X is Hilbert,
x
we take L (x) = ·, ∥x∥ ). We substitute that into (∗) and get:

∥x∥ ≤ 1 lim inf ∥xn ∥ .

n→∞

82
d) Suppose xj ∈ U with U ⊆ X a closed subspace and xj ⇀ x with x ∈ / U.
⊥
If X is Hilbert, we take L = ·, x where x is the component of x in U ⊥ (which exists by the
⊥

Projection Theorem). We then have:

D E D E
0 = xj , x⊥ = L (xj ) → L (x) = 0 = x, x⊥ = ∥x∥2

and thus x⊥ = 0 and x ∈ U which is a contradiction.

If X is Banach, we again use Hahn-Banach.
This tells us that there exists L ∈ X ∗ such that L U = 0 with L (x) ̸= 0 (“X ∗ separates points
and closed subspaces”). This has the same properties as the explicit L from the Hilbert case. We
therefore repeat the argument:
0 = L (xj ) → L (x) ̸= 0
xj ∈U,
L =0
U

by construction of L, which is a contradiction.

Remark (On d)). One can show: For xj ∈ A and A ⊆ X closed and convex, we have

xj ⇀ x =⇒ x ∈ A.

Example 11.3. A closed subspace (see proposition 11.2 d)).

Example 11.4. A = {x ∈ X : ∥x∥ ≤ r} a closed ball (see proposition 11.2 c)). This belongs to the
field of convex analysis.

Now our goal is to restore Bolzano-Weierstrass using weak or weak∗ convergence (i.e. we would
like to show that bounded sequences have weak or weak∗ convergent subsequences).

Definition 11.2. A Banach space is called reflexive, if every continuous linear functional on X ∗ is
of the form
f (L) = L (x)
for some x ∈ X and all L ∈ X ∗ (i.e. if every continuous linear functional on X is an “evaluation
functional” xev for some x ∈ X). We use the informal notation X ∗∗ ∼
= X.
Remark.

X∼
= X ∗ Hilbert, very nice
X∼
= X ∗∗ reflexive Banach, still nice
X∼
̸ X ∗∗ nasty
=
X∼
= X ∗∗∗ again very nice, since X ∗∗∗ ∼
= X ∗.

Space Separable Reflexive

X Hilbert No Yes

c0 , l1 Yes No

lp , 1 < p < ∞ Yes Yes

l∞ No No

C (Ω), Ω ⊆ Rn ; L1 (Ω), Ω ⊆ Rn measurable Yes No

Lp (Ω), Ω ⊆ Rn measurable, 1 < p < ∞ Yes Yes

L∞ (Ω), Ω ⊆ Rn measurable No No

83
 n o
Remark. L∞ ((0, 1)) is not separable, since 1(0,a) is uncountable but 1(0,a) − 1(0,a′ ) ∞
=1
a∈(0,1)
for a ̸= a′ .
Furthermore, lp , 1 < p < ∞ is reflexive, since
1 1
(lp )∗ ∼
= lq + =1
q p
∼ 1 1
(lp )∗∗ = (lq ) ∼
∗
= lr + =1
r q

and thus p = r and (lp )∗∗ ∼

= lp . The same argument works for Lp with 1 < p < ∞.
We also have:
l1 ∗ ∼
= l∞ , but (l∞ ) ̸∼
∗
= l1 ,

which we will see in class.

Theorem 11.1 (Banach-Alaoglu Theorem). Let X be a separable Banach space. Then every bounded
sequence in X ∗ has a weak∗ convergent subsequence.

Theorem 11.2 (Weak Sequential Compactness). Let X be a Banach space. Then if X is reflexive,
every bounded sequence in X has a weakly convergent subsequence.

Remark. In theorem 11.2 the converse is also true, which is called the Eberlein-Shmulyan theorem.

Example (Bounded sequence which does not have a weakly convergent subsequence).

X = L1 ([0, 1]).
We take uε (x) = 1ε 1([0,ε]) (x) with ε = 31j , j ∈ N. Then we have ∥uε ∥1 = 0 uε = 1.
1

∗
Suppose there exists a weakly convergent subsequence uεn ⇀ u in L1 . Since L1 ∼ = L∞ , it follows
that   1 1
uεj f → uf ∀ f ∈ L∞ ([0, 1]) .
0 0
We define f such that it takes the values 1 and −1, changing over at points ε′j = εnj (see also figure
??). We then have:

 1  ε′j  ε′j+1  ε′j

1 1 1
uε′j f = f= f+ f .
0 ε′j 0 ε′j 0 ε′j ε′j+1
| {z }
=(−1)j (ε′j −ε′j+1 )

ε′
The first summand has an absolute value of no less than j+1 ε′j
≤ 13 and the second summand can be
ε′
 
written as (−1)j 1 − j+1
ε ′ = (−1)j + bj where |bj | ≤ 13 . Thus we get:
j

= (−1)j + aj ,
1
where |aj | ≤ 3 + 13 . This does not converge.

Proof of Theorem 11.1. Let {Lj }j∈N be a bounded sequence in X ∗ . Since X is by assumption separa-
ble, there exists a countable dense subset {x1 , x2 , . . . } of X. Our Idea is to first construct a sbsequence
such that limj→∞ Lj (xk ) exists for all xk .
Claim 1: {Lj (xk )}j∈N with k fixed is a bounded sequence in K.

Proof of Claim 1. We have:

|Lj (xj )| ≤ ∥Lj ∥ ∥xj ∥ .

| {z } |{z}
bounded by assumption fixed

84
 By Bolzano-Weierstrass in K there exists a convergent subsequence {Ljn (xk )}n∈N . By the
Diagonal Method (see e.g. the proof of compactness of bounded sets in C 0,α in C) we have
subsequences such that Ljn (xk ) that are convergent for all k as n goes to infinity.
We define on {x1 , x2 , . . . }:
L (xk ) := lim Ljn (xk ) .
n→∞

Claim 2: L : Span {x1 , x2 , . . . } → K is linear.

Claim 3: L is continuous.

Proof of Claim 2. The Ljn are linear, thus the limit commutes with linear combinations.

Proof of Claim 3. We have:

|L (xk )| = lim |Ljn (xk )| ≤ sup ∥Lj ∥ · ∥xk ∥ .

n→∞ j
| {z }
=:M <∞

We now need a lemma to continue with the proof:

Lemma 11.1 (Continuous Extension of Linear Operators). Let D ⊆ X be a dense subspace of a

normed vector space X, Y a Banach space and T ∈ L (D, Y ). Then there exists a unique Te ∈ L (X, Y )
with Te D = T .

Remark. Te is called the continuous extension of T to X.

Proof. The key point is to define Te (x) for x ∈ X. Let x ∈ X. We then take a sequence {xj } in D
with xj → x.
Since {xj } is convergent, it is Cauchy and since T is continuous on D, {T (xj )} is Cauchy as well. Y
is Banach and thus {T (xj )} is convergent.
We define Tex := limj→∞ T (xj ) and it is not difficult to check that Te has the has the asserted
properties.

So now, by lemma 11.1 there exist a unique continuous extension

L : Span {x1 , x2 , . . . } → K.
| {z }
=X

We still need to show that Ljn (x) → L (x) for all x ∈ X, not just for the x in Span {x1 , x2 , . . . }.
Our idea is to approximate x by xk ∈ {x1 , x2 , . . . }.
For x ∈ X and ε > 0, there exists xk ∈ {x1 , x2 , . . . } such that ∥xk − x∥ < ε. We get:

|Ljn (x) − L (x)| = |Ljn (x) − Ljn (xk ) + Ljn (xk ) − L (xk ) + L (xk ) − L (x)|
≤ |Ljn (x) − Ljn (xk )| + |Ljn (xk ) − L (xk )| + |L (xk ) − L (x)|
≤ sup ∥Ljn ∥ · ∥x − xk ∥ + |Ljn (xk ) − L (xk )| + ∥L∥ ∥xk − x∥ .
n | {z } |{z}
| {z } →0 <∞
<∞

We thus get:  
lim sup |Ljn − L (x)| ≤ sup ∥Ljn ∥ + ∥L∥ · ε.
n→∞ n

Since ε > 0 was arbitrary, |Ljn − L (x)| → 0 for n → ∞, which was what we needed to show. Thus
∗
Ljn ⇀ L in X ∗ .

85
Proof of Theorem 11.2 for X separable. 1. Using the definitions, we get:

xj ⇀ x ∈ X ⇐⇒ (xj )ev = L (xj ) → L (x) = xev (L)

∗
⇐⇒ (xj )ev ⇀ xev ∈ (X ∗ )∗ ,

where xev : X ∗ → K is the evaluation functional xev (L) = L (x).

2. We furthermore have xj bounded in X ⇐⇒ (xj )ev is bounded in X ∗∗ , since

|yev (L)| |L (y)|

∥yev ∥ = sup = sup = = ∥y∥ ,
L∈X ∗ \{0} ∥L∥ L∈X ∗ \{0} ∥L∥ 10.1

i.e. X and X ∗∗ are even isometric.

3. For the next step we need a lemma:

Lemma 11.2. Let Y be a Banach space and Y ∗ separable. Then Y is separable.

Proof. See class.

We apply this to Y = X ∗ and get that if X ∼ = X ∗∗ is separable, so is X ∗ . xj is bounded in X

and thus by 2. (xj )ev is bounded in X . By Banach-Alaoglu, which is applicable since X ∗ is
∗∗
∗
separable, there exists (xjn )ev ⇀ L in X ∗∗ , which is xev for some x ∈ X, since X is reflexive.
By 1. we get that xjn ⇀ x.
When X is not separable, the theorem is a statement about sequences {x1 , x2 , . . . }. Every sequence
automatically lies in a separable Banach space, namely

U := Span {x1 , x2 , . . . }.

Since xj is a bounded sequence in X, it is a bounded sequence in U .

Lemma 11.3. Let X be a reflexive Banach space and U a closed subspace. Then U is reflexive.

Remark. Compare: When X is Hilbert with closed subspace U , U is Hilbert.

Proof. Ommited.

By theorem 11.2 in the separable case, there exists a subspace xjn ⇀ x ∈ U in U . That is,
n→∞

L (xjn ) → L (x) ∀ L ∈ U ∗.

e ∈ X ∗ . Claim: L
Let now L e ∈ U ∗.
U

Proof of the Claim. We know that L

e
U
: U → K is linear and:

L
e (u) L
e (x)
L
e
U
= sup ≤ sup e <∞
=: L
u∈U \{0} ∥u∥ x∈X\{0} ∥x∥
L
e (xjn ) = L
e (xjn ) → L
e (x) = L
e (x) e ∈ X ∗,
∀L
xjn U xjn ⇀x∈U U x∈U

so xjn ⇀ x ∈ X.

86
Substantial Examples for Weak Convergence
Example 11.5 (Trick-Free Existence Proof for Riesz Representation Theorem and Lp Duality). For
Riesz, we have the following: Let X be a Hilbert Space and L ∈ X ∗ . Then there exists a minimizer
2
x0 of F (x) = ∥x∥
2 − L (x) (implying that L (x) = ⟨x, x0 ⟩ for all x). We can then do an easy proof of
existence via weak convergence (see class).
For Lp (Ω), we have that for X = Lp (Ω) with Ω ⊆ Rn measurable and L ∈ X ∗ there exists a minimizer
∥f ∥p
of I (f ) = p p − L (f ) (where if f is a minimizer and g := |f |p−1 sgn f ∈ Lq for 1
+ 1
= 1 we have

q p
that L (φ) = Ω ϕg for all ϕ).
To prove the existence of a minimizer, we do the following:

1. Choose a minimizing sequence {fj } of I

∥f ∥pp
2. I (f ) ≥ p − ∥L∥ ∥f ∥p → ∞ for ∥f ∥p → ∞, so {fj } are bounded.

3. Lp is reflexive and has weak sequential compactness (theorem 11.2) (we only need the separable
case). This implies that there exxists a weakly convergent subsequence fjn ⇀ f ∈ Lp
n→∞

4. By weak lower semi-continuity of the norm, we have:

I (f ) ≤ lim inf I (fjn ) inf I,

n→∞ fj minimizing sequenceX

so f is a minimizer.

This is a template for very general existence proofs in mathematics, e.g. Partial Differential Equations,
Convex Analysis, Optimization etc.

Example 11.6 (Paul Dirac, 1930, Nobel Prize Winner in Physics). We define δ : R → R such that:

δ (0) = ∞

δ (x) = 0, x ̸= 0 .



Rδ = 1

The orthodox

mathematician would say that this cannot exist, since f (x) = 0 almost everywhere and
therefore R f = 0 (if the integral is defined in the usual way).

Consequences of the Above (According to Dirac)

(A) R δ (x) f (x) dx = f (0) for f continuous


(B) If φ : R → R continuous with φ = 0 outside of [−1, 1] and R φ = 1, then 1ε φ ε· “converges to”

δ.

Dirac sequence; note that if φ > 0 in (−1, 1), then φε (0) → ∞, φε (x) → 0
This is a so called
(for x ̸= 0) and φε = 1 for all ε.

Functional Analytic Viewpoint

1. Delta is not a function, but a linear functional on Cb (R), which is the space of bounded continuous
functions on R, i.e.
φ : Cb (R) → R, δf := f (0) .
δ is continuous, since |δf | = |f (0)| ≤ supx∈R |f (x)| = ∥f ∥∞ , so δ ∈ Cb (R)∗ .

2. Let’s now take a look at the Dirac sequences: The 1ε φ ε· can also be viewed in a natural way

as elements of Cb (R)∗ . Define for ψ ∈ L1 (R)


Lψ : Cb (R) → R, Lψ f := ψf.
R

87
 Lψ is the
functional “integrate against ψ”. It is in Cb (R)∗ , since it is obviously linear and
|Lψ f | = ψf ≤ ∥ψ∥1 ∥f ∥∞ .
∗
Claim: L 1 φ( · ) ⇀ δ in Cb (R)∗ .
ε ε

Proof of the Claim. Let f ∈ Cb (R). We then have:

  ε
1 x 1 x
L 1 φ( · ) = φ f (x) dx + φ f (x) dx
ε ε ε ε −ε ε ε
R ε  ε
1 x 1 x
= φ f (0) dx + φ (f (x) − f (0)) dx
−ε ε ε −ε ε ε

ε
| {z }
=f (0) −ε φ( xε ) dx=f (0)·1
 ε
1 x
≤ f (0) + φ dx · sup |f (x) − f (0)|
−ε ε ε x∈[−ε,ε]
| {z } | {z }
=1 →0 for ε → 0 since f continuous

and thus L 1 φ( · ) f → f (0) = δf for ε → 0.

ε ε

Summary

Interpret the participating “functions” δ and L 1 φ( · ) as elements of a suitable dual space. Then
ε ε
(B) is valid as weak∗ convergence. Even though the approximating functionals L 1 φ( · ) have an

ε ε
integral representation of form Lψ = ψ ·
with ψ ∈ L1 , this is not thrue for the limit functional,
i.e. there exists no ψ ∈ L1 such htat δ = ψ ·.

Outlook: δ shows that C ([a, b])∗ ⫌ L1 ([a, b]). C ([a, b]) is indeed isomorphic to the signed Borel

b
measures of finite mass, i.e. Lf = a f dµ.
Example 11.7 (Riemann-Lebesgue Lemma from Fourier Analysis). For f ∈ L2 ([−π, π]) this states:
 π
lim e±inx f (x) dx = 0. (∗)
n→∞ −π

We can rewrite this in functional analysis language:

(∗) ⇐⇒ e±in· ⇀ 0 ∈ L2 ([−π, π]) .
def. of weak conv. n→∞
n o
Proof. √1 ein is orthonormal in L2 ([−π, π]) and by problem 56(e) from class, we know that if
2π n∈Z
{en } is orthonormal in a Hilbert space, ej ⇀ 0.
(N ) (N ) (N ) (N )
Example 11.8 (Riemann-Integral). Let f ∈ C ([a, b]), a = xo < x1 < · · · < xN −1 < xN =b
(N ) b−a
and xj = a + jh with h = N . We define:
N   b−a
(N )
X
SN f := f xj · ,
N
j=1

b
which we call a Riemann sum which then converges to a f for all f ∈ C ([a, b]).

Functional Analytic Formulation

b−a PN
SN : C ([a, b]) → R is a continuous linear functional, SN = N j=1 δx(N ) , δx f := f (x) is the translated
j

b
δ-“function”. L1 : C ([a, b]) → R is a continuous linear functional with L1 f = 1 1 · f . We have the

b
special case of Lψ f = a ψf and the aforementioned convergence of the Riemann Sum is equivalent
∗
to SN ⇀ L1 ∈ C ([a, b])∗ .

88
References
[1] Hans Wilhelm Alt. Linear Functional Analysis. Universitext SpringerLink. London: Springer,
2016. isbn: 9781447172802.
[2] Jean Alexandre Dieudonné. Foundations of modern analysis. Pure and applied mathematics; 10,1.
New York.
[3] Friedrich Hirzebruch and Winfried Scharlau. Einführung in die Funktionalanalysis. unchanged
reprint of the 1st ed. from 1971. Hochschultaschenbuch. Heidelberg: Spektrum, Akad. Verl., 1996.
isbn: 9783860254295; 3860254294.
[4] Michael Reed and Barry Simon. Methods of Modern Mathematical Physics. New York.
[5] Dirk Werner. Funktionalanalysis. 8. ed. 2018. Springer-Lehrbuch. Berlin, Heidelberg: Springer
Spektrum, 2018. isbn: 9783662554074.
[6] Kōsaku Yoshida. Functional analysis. Reprint of the 1980 ed. Classics in mathematics. Berlin:
Springer, 1995. isbn: 3540586547; 9783540586548.

89