Functional Analysis: Prof. Dr. Patrick Dondl Adapted From Lecture Notes by Josias Reppekus (TU-München) 23rd July 2016

Functional Analysis
Prof. Dr. Patrick Dondl

adapted from lecture notes by Josias Reppekus (TU-München)
23rd July 2016
Contents
Introduction 1
1 Basic structures 2
1.1 Topological spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Metric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Vector spaces and norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 Scalar product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.5 Example spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.5.1 The finite dimensional vector spaces Kn . . . . . . . . . . . . . . . . . . . . 10
1.5.2 The sequence spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.5.3 Bounded Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.5.4 Continuous and Differentiable Functions . . . . . . . . . . . . . . . . . . . . 11
1.6 Compactness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2 Lebesgue-Spaces, Part 1 14
2.1 A reminder of theorems from measure theory . . . . . . . . . . . . . . . . . . . . . 14
2.2 Definition and basic properties of Lp . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3 Density of smooth functions and separability . . . . . . . . . . . . . . . . . . . . . 18
3 Continuous linear maps 21
4 Hahn-Banach theorem and some consequences 23

4.1 Analytic version of the theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.2 Separation of convex sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5 Baire category argument 31

5.1 Banach-Steinhaus theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
5.2 Open mapping and closed graph theorems . . . . . . . . . . . . . . . . . . . . . . . 33
6 The weak topology 36

6.1 The weak topology σ (X, X 0 ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
6.2 The weak* topology σ(X 0 , X) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
6.3 Reflexive spaces and separable spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 47
7 Lebesgue-Spaces Part II 56
7.1 The Dual of Lp . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
7.1.1 Case 1 < p < ∞ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
7.1.2 The space L1 (µ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
7.1.3 Study of L∞ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
7.2 Weak convergence in Lp (µ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
8 Hilbert spaces 62
8.1 Projection onto convex sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
8.2 Dual spaces of Hilbert spaces and the theorem of Lax and Milgram . . . . . . . . . 64
8.3 Orthonormal basis in Hilbert spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 66
9 Some theory of compact operators 69

9.1 Compact Operators and the adjoint Operator . . . . . . . . . . . . . . . . . . . . . 69
Introduction
• Functional Analysis: Analysis in infinite dimensional spaces.
• Linear Functional Analysis: Linear Algebra in infinite dimensional spaces.
• Typical question: X, Y vector spaces; A : X → Y linear continuous. Does A have a continu-

ous inverse? (If not, can we somehow measure the defect?)
This leads to the so-called Fredholm-Alternative
Example. U ⊂ Rn open, with smooth boundary,
X := u ∈ C 2 (U ) ∩ C U | u = 0 on ∂U ,

Y := C (U ) ,
n
X ∂2
A := −∆ = .
i=1
∂x2i
A invertible ⇔ The equation

−∆u (x) = f (x) , x ∈ U,
(0.1)
u (x) = 0, x ∈ ∂U.
has a solution for any f ∈ Y.
Classically: The equation is a pointwise condition on u.
Modern: u is a point (or vector) in a function space, −∆ is a linear mapping between function
spaces.
Fact: A is continuous and injective. However, A is not surjective.
In particular, there exist f ∈ C (U ) : no u ∈ Y satisfies −∆u = f. We will work out
this example in the first tutorial.
Problem: C 2 is a terrible function space for the Laplacian. Much better: Hölder spaces (C 2,α )
or Sobolev spaces.
Topics of the class

• Spaces, norms, topologies,
• Linear operators and their properties,
• General theory ↔ concrete spaces, e.g. Lp .
1
1 Basic structures
Outline
• Basic structures and spaces,
• Continuous linear operators,
• Hahn-Banach,
• Baire Lemma,
• Weak topology,
• Lp and Sobolev spaces,
• Hilbert space theory,
• Spectral theory.
1 Basic structures
1.1 Topological spaces
Definition 1.1 (Topology). Let X be a set and let T be a set of subsets of X (Notation: T ⊂ 2X )
with the properties
1. ∅, X ∈ T
2. T 0 ⊂ T ⇒
S
U ∈T 0 U ∈T
3. U1 , U2 ∈ T ⇒ U1 ∩ U2 ∈ T .
Then, (X, T ) is called a topological space, T is called topology and sets U ∈ T are called open.
If in addition to that, we have that for all x1 , x2 ∈ X, x1 6= x2 there exist U1 , U2 ∈ T with the
property that
x1 ∈ U1 , x2 ∈ U2 , U1 ∩ U2 = ∅, (separation axiom)
then (X, T ) is called a Hausdorff space.
Definition 1.2. A subset A ⊂ X of a topological space is called closed if
∃U ∈ T such that A = X\U =: U c .
Let X henceforth be a topological space.

Definition 1.3. Let A ⊂ X.
1. A◦ := {x ∈ X | ∃U ⊂ A, U ∈ T s.t. x ∈ U } is the interior of A.
2. A := {x ∈ X | ∀U ∈ T , x ∈ U we have U ∩ A 6= ∅} is the closure of A.
3. ∂A := A\A◦ is the boundary of A.
Definition 1.4. 1. A set A ⊂ X is called dense in X, if A = X.
2. X is called separable, if there exists A ⊂ X that is dense and countable.
Proposition 1.5. 1. A◦ ⊂ A ⊂ A.
2. A = A◦ ⇔ A open, A = A ⇔ A closed.
3. A◦ is open, A is closed.
◦
4. X\A = (X\A) .
Proof. Analysis 2 or exercise.
2
1 Basic structures 1.1 Topological spaces
Proposition 1.6. Let A ⊂ X, (X, T ) be a topological space. Then (A, TA ) with
TA := {U ∩ A | U ∈ T }
is a topological space. TA is called a subspace topology (relative topology).

Proof. Should be pretty clear.
Definition 1.7. Let (X, T ) be a topological space.
• B ⊂ T is called basis of T , if any set in T can be written as a union of sets in B.
• S ⊂ T is called a subbasis of T , if the set of all finite intersections of sets in S is a basis of T .
Proposition 1.8. Let X be a set, S ⊂ 2X a collection of subsets of X. Let now B ⊂ 2X be the
set of subsets of X generated by taking any finite intersections between sets in S. Then the set
generated by taking arbitrary unions of sets in B is a topology.
Proof. Exercise.
Definition 1.9. Let T1 , T2 be topologies on X. T2 is called stronger (or finer ) than T1 and T1
weaker (or coarser ) than T2 , if
T1 ⊂ T2 .
Example. 1. The indiscrete topology: T = {∅, X} ,
2. The discrete topology: T = 2X := P (X) ,

3. The cofinite topology: X = N, T := U ∈ 2N | # (X\U ) < ∞ ∪ ∅,
4. Topologies induced by a metric.
Definition 1.10 (Neighborhood). Let A ⊂ X. A set N is called a neighborhood of A if there
exists an open set U such that A ⊂ U ⊂ N . Note that A can be a singleton. We write N (A).
Definition 1.11 (Continuity). Let (X, TX ) , (Y, TY ) be topological spaces.
1. A mapping f : X → Y is called continuous if
∀V ∈ TY we have f −1 (V ) ∈ TX .
2. A mapping f : X → Y is called continuous in x ∈ X, if for any neighborhood NY of f (x)

there exists a neighborhood NX of x such that f (NX ) ⊂ NY .
Proposition 1.12. f : X → Y is continuous, if and only if f : X → Y is continuous in all x ∈ X.
Proof. “⇒” Fairly easy, just take the preimage of the open set in NY as NX .
“⇐” Consider V ⊂ Y open. If the preimage is empty, we are done. Otherwise, take x ∈
f −1 (V ) =: A. By the definition of continuity in a point, there exists an open set U
containing x such that U ⊂ A. So A = A◦ and thus open.
Definition 1.13 (Convergence). Let (X, T ) be a topological space, let x ∈ X. We say that a
sequence (xk )k∈N converges to x, if
for any neighborhood N (x) we have that ∃k0 ∈ N such that ∀k > k0 we have xk ∈ N.
We write
T
xk −
→ x.
Proposition 1.14. If (X, T ) is Hausdorff, then the limit x is uniquely determined. We then write
x = lim xk w.r.t. (with respect to) T .

k→∞
3
1 Basic structures 1.2 Metric
Proof. Assume xk → x1 , xk → x2 6= x1 . Then
∃U1 3 x1 , U2 3 x2 with U1 ∩ U2 = ∅.
But: ∃k0 : ∀k > k0 we have xk ∈ U1 and at the same time xk ∈ U2 . This is a contradiction.
Example. 1. TX = 2X . Any function f : X → Y is continuous.
However, xk → x only if ∃k0 s.t. ∀k > k0 we have xk = x.
2. TX = {∅, X}, TY Hausdorff. Then we have: f is continuous only if f is constant.

However, any sequence in X converges to every element in X.
Remark (Direct method of the calculus of variations). Goal: Find x ∈ X :
f (x) = inf f (y) for some f : X → [0, ∞).

y∈X
1. ∃ (xk )k∈N : f (xk ) → inf f. We have f (xk ) < C for k > k0 .

2. We want that a subsequence of (xk )k∈N converges to some x ∈ X.
3. Then we need to show that f (x) = inf f. (⇐ xj → x ⇒ f (x) ≤ lim inf f (xj ))
| {z }
lower semicontinuity
Remark 1.15. i) One topic we left

Qnuntouched is the product topology. For finite products it suffices
to consider the box topology on i=1 Xi , namely the topology to the basis consisting of all sets of
the form ( n )
Y
Ui | Uj ∈ Tj .
i=1
ii) A further point we have not discussed here is the issue of compactness. In purely topological
spaces, this is somewhat tricky and we will treat compactness in more detail later.
1.2 Metric
Definition 1.16. Let X be a set and d : X × X → R, such that for all x, y, z ∈ X we have
1. d (x, y) ≥ 0 and d (x, y) = 0 ⇔ x = y.

2. d (x, y) = d (y, x) .
3. d (x, z) ≤ d (x, y) + d (y, z) .
Then we call (X, d) a metric space, d a metric or a distance.

Remark. • Without the requirement that d (x, y) = 0 ⇔ x = y we call d a pseudo metric.
• By taking a modulo
x=y
ˆ ⇔ d (x, y) = 0
we can transform a pseudo metric space into a metric space.
• The only thing that is to show for this claim is
x1 , y1 , x2 , y2 ∈ X, x1 =x
ˆ 2 , y1 =y
ˆ 2 ⇒ d (x1 , y1 ) = d (x2 , y2 ) .
This, however follows from the triangular inequality [3. of the definition of a (pseudo) metric].
Example. 1. X = R, d (x, y) = |x − y| .
|x−y|
2. X = R, d (x, y) = 1+|x−y| .
4
3. (X, d) metric space, h : Y → X injective,
dY (y1 , y2 ) := d (h (y1 ) , h (y2 ))
is called the pullback metric.

4. Discrete metric: (
0, x = y
d (x, y) := .
6 y
1, x =
Definition 1.17. Let (X, d) be a metric space,
Br (x) := {y ∈ X | d (x, y) < r} for r > 0, x ∈ X.
We call U ⊂ X open with respect to the metric d, if ∀x ∈ U there exists r > 0
Br (x) ⊂ U.
The empty set is open.

Proposition 1.18. Let (X, d) be a metric space, and let
T := {U ⊂ X | U is open w.r.t. d} .
Then (X, T ) is a Hausdorff space. We call T the by d induced topology.

Proof. 1. ∅, X ∈ T : clear.
2. Let T 0 ⊂ T and W := U ∈T 0 U . We need to show that for all x ∈ W there exists r > 0 with
S
Br (x) ⊂ W . This is obvious, though, since there is some U ∈ T 0 with x ∈ U and we can
just take the r for that U .
3. Let U1 , U2 ∈ T , x ∈ U1 ∩ U2 . We have r1 , r2 > 0 :
Br1 (x) ⊂ U1 and Br2 (x) ⊂ U2 .
It follows by the triangle inequality that
Bmin(r1 ,r2 ) (x) ⊂ U1 ∩ U2 .
4. Let x 6= y. We have that d (x, y) = c > 0.

c
Br (x) ∩ Br (y) = ∅, r= > 0.
2
Definition 1.19. Let d1 , d2 both be a metric on X.
• d1 is called stronger than d2 or d2 weaker than d1 , if the same holds for the induced topologies.
• d1 is called equivalent to d2 , if the induced topologies are equal.
Proposition 1.20 (Continuity in metric spaces). Let (X, dX ) , (Y, dY ) be metric spaces and let
f : X → Y. Then f is continuous in x ∈ X, if and only if
∀ε > 0 : ∃δ > 0 : dX (x, y) < δ ⇒ dY (f (x) , f (y)) < ε.
Proof. Analysis 2.
Proposition 1.21 (Convergence in metric spaces). Let (X, dX ) , (Y, dY ) be metric spaces.
1. Let (xj )j∈N be a sequence in X, then
d
xk −
→x ⇔ ∀ ∃k : d (xk , x) ≤ ( for k ≥ k ) .
5
2. A mapping f : X → Y is continuous in x, iff for all sequences (xk )k∈N with xk → x we have
f (xk ) → f (x) .
Remark. The second condition is called sequential continuity.

Warning. This is generally not true in purely topological spaces.
Proof. 1. Clear.
2. “⇒” Clear (somewhat).
“⇐” Assume that f is not continuous. Then ∃ε > 0 : ∀δ > 0∃xδ :
dX (x, xδ ) < δ, but dY (f (x) , f (xδ )) > ε.

1 1
δ := k ; ∃xk : dX (x, xk ) < k → 0.
dY (f (x) , f (xk )) ≥ ε, xk → x, but f (xk ) 6→ f (x) .
Definition 1.22. Let (X, d) be a metric space.

1. (xk )k∈N is called a Cauchy sequence, if ∀ε > 0∃k0 ∈ N : ∀k, l > k0 , we have d (xk , xl ) < ε.
2. The space (X, d) is called complete, if every Cauchy sequence admits a limit in X.
Proposition 1.23 (sequential criterion). Let A ⊂ (X, d) (metric space). A is closed, iff for all
sequences (xj )j∈N with xj ∈ A and xj → x ∈ X, we have x ∈ A.
Proof. ⇒ Assume there was a limit point x in Ac , which is open. Thus there exists a neighbor-
hood N (x) ⊂ Ax . Then points in the sequence would have to lie in this neighborhood and
outside of A, which is a contradiction.
⇐ Assume A is not closed and consider x ∈ Ac \ A. By the definition of the closure, for any
r > 0 we have Br (x) ∩ A 6= ∅. Now take a sequence of radii going to zero and pick as xk any
point in the intersection of the respective ball with A.
Proposition 1.24. Let (X, d) be a complete metric space, and A ⊂ X be closed. Then (A, d) is a
complete metric space.
Proof. It’s clear that (A, d) is a metric space. For completeness, consider a Cauchy sequence in A.
By completeness of (X, d), it admits a limit in X, this limit must lie in A by the above sequential
criterion.
Example. 1. Q (the rationals) with the usual distance are a not complete metric space.
2. Take I = [0, 1] and
Pn := {f : I → R | f is a ploynomial of degree deg (f ) ≤ n} ,

[
P := Pn ,
n∈N
d (f, g) := sup |f (x) − g (x)| .
x∈I
(P, d) is a metric space. We take

n
X xk
f (x) := exp (x) and fn (x) := ,
k!
k=0
noting that f ∈
/ P . We have
sup |fn (x) − f (x)| → 0.
x∈I
6
(And also supx∈I |fn (x) − fk (x)| −−−−−→ 0.) However, if

n,k→∞
g = lim fk
k→∞
would exist in P , then necessarily, we would need
|fn (x) − g (x)| ≤ d (fn , g) → 0 ∀x ∈ I,
so we would have
g = f = exp .
But again, f ∈
/ P.
Theorem 1.25 (Completion). Let (X, d) be a metric space and consider the set X N of all sequences
in X. Let n o.
X̃ := x̃ = (xj )j∈N ∈ X N | (xj )j∈N is a Cauchy-sequence
=
endowed with the equivalence relation
(xj )j∈N = (yj )j∈N :⇔ d (xj , yj ) → 0.

Then, X̃, d˜ is a complete metric space with

d˜ (xj )j∈N , (yj )j∈N := lim d (xj , yj ) .
j→∞
Furthermore, the mapping
J : X → X̃, J (x) := (x)j∈N (constant sequence)
defines an (therefore injective) isometry, that is
d˜(J (x) , J (y)) = d (x, y) ∀x, y ∈ X,
and for any (xj )j∈N ∈ X̃, we have

d˜ (xj )j∈N , J (xk ) −−−−→ 0,
k→∞
thus J (X) is dense in X̃.

˜ Let x̃ = (xj )
• Well-definedness of d:
Proof. j∈N , ỹ = (yj )j∈N be in X̃. We have
|d (xj , yj ) − d (xi , yi )| ≤ |d (xj , yj ) − d (xi , yj )| + |d (xi , yj ) − d (xi , yi )|

≤ d (xj xi ) + d (yj , yi ) −−−−→ 0.
i,j→∞
So this was a Cauchy sequence in R and the limit
d˜(x̃, ỹ) = lim d (xj , yj )

j→∞
exists. By analogous arguments,
d x2i , yi2 − d x1i , yi1 → 0,

if x̃1 = x̃2 ∈ X̃ and ỹ 1 = ỹ 2 ∈ X̃.

• Is d˜ a metric? Consider the axioms
1. d˜ ≥ 0 clear.
d˜(x, y) = 0 ⇔ x = y (clear by definition of “=” on X̃).
7
1 Basic structures 1.3 Vector spaces and norms
2. Symmetry carries over in the limit.

3. The triangle inequality also carries over in the limit.

• Completeness: Consider xk k∈N a Cauchy sequence in X̃, xk = xkj j∈N ∈ X̃. For k ∈ N,
we pick a jk , such that
1
d xki , xkj ≤ ∀i, j ≥ jk .
k
Now for j ≥ jk , jl , k, l ∈ N, we have
d xkjk , xljl ≤ d xkjk , xkj + d xkj , xlj + d xlj , xljl

1 1
≤ + d xkj , xlj +
k l
1 ˜ k l
1
−−−→ + d x , x +
j→∞ k l
−−−−→ 0.
k,l→∞
So we define
x∞ := xljl

l∈N
∈ X̃.
Claim. xl → x∞ .
Proof. d˜ xl , x∞ ←−−−− d xlk , x∞

k , however for k ≥ jk :
k→∞
d xlk , x∞ ≤ d xlk , xljl + d xljl , xkjk

k
1
≤ + d xljl , xkjk

l
−−−−→ 0.
k,l→∞
• The statements about J are easy to verify.
1.3 Vector spaces and norms

We only consider vector spaces over the fields K ∈ {R, C} and consider those fields metric (or
topological) spaces with the usual distance.
Definition 1.26. Let X be a vector space over K and let X at the same time a topological space.
If vector addition and scalar multiplication are continuous, then X is called a topological vector
space.
Remark. Note that we need to use the box topology where appropriate here.
Definition 1.27 (Norm).

• Let X be a vector space over K. A mapping k·k : X → R is called a norm, if we have
1. kxk ≥ 0 ∀x ∈ X.
2. kxk = 0 ⇒ x = 0.
3. kαxk = |α| kxk .
4. kx + yk ≤ kxk + kyk .
• If k·k is a norm on X, then (X, k·k) is called a normed space.
• A mapping with properties 1., 3., 4. is called semi-norm.

Remark. Again, we can turn a space with a semi-norm into a normed space by taking the modulo.
8
1 Basic structures 1.3 Vector spaces and norms
Proposition 1.28. Let (X, k·k) be a normed space. By taking
d : X × X → R, d (x, y) := kx − yk ,
we get a metric space (X, d) .

Proof. Analysis II.
Definition 1.29. A complete normed space is called Banach space.
Proposition 1.30. A normed vector space is a topological vector space and the topology is Haus-
dorff.
Proof. Exercise.
Proposition 1.31. Let k·k1 , k·k2 be two norms on X, d1 , d2 the induced metrics and T1 , T2 the
induced topologies. We have
1. d2 is stronger than d1 , iff ∃C > 0 :
kxk1 ≤ C kxk2 ∀x ∈ X.
2. Two norms are equivalent (:⇔ T1 = T2 ), iff ∃C, c > 0 :
c kxk1 ≤ kxk2 ≤ C kxk1 .
Proof. ⇐ Take U open w.r.t. T1 , and take x ∈ U . Then there exists a d1 -r-ball around x which
still lies in U . The d2 - Cr -ball then also lies in U .
⇒ Consider T1 3 Brd1 , an open d1 -r-ball around the origin and, for any δ > 0, the d2 -δ-ball Bδd2 .
Since the inequality is not satisfied for any C, we can always find an x ∈ Bδd2 with x ∈ / Brd1 ,
which is therefore not open wrt. d2 .
Theorem 1.32. All norms on a finite dimensional vector space are equivalent.
Proof. Use the basis, Luke.
Example. The norms
ˆ 1
kf k∞ := sup |f (t)| and kf k1 := |f (t)| dt
t∈[0,1] 0
on the vector space C ∞ ([0, 1]) are not equivalent. Take e.g. fn (t) := tn . We have
kfn k∞ = 1, however
1
kfn k1 = → 0.
n+1
Theorem 1.33. A finite dimensional subspace of a normed space is complete and closed.
Proof. Completeness follows from the completeness of Kn , closedness from the sequence-criterion
for closedness from before.
Remark. • This does not generally hold for infinite dimensional subspaces.
Remark. If X, Y are normed spaces, Z = X × Y , then
p p
kξkZ,p := (kζkX + kηkY ) , ξ = (ζ, η) ∈ Z
and
kξkZ,∞ := max (kζkX , kηkY )
are all equivalent norms on Z. Z is a Banach space, iff X and Y are Banach.
9
1 Basic structures 1.4 Scalar product
1.4 Scalar product

Definition 1.34. Let X be a K-vector space. A mapping
(·, ·) : X × X → R
is called a sesquilinear form, if ∀x, x1 , x2 , y, y1 , y2 ∈ X, α ∈ R we have

1. (αx, y) = α (x, y) = (x, ᾱy)
(
(x1 + x2 , y) = (x1 , y) + (x2 , y)
2.
(x, y1 + y2 ) = (x, y1 ) + (x, y2 )
A sesquilinear form is symmetrical, if
(x, y) = (y, x).
A symmetrical sesquilinear is called positive semidefinite, if
(x, x) ≥ 0 ∀x ∈ X.
A positive semidefinite sesquilinear form is called positive definite, if
(x, x) = 0 ⇒ x = 0.
Remark. If K = R, we use the phrase bilinear.

Definition 1.35. A positive definite (symmetrical) sesquilinear form is called scalar product. The
pair (X, (·, ·)) is called pre-Hilbert space.
Lemma 1.36. Let (·, ·) be a scalar product and define

p
kxk := (x, x) ∀x ∈ X.
p
Then kxk := (x, x) is a norm on X. Furthermore we have
1. |(x, y)| ≤ kxk · kyk (Cauchy-Schwarz inequality)

2 2 2 2
2. kx + yk + kx − yk = 2 kxk + 2 kyk (Parallelogram identity)
Proof. Linear Algebra
Remark. Central Property: Orthogonality.

Definition 1.37. 1. Two vectors x, y are called orthogonal, if (x, y) = 0.
2. Two subspaces U, V ⊂ X are called orthogonal, if (x, y) = 0 ∀x ∈ U, y ∈ V.
Definition 1.38. A complete pre-Hilbert space is called Hilbert space.
´1 ´1
Example. X = C 1 ([0, 1]) with (f, g) := 0 f g + 0 f 0 g 0 is a pre-Hilbert space.
p
The completion of X with respect to kxk = (x, x) is called Sobolev space H 1 .
1.5 Example spaces

1.5.1 The finite dimensional vector spaces Kn
A finite dimensional vector space, in fact the canonical finite dimensional
Pnvector space over the field
K. Becomes a Banach space with any of the usual norms ||x||p = ( i=1 |xj |p )1/p and a Hilbert
space with the Euclidean norm. Do you remember the triangle inequality?
10
1 Basic structures 1.5 Example spaces
1.5.2 The sequence spaces

We consider KN , the space of sequences with values in K. The i-th canonical unit vector is denoted
by ei , the vector where the i-th entry is 1, all others are zero.
Theorem 1.39. We have
P∞ |xi |
1. KN is a metric space with the Fréchet-Metric d(x, y) = ρ(x − y) with ρ(x) = j=1 2−j 1+|x i
.
2. Entrywise convergence is the same as convergence with respect to this metric

3. The space KN is complete.
4. `p , 1 ≤ p ≤ ∞ are Banach-spaces.
5. `2 is a Hilbert-space.
Proof. See Alt, Lineare Funktionalanalysis.
1.5.3 Bounded Functions

For a set S and a K-Banach-space Y we consider the set of all bounded maps from S to Y ,
B(S, Y ) = {f : S → Y ; f (S) is a bounded subset of Y}.
Theorem 1.40. With the supremumsnorm ||f ||B(S,Y ) = supx∈S ||f (x)||Y , this space becomes a
Banach-space.
Proof. The properties of being a norm are easy to verify. For any x in S, every Cauchy-sequence
in B(S, Y ) is a Cauchy-sequence in Y and thus admits a limit (by assumption of Y being Banach)
f (x). We have
|f (x) − fk (x)| = lim |fl (x) − fk (x)| ≤ lim inf ||fl − fk ||B(S,Y ) .
l→∞ l→∞
Thus, f − fk is a bounded function and so is f . Furthermore, for k → ∞, the right hand side of
the above goes to zero, so fk converges to f with respect to the supremums-norm, as the bound is
independent of x.
1.5.4 Continuous and Differentiable Functions

Theorem 1.41. For S ⊂ Rn closed and bounded, Y a K-Banach-space, the space of continuous
functions from S to Y endowed with the supremums-norm is a Banach-space.
Proof. Identification of the limit as in Theorem 1.40. We also know that uniform limits of con-
tinuous functions are continuous.
Remark. 1. On unbounded or not closed sets S ⊂ Rn , since then continuous functions may
not be bounded, one can either simply look at the bounded continuous S functions and derive
the same theorem as above, or for compact sets (Kj )j∈N such that j Kj = S define a
P∞ ||f ||
Fréchet-metric by taking ρ(x) = j=1 2−j 1+||f C(K i ,Y )
||C(Ki ,Y ) . This results in a complete metric
space.
2. On closed and bounded sets S, we consider m-times continuously differentiable functions
to be the functions whose derivatives, which are naturally defined on the interior of S,
can be continuously extended to the boundary. These spaces form Banach-spaces with the
supremums-norm on all derivatives (proof is the same as for continuous functions, the uniform
convergence also ensures that the limits of the derivatives are again the respective derivatives
of the limits).
3. There is no norm such that induces convergence on each derivative on the space of infinitely-
differentiable functions (see exercises).
4. We also know the spaces of continuous (or differentiable, etc.) functions with compact
support.
11
1 Basic structures 1.6 Compactness
1.6 Compactness
Theorem 1.42 (Compactness in metric spaces). Let A be a subset of a metric space (X, d) . Then
the following are equivalent:
1. A is (covering) compact, that is, every open cover of A contains a finite subcover.
2. A is sequentially compact, that is, every sequence in A admit a converging subsequence.
3. (A, d) is complete and precompact, that is, for all ε > 0 there is a finite cover of A with
ε-balls.
Remark. compact =
ˆ covering compact.
Proof. 1. ⇒ 2. Let (xk )k∈N be a sequence without accumulation point. Then, ∀y ∈ A : ∃ry > 0 :

Ny := k ∈ N | xk ∈ Bry (y) is finite.
N
The balls Bry (y) comprise an open cover of A. Therefore ∃ {yk }k=1 :
N
[
Bryk (yk ) ⊃ A.
k=1
This is a contradiction to our assumption, since otherwise there would only be finitely
many elements in the sequence.
2. ⇒ 3. Fist, we show completeness: By 2., we have that every Cauchy sequence in A admits
an accumulation point. However, a Cauchy sequence admits at most one accumulation
point. By 2., the accumulation point is in A. Therefore, the sequence converges to an
element of A.
For compactness, we argue by contradiction: Assume ∃ε > 0 : no finite cover with
ε-balls exists. Therefore there exists a sequence (xk )k∈N ⊂ A :
k
[
xk+1 ∈ A\ Bε (xj ) .
j=1
So (xk ) has no accumulation point.

3. ⇒ 1. Let (Ui )i∈I be an open cover of A. We define
( )
[
B := B ⊂ A | J ⊂ I, B ⊂ Ui ⇒ |J| = ∞ .
i∈J
We want to show A ∈
/ B. A is precompact, so ∀B ∈ B, ε > 0 there is a cover
n%
[
B⊂ Bε (xi ) .
i=1
Hence, for some i (depending on ε):
Bε (xi ) ∩ B ∈ B.
Assume A ∈ B. So inductively, we can take ε = k1 , k ∈ N and get existence of xk ∈ X

and sets:
B1 := A, Bk := B1/k (xk ) ∩ Bk−1 ∈ B ∀k ≥ 2.
Take yk ∈ Bk , k ∈ N. For k ≤ l, we have
2
yk , yl ∈ B1/k (xk ) ⇒ d (yk , yl ) ≤ .
k
12
1 Basic structures 1.6 Compactness
⇒ (yk )k∈N is a Cauchy sequence. Since A is complete, ∃y ∈ A :
εk := d (yk , y) −−−−→ 0.
k→∞
But we have y ∈ Ui0 for some i0 and hence for k large enough
Bk ⊂ B1/k (xk ) ⊂ B2/k (yk ) ⊂ B k2 +εk (y) ⊂ Ui0 .
Therefore Bk ∈
/ B, which is a contradiction.
Theorem 1.43 (Riesz). Let X be a Banach space. then B1 (0) is compact, if and only if X is
finite dimensional.
Proof. “⇐” Heine-Borel.
“⇒” By precompactness of B1 (0), there exists a cover
m
[
B1 (0) ⊂ B1/2 (yk ) .
k=1
Take
m
Y := span ({yk }k=1 ) .
Y is finite dimensional, so by Theorem 1.33 it is closed in X. We assume Y 6= X.
Claim. We have ∀θ ∈ (0, 1) : ∃xθ ∈ X, kxθ k = 1 :
dist (xθ , Y ) ≥ θ.
Proof. Take x ∈ X\Y. Then

dist (x, Y ) > 0.
There exists y0 ∈ Y :
1
0 < kx − y0 k ≤ dist (x, Y ) .
θ
Take
x − y0
xθ := .
kx − y0 k
Then, for all y ∈ Y :
∈Y
1 z }| {
kxθ − yk = kx − (y0 + kx − y0 k y)k
kx − y0 k
1
≥ dist (x, Y )
kx − y0 k
dist (x, Y )
≥ 1
θ dist (x, Y )
= θ.
This proves the claim.
Remark. Such an xθ is called almost orthogonal element.)

However, for some j ∈ {1, ..., n} :
xθ ∈ B1/2 (yj ) .
1
With 2 < θ < 1 there is a contradiction to
dist (xθ , Y ) ≥ θ.
13
2 Lebesgue-Spaces, Part 1
2 Lebesgue-Spaces, Part 1
In short, these are the spaces of functions f , such that
ˆ
p
|f | dµ < ∞.
2.1 A reminder of theorems from measure theory

Proofs and theorems for this paragraph and further read on measure theory:
• Fonseca & Leoni: Modern Methods in the calculus of of variations: Lp spaces
• Brokate & Kersting: Maß & Integral
• Evans & Gariepy: Measure Theory and Fine Properties of Functions (“tough read”)
We consider (Ω, M, µ) be a σ-finite measure space, i.e. Ω is a set and M is a σ-algebra in Ω,
µ is a measure and (Ω, M, µ) is σ-finite, i.e. ∃ (Ωn )n countable family of sets in M such that
∞
[
Ω= Ωn and µ (Ωn ) < ∞ ∀n.
n=1
Sets E ∈ M such that µ (E) = 0 are called null sets. We say that the same property holds almost
everywhere (a.e.) on Ω, if the property holds on all x ∈ Ω\E for some
´ null set E. In
´ this chapter,
we identify functions that agree almost everywhere. We often write f instead of Ω f dµ. for the
integral over a measurable function.
We will need the following facts about integration.
´
Definition. L1 (µ) := f : Ω → K, f measurable : Ω |f | dµ < ∞ .
Theorem (Monotone convergence theorem (Beppo-Levi)). Let (fn )n∈N be a sequence in L1 :
1. f1 ≤ f2 ≤ ... ≤ fn ≤ ... a.e. on Ω.
´
2. supn fn < ∞.
Then fn (x) converges a.e. on Ω to a finite limit f (x) . We have f ∈ L1 and
ˆ
kfn − f kL1 := |fn − f | −−−−→ 0.
n→∞
Theorem (Dominated convergence (Lebesgue)). Let (fn )n∈N be a sequence of functions in L1 ,

such that
1. fn (x) → f (x) a.e. on Ω.
2. ∃g ∈ L1 : |fn (x)| ≤ g (x) ∀n ∈ N, a.e. on Ω.
Then, f ∈ L1 and
kfn − f kL1 −−−−→ 0.
n→∞
Lemma (Fatou’s lemma). Let (fn )n∈N be a sequence in L1 :

1. ∀n : fn ≥ 0 a.e. on Ω.
´
2. supn fn < ∞.
For almost all x ∈ Ω, set
f (x) := lim inf fn (x) ≤ ∞.
n→∞
1
Then, f ∈ L and ˆ ˆ
f ≤ lim inf fn .
n→∞
14
2 Lebesgue-Spaces, Part 1 2.2 Definition and basic properties of Lp
Now let (Ω1 , M1 , µ1 ) and (Ω2 , M2 , µ2 ) be two measure spaces. There is a canonical way to
define a measure on the product space Ω = Ω1 × Ω2 :
Theorem (Tonelli). Let F : Ω1 × Ω2 → R be a measurable function, such that
´
1. Ω2 |F (x, y)| dµ2 < ∞ for a.e. x ∈ Ω1 and
´ ´
2. Ω1 Ω2 |F (x, y)| dµ2 dµ1 < ∞.
Then F ∈ L1 (Ω1 × Ω2 ) .
Theorem (Fubini). Let F ∈ L1 (Ω1 × Ω2 ). Then
F (x, ·) ∈ L1 (Ω2 ) for a.e. x ∈ Ω1
and ˆ
F (x, y) dµ2 ∈ L1 (Ω1 )
Ω2
end vice versa. Moreover

ˆ ˆ ˆ ˆ ¨
F (x, y) dµ2 dµ1 = F (x, y) dµ1 dµ2 = F (x, y) d (µ1 ⊗ µ2 ) .
Ω1 Ω2 Ω2 Ω1 Ω1 ×Ω2
Notation. A basic example is the case, where Ω = Rn and µ is the Lebesgue measure on Rn .
We denote by Cc (Rn ) the continuous functions with compact support on Rn . The support of a
(continuous) function f is the closure of the set {f 6= 0}
Theorem (density of continuous functions). The space Cc (Rn ) lies dense in L1 (Rn ) , i.e. ∀f ∈
L1 (Rn ) , ε > 0∃f¯ ∈ Cc (Rn ) such that

f − f 1 ≤ ε.
L
2.2 Definition and basic properties of Lp

Definition 2.1. Let p ∈ (1, ∞) . Set
p
Lp (µ) := f : Ω → K | f is measurable and |f | ∈ L1 (µ)

with ˆ 1/p
p
kf kLp := kf kp := |f | dµ .
Ω
Definition 2.2. We set
L∞ (µ) := {f : Ω → K | f is measurable and ∃C > 0 : |f (x)| ≤ C a.e. on Ω} ,
where
kf kL∞ := kf k∞ := inf {C > 0 | |f (x)| ≤ C a.e. on Ω} .
Remark. For Ω ⊂ R we often write Lp (Ω) for the Lp -space on Ω with the Lebesgue measure. We
n
sometimes implicitly consider a function in Lp (Ω) to be a function of Lp (Rn ) by continuation by

zero outside Ω. If no ambiguity should occur, we sometimes leave out the measure or set altogether.
Definition 2.3. We set for 1 ≤ p ≤ ∞
Lploc (Ω) := {f : Ω → K|f ∈ Lp (K) for any K ⊂ Ω compact}
and we write fn → f in Lploc if fn converges to f in Lp on any compact subset of Ω.

Notation. Let 1 ≤ p ≤ ∞. We denote by p0 the conjugate exponent:
1 1
+ = 1.
p p0
15
0
Theorem 2.4 (Hölders’s inequality). Assume f ∈ Lp , g ∈ Lp , 1 ≤ p ≤ ∞. Then f · g ∈ L1 with
ˆ
|f · g| ≤ kf kLp · kgkLp0 .
Remark. 1. A useful consequence of Hölders’s inequality is the following:

Take f1 , f2 , ..., fk funktions: fi ∈ Lpi :
1 1 1 1
= + + ... + ≤n
p p1 p2 pk
Then f = f1 · f2 · · · fk ∈ Lp and
kf kLp ≤ kf1 kLp1 · · · kfk kLpk .
In particular, if f ∈ Lp ∩ Lq , 1 ≤ p ≤ q ≤ ∞, then
f ∈ Lr ∀p ≤ r ≤ q
and we have the following “interpolation inequality”

α 1−α
kf kLr ≤ kf kLp kf kLq ,
where
1 α 1−α
= + , 0 ≤ α ≤ 1.
r p p
2. We also note that for 1 ≤ p < ∞ we have ((|a| + |b|)p ≤ 2p (|a|p + |b|p ).
Proof of Theorem 2.4. See exercises.
Theorem 2.5. Lp is a vector space and k·kLp is a norm for any 1 ≤ p ≤ ∞.

Proof. The only difficult thing for to check p < ∞ is (again) triangle inequality, the checking of
which is an exercise. For p = ∞ the triangle inequality is easy, but make sure that zero norm
implies the function vanishes a.e.
Theorem 2.6 (Fischer-Riesz). Lp is a Banach space for any 1 ≤ p ≤ ∞.

Proof. p = ∞. Take a Cauchy sequence (fn )n∈N in L∞ . For all k ≥ 1 there exists Nk :
1
kfm − fn kL∞ < ∀n, m ≥ Nk .
k
Thus there exist null sets Ek :
1
|fm (x) − fn (x)| ≤ ∀x ∈ Ω\Ek , n, m ≥ Nk . (∗)
k
Take [
E= Ek .
k∈N
We have µ (E) = 0. We see that for all x ∈ Ω\E, fn (x) is Cauchy in K. Thus
fn (x) → f (x) ∀x ∈ Ω\E.
Taking the limit n, m → ∞ in (∗) yields

1
|f (x) − fn (x)| ≤ ∀x ∈ Ω\E, n ≥ Nk .
k
L∞
⇒ fn −−→ f.
16
1 ≤ p < ∞. Let (fn ) be a Cauchy sequence in Lp . It is enough to show convergence on a sub-

sequence (why?), so wlog. we may assume
1
kfk+1 − fk kLp ≤ ∀k ≥ 1.
2k
Let
n
X
gn (x) := |fk+1 (x) − fk (x)| .
k=1
We have
kgn kLp ≤ 1.
g is also increasing, so by the monotone convergence theorem (Beppo-Levi), (gn (x))n∈N
has a finite limit g (x) a.e. on Ω such that g ∈ Lp . We have for m ≥ n ≥ 2 :
|fm (x) − fn (x)| ≤ |fm (x) − fm−1 (x)| + ... + |fn+1 (x) − fn (x)| ≤ g (x) − gn−1 (x) .
Therefore, we have that fn (x) is a Cauchy sequence a.e. on Ω. It follows that fn (x)
converges to a limit f (x). So a.e. on Ω :
|f (x) − fn (x)| ≤ g (x) ∀n ≥ 2.
Thus, f ∈ Lp and by dominated convergence,
kfn − f kLp → 0,
p p
since |fn (x) − f (x)| → 0 a.e. and |fn − f | ≤ g p ∈ L1 .
Theorem 2.7. Let fn be a sequence in Lp and let f ∈ Lp :
kfn − f kLp −−−−→ 0.
n→∞
Then there exists a subsequence (fnk )k∈N and h ∈ Lp :

1. fnk → f a.e. on Ω.
2. |fnk | ≤ h a.e. on Ω.
Example. Take (
1, x ∈ In
fn (x) := ,
0, otherwise
where (In )n∈N is a sequence of intervals that repeatedly move through all of [0, 1] (i.e. every point
of [0, 1] is in at least one of the intervals of each “round”) while steadily getting smaller.
This sequence converges to zero in Lp , p < ∞, but it converges nowhere pointwise.
Proof of Theorem 2.7. p = ∞ is obvious. Take 1 ≤ p < ∞. Since (fn )n∈N is a Cauchy sequence,
we consider a relabelled subsequence (fk )k∈N :
a.e.
fk −−−−→ f¯ ∈ Lp .
k→∞
From the previous proof, we know

f¯ − fk ≤ g ∈ Lp

a.e. on Ω.
By dominated convergence, we know that
p
L
fk −−→ f¯
and thus
f = f¯ a.e.
In addition, we have
|fk (x)| ≤ f¯ (x) + g (x) ,

which implies 2.
17
2 Lebesgue-Spaces, Part 1 2.3 Density of smooth functions and separability
2.3 Density of smooth functions and separability

Theorem 2.8. Let Ω ⊂ Rn open. The space Cc (Ω) is dense in
Lp (Ω) := Lp (Ω, µ) ∀1 ≤ p < ∞, (µ Lebesgue measure).
Notation. We define the complex sign function

(
z
|z| , z 6= 0
sgn : C → B1 (0) , sgn z := ,
0, z=0
the truncation operator

Tn : C → C, Tn z := sgn z · min {|z| , n} ,
and for a set E ⊂ Ω, the characteristic function
(
1, x∈E
χE : Ω → R, χE (x) := .
0, x∈
/E
Proof of Theorem 2.8.

Claim. Given f ∈ Lp (Ω) , ε > 0, there exists g ∈ L∞ (Ω) and a compact set K ⊂ Ω such that
g (x) = 0 ∀x ∈ Ω\K, and kg − f kLp < ε.

S
Proof. Take (Kn )n∈N an increasing sequence of compact subsets of Ω such that n∈N Kn = Ω \ E
for some null set E (this is possible since Lebesgue measurable sets can be approximated from the
inside by compact sets).
χn := χKn , fn := χn Tn f.
We have
a.e.
fn −−→ f.
By dominated convergence, we see
kfn − f kLp −−−−→ 0.

n→∞
So it suffices to take g = fn for n large enough.

By the density of Cc (Rn ) in L1 (Rn )), ∀δ > 0 : ∃g1 ∈ Cc (Ω) such that
kg − g1 kL1 < δ.
To see this if Ω 6= Rn , first consider the function g as a function on all of Rn (by continuation by
zero outside Ω), then approximate with a functio g̃1 by density of Cc . Then, since Ω is open and
g vanishes outside a compact subset Kn of Ω, we have dist(Kn , ∂Ω) = c > 0. We can thus find K
compact and U open with, Kn ⊂ U ⊂ K ⊂ Ω and have space to cut off g̃1 in a continuous fashion
in between those two sets such that supp(g1 ) ⊂ K by taking g1 (x) = φ(x)g̃1 (x) where φ = 1 on
Kn , φ = 0 on Ω \ K, and φ continuous. Note that this cutting off can not increase the L1 -distance
to g as g vanishes outside Kn anyhow.
We can furthermore assume
kg1 kL∞ ≤ kgkL∞
(otherwise, replace g1 with T||g||L∞ g1 and note that the distance again only decreases) and get
1/p 1−1/p
kg − g1 kLp ≤ kg − g1 kL1 · kg − g1 kL∞ (the useful interpolation inequality)
1/p 1−1/p
<δ (2 kgkL∞ ) ,
so by picking δ small enough, the theorem is proven.

Corollary 2.9. Lp (Ω, µ), Ω ⊂ Rn measurable (not necessarily open), µ Lebesgue, is separable for
1 ≤ p < ∞.
18
Proof. First, consider the functions defined on all of Rn by continuation by zero. Then approx-
imate by a function in Cc (Rn ). Functions in Cc (Rn ) are uniformly continuous, so the can be
approximated in Lp by finite step functions. The finite step functions taking (complex) rational
values on appropriate sets form a countable dense subset of the finite step functions. If necessary,
cut off the step functions outside of Ω by multiplication with χΩ .
Remark. Separability also holds for a more general measure space Ω, if that measure space is
separable (i.e. its σ-algebra is countably generated).
Remark. L∞ ((0, 1)) is not separable. To see this, consider fα = χ(α,1) for α ∈ (0, 1) and note that
||fα − fβ ||L∞ = 1 for α 6= β and that there are more than countably many real numbers α ∈ (0, 1).
Therefore, B = {B1/2 (fα )|α ∈ (0, 1)} is an uncountable and non-intersecting collection of 1/2-balls
in L∞ ((0, 1)). Thus there can not exist a countable, dense subset. Similarly, L1 (Ω) is not separable
for any open set Ω ⊂ Rn .
Notation. In the following we write V ⊂⊂ Ω for V open and V ⊂ K ⊂ Ω for some compact set
K ⊂ Ω.
Definition 2.10. 1. We define η ∈ C ∞ (Rn ) by
(
C exp |x|21−1 if |x| < 1
η(x) =
0 otherwise
´n
where C > 0 is such that R η = 1.
2. For > 0 set
1 x
η = η .
n
We call η standard mollifier.
´
Remark. Note that Rn η = 1 and supp(η ) = B (0).
Definition 2.11. For f ∈ L1loc (Ω), we define its mollification
ˆ ˆ
f (x) = η (x − y)f (y)dy = η (y)f (x − y)dy
Rn B (0)
for x ∈ Ω. Note that we implicitly continued f by zero outside of Ω for the integration.
Theorem 2.12 (Properties of mollifiers). Consider Ω ⊂ Rn open, f ∈ L1loc (Ω). We have
1. f ∈ C ∞ (Ω)
2. If f ∈ C(Ω), then f → f uniformly on compact subsets of Ω.
3. For 1 ≤ p < ∞, f ∈ Lploc (Ω), f → f in Lploc (Ω).
Proof. 1. Let x ∈ Ω, i ∈ {1, ..., n}, h > 0, so that x + hei ∈ Ω.
We have
ˆ
f (x + hei ) − f (x)

1 1 x + hei − y x−y
= n η −η f (y)dy
h Rn h
ˆ
1 1 x + hei − y x−y
= n η −η f (y)dy
V h
for some V ⊂⊂ Rn .
We have
1 x + hei − y x−y 1 ∂ x−y
η −η → η
h ∂xi
∂f
uniformly on V , ∂xi therefore exists and equals
ˆ
∂η
(x − y)f (y)dy
Ω ∂xi
The same argument holds for any other derivative.
19
2. Let K ⊂ V ⊂⊂ Ω be compact and > 0. By uniform continuity of f on V , there exists δ > 0

with
|f (x − y) − f (x)| < ∀x ∈ K, y ∈ Bδ (0).
S
For x ∈ K and δ so small that z∈K Bδ (z) ⊂ V we now obtain
ˆ
δ
|f (x) − f (x)| ≤ η = .
3. Let 1 ≤ p < ∞, f ∈ Lploc (Ω) and V ⊂⊂ W ⊂⊂ Ω. Claim: For > 0 sufficiently small we have
kf kLp (V ) ≤ kf kLp (W )
Let x ∈ V . We have
ˆ ˆ

|f (x)| = η (x − y)f (y)dy ≤ B (x)η1−1/p (x − y)η1/p (x − y)|f (y)|dy

B (x)
ˆ 1−1/p ˆ 1/p

≤ η (x − y)dy η1/p (x − y)|f (y)|p dy

B (x) B (x)
| {z }
=1
ˆ ˆ ˆ !
p
⇒ |f (x)| dx = η1/p (x p
− y)|f (y)| dy dx
V V B (x)
ˆ ˆ ! ˆ
p
≤ |f (y)| η (x − y)dx dy = |f (x)|p dx
W B (y) W
if B (x) ⊂ W, ∀x ∈ V .
Now finally choose again V ⊂⊂ W ⊂⊂ Ω, δ > 0 and g ∈ Cc (W ) (by Theorem 2.8), so that
kf − gkLp (W ) < δ
Then
kf − f kLp (V ) ≤ kf − g kLp (V ) + kg − gkLp (V ) + kg − f kLp (V )
≤ 2kf − gkLp (W ) + kg − gkLp (V ) ≤ 2δ + kg − gkLp (V )
kg − gkLp (V ) → 0, ( → 0),
since g → g uniformly on V .
ˆ
|g (x) − g(x)| ≤ C |g(x) − g(y)|dy → 0
B (x)
And the the claim follows.
Remark. By first multiplying with χn as in the proof of Theorem 2.8, we also get density of
Cc∞ (Ω) in Lploc (Ω) for Ω ⊂ Rn open. Remember that we use the Fréchet-Metric on Lploc , i.e.,
||g|Kj ||Lp
d(f, g) = j∈N 2−j 1+||g|K
P S
j
||Lp for compact Kj ⊂ Ω such that Ω = j∈N Kj modulo a null set. For
bounded Ω this is equivalent to the usual distance from || · ||Lp (Ω) (try it!).
20
3 Continuous linear maps

We take (X, k·kX ) and (Y, k·kY ) to be normed spaces and consider a linear map A : X → Y .
Definition 3.1. A is called bounded, if
sup kAxkY < ∞.

kxkX ≤1
Theorem 3.2. The following are equivalent:

1. A is continuous in 0 ∈ X.
2. A is continuous.
3. A is Lipschitz continuous (i.e. ∃L > 0 : ∀x, y ∈ X : kAx − Ayk ≤ L kx − yk).
4. A is bounded.
Proof. 4. ⇒ 3. Let x1 + x2 ∈ X.
kAx1 − Ax2 k = kA (x1 − x2 )k

x1 − x2
= A
· kx1 − x2 k
kx1 − x2 k
≤ sup kAzk kx1 − x2 k .
kzk≤1
| {z }
=:L≤∞
3. ⇒ 2. Trivial.
2. ⇒ 1. Trivial.
1. ⇒ 4. By contradiction: Assume A is not bounded, that is
∃ (xk )k∈N in X : kxk k = 1, kAxk k → ∞.
We can construct
−1
zk := kAxk k xk → 0.
However, we have kAzk k = 1 ∀k, contradicting sequential continuity of A at 0.
Corollary 3.3. If dim X < ∞, then all linear operators A : X → Y are continuous.
Example. Take
(X, k·kX ) := (C ([0, 1]) , k·kL1 ) ,

(Y, k·kY ) := (C ([0, 1]) , k·k∞ ) .
Then
id : X → Y, f 7→ f
is not continuous. To see this, consider fk (t) = tk .
Definition 3.4.
L (X, Y ) := {A : X → Y | A is linear and continuous} ,

kAkL (X,Y ) := sup kAxkY .
kxkX ≤1
If Y = K, then we call X 0 := L (X, K) , the space of (continuous linear) functionals or the

(topological) dual space of X.
21

Proposition 3.5. L (X, Y ) , k·kL (X,Y ) is a normed space. For all A ∈ L (X, Y ) , x ∈ X :
kAxkY ≤ kAkL (X,Y ) kxkX .
Proof. Exercise.
Theorem 3.6. Let X, Y, Z be normed spaces, A ∈ L (X, Y ) , B ∈ L (Y, Z) . Then BA ∈ L (X, Z)
with
kBAkL (X,Z) ≤ kBkL (Y,Z) kAkL (X,Y ) .
Furthermore, the mapping
(A, B) 7→ BA
is continuous.
Proof. Let kxk = 1. Then we have
kBAxkZ ≤ kBkL (Y,Z) · kAxkY

≤ kBkL (Y,Z) · kAkL (X,Y ) · kxkX .
| {z }
=1
Continuity of (A, B) 7→ BA follows from
kB1 A1 − B2 A2 kL (X,Z) = kB1 (A1 − A2 ) + (B1 − B2 ) A2 kL (X,Z)

≤ kB1 kL (Y,Z) kA1 − A2 kL (X,Y ) + kA2 kL (X,Y ) kB1 − B2 kL (Y,Z)
−−−−−−−−−−−−−−−−−−−−−−−−→ 0.
kA1 −A2 kL (X,Y ) ,kB1 −B2 kL (Y,Z) →0
Theorem 3.7. If Y is a Banach space, then so is L (X, Y ) .

Remark. Note that we do not require completeness of X.
Proof of Theorem 3.7. Let (Ak )k∈N be a Cauchy sequence. It follows that
sup kAk x − Al xkY −−−−→ 0

kxkX ≤1 k,l→∞
and for any fixed x ∈ X\ {0} , we have

1
kAk x − Al xkY −−−−→ 0.
kxkX k,l→∞
That means, (Ak x)k∈N is a Cauchy sequence in Y . Thus there exists a limit
Y 3 y (x) := lim Ak x.
k→∞
Claim (1). y (x) is linear in x.

Proof (1). Take α ∈ K, x1 , x2 ∈ X. We have
y (αx1 + x2 ) = lim Ak (αx1 + x2 )

k→∞
= lim (αAk x1 + Ak x2 )
k→∞
= α lim Ak x1 + lim Ak x2
k→∞ k→∞
= αy (x1 ) + y (x2 ) .
So y is a linear map and we write

y = Ax.
Claim (2). A is continuous.
22
4 Hahn-Banach theorem and some consequences
Proof (2).

kAxkY = lim A
k→∞ k x
Y
= lim kAk xkY
k→∞
≤ lim sup kAk kL (X,Y ) kxkX
k→∞
≤ L kxkX ,
since Cauchy sequences are bounded.

Claim (3). kAk − AkL (X,Y ) −−−−→ 0.
k→∞
Proof (3).

kAx − Ak xkY = lim Al x − Ak x

l→∞
Y
≤ lim sup kAl − Ak kL (X,Y ) kxkX .
l→∞
If we take the sup over kxk ≤ 1, we get
kAl − Ak kL (X,Y ) ≤ lim sup kAl − Ak k → 0,

l→∞
since (Ak )k∈N is a Cauchy sequence.
Remark. For an interesting application see the exercise with the exponential and Neumann series
and the connection to Voterra’s integral equation.
4 Hahn-Banach theorem and some consequences

Question. Is the dual space rich enough?
Example. X vector space, take x, y ∈ X, x 6= y. Does there exist some f ∈ X 0 : f (x) 6= f (y) .
4.1 Analytic version of the theorem

Theorem 4.1 (Hahn-Banach). Let X be a vector space over the real numbers and take p : X → R
with the following properties:
1. p (λx) = λp (x) ∀λ > 0, x ∈ X (positive homogeneity)
2. p (x + y) ≤ p (x) + p (y) ∀x, y ∈ X (sublinearity)
Furthermore, let G ⊂ X be a linear subspace and let g : G → R be linear with
g (x) ≤ p (x) ∀x ∈ G.
Then there exists a linear map

f :X→R
such that
1. f |G = g (that is f (x) = g (x) ∀x ∈ G).
2. f (x) ≤ p (x) ∀x ∈ X.
We say that f is a continuation (extension) of g.
23
4 Hahn-Banach theorem and some consequences 4.1 Analytic version of the theorem
Remark. On a normed vector space, the norm satisfies the requirements on p, whence the extension
is a bounded linear functional.
Example. Series with increasing difficulty:
1. X = R2 = span {(1, 0) , (0, 1)} .
x̃ ∈ G := span {(1, 0)} ,
g (x̃) = x̃,
1/2
p (x) = kxk2 = x21 + x22 .
Clearly, we have an extension f of g, namely
f (x) = x1 .
2. X = L2 ((0, 1)) Take

φ ∈ X such that kφk2 = 1,
G := span {φ} := {x̃ ∈ X | x̃ = λφ, λ ∈ R} ,
g (x̃) := λ ∀x̃ ∈ G,
p (x) := kxk2 .
We can take ˆ 1
f (x) = x (t) φ (t) dt = (x, φ)L2
0
(bound follows from Cauchy-Schwarz).
3. X := L1 ((0, 1))
ˆ 1
kxk1 = |x (t)| dt.
0
1
Take φ ∈ X, φ (t) = √
2 t
. We thus have kφk1 = 1.
G := span {φ} = {x̃ ∈ X | x̃ = λφ, λ ∈ R} ,

g (x̃) := λ ∀x̃ ∈ G,
p (x) := k·k1 .
Note that ˆ 1
f (x) = x (t) φ (t) dt
0
does not work. However, the Hahn-Banach theorem guarantees the existence of such a
functional
f : X → R, kf k ≤ 1, f |G = g.
4. X = L∞ ((0, 1)) Take the linear subspace of L∞
G := (C ((0, 1)) , k·k∞ ) ,

1
g (x̃) := x̃ ,
2
p (x) := kxkL∞ .
Existence of a continuation is guaranteed, even though a general function in L∞ does not
admit a “value at 1/2”.
Idea of the proof. We consider the set P of all continuations h of g on D (h) (domain of h,
subspace of X) and order the elements of P by set inclusion with respect to their domain
D (h) , that is
h ≤ h0 :⇔ D (h0 ) ⊃ D (h) and h (x) = h0 (x) ∀x ∈ D (h) .
”All” that is left to do then is to find the “largest” extension.
24
Definition 4.2. 1. A set P with a relation ≤ is called partially ordered, if ∀a, b, c ∈ P :
(a) a ≤ a (Reflexivity)
(b) (a ≤ b and b ≤ c) ⇒ a ≤ c (Transitivity)
(c) (a ≤ b and b ≤ a) ⇒ a = b (Antisymmetry)
2. A set Q with a relation ≤ is called totally ordered, if it is partially ordered and we further
have
a ≤ b ∨ b ≤ a ∀a, b ∈ Q.
3. Let R ⊂ P , P partially ordered. An element c ∈ P is called an upper bound of R, if
a ≤ c ∀a ∈ R.
4. An element m ∈ P is called maximal, if for all a ∈ P , we have
m≤a ⇒ m = a.
Remark. • The canonical example for a partial order is set inclusion.

• If for a, b ∈ P we neither have a ≤ b nor b ≤ a, then a and b are called not comparable.
Zorn’s Lemma. Let (P, ≤) be not empty, partially ordered and assume that any totally ordered
subset Q of P admits an upper bound in P. Then P admits a maximal element.
Remark. Zorn’s Lemma is equivalent to the axiom of choice:
Axiom of choice. Take A to be a set of non-empty sets ,then there exists a choice function F on
A, such that
F (X) ∈ X ∀X ∈ A.
Proof of equivalence. Not here.
Proof (Theorem 4.1). Consider
P := {h : D (h) → R | G ⊂ D (h) , h is a continuation of g onto D (h) , h (x) ≤ p (x) ∀x ∈ D (h)}
with the relation (

D (h1 ) ⊂ D (h2 )
h1 ≤ h2 :⇔ .
h1 (x) = h2 (x) ∀x ∈ D (h1 )
Claim (1). P is not empty.
Proof (1). g ∈ P.
Claim (2). “≤” is a partial order on P .
Proof (2). 1. a ≤ a is clear by definition.
2. a ≤ b ∧ b ≤ c ⇒ D (a) ⊂ D (b) ⊂ D (c) and c |D(a)⊂D(b) = a.
3. a ≤ b ∧ b ≤ a ⇒ D (a) = D (b) ∧ a |D(b) = b ⇒ a = b.
Claim (3). Every totally ordered subset of P admits an upper bound in P .
Proof (3). Let
Q := {hj }j∈I ,
[
D (h) := D (hj ) ,
j∈I
h (x) := hj (x) ∀x ∈ D (hj ) .
25
Claim (3a). h is well defined.
Proof (3a). h1 , h2 ∈ Q, x ∈ D (h1 ) ∩ D (h2 ) . Since Q is totally ordered, we have either
h1 ≤ h2 or h2 ≤ h1 .
W.l.o.g. let’s assume the former. ⇒
x ∈ D (h1 ) ⇒ x ∈ D (h2 ) .
Claim (3b). h ∈ P.
Proof (3b). • D (h) is a subspace of X, because
x, y ∈ D (h) ⇒ ∃j ∈ I : x, y ∈ D (hj ) .
Therefore, we have
(αx + y) ∈ D (h) ∀α ∈ R.
Linearity of h follows in the same way.
• h (x) ≤ p (x) on D (h) is also clear, since
∃j ∈ I : x ∈ D (hj ) ⇒ h (x) = hj (x) ≤ p (x) .
• h (x) = g (x) on G is again clear, since for any j ∈ I, we have G ⊂ D (hj ) and hj = g on G.
• hj ≤ h for all hj ∈ Q follows from the definition of h.
With that, all requirements of Zorn’s Lemma are fulfilled and we can deduce existence of a maximal
element f ∈ P.
Claim (4). f is the sought after continuation of g.

Proof (4). Since f ∈ P, the only thing that remains to show is that D (f ) = X. We argue by
contradiction.
Let x0 ∈ X such that x0 ∈
/ D (f ) . Take

D f˜ := D (f ) + span ({x0 }) .
We have
x̃ = x + tx0 ∈ D f˜ ∀t ∈ R.
Set
f˜ (x̃) := f (x) + αt
for a suitable α ∈ R such that f˜ ∈ P .
Claim (4a). Such an α exists.
Proof (4a). We only have to show that f˜ ≤ p which is nothing but
f (x) + αt ≤ p (x + tx0 ) ∀x ∈ D (f ) , t ∈ R
x x
⇔ f +α≤p + x0 (t > 0)
xt xt
⇔ f −α≤p − x0 (t < 0)
t t
by positive homogeneity. This is equivalent to showing that
(
f (x) + α ≤ p (x + x0 )
∀x ∈ D (f ) : (∗)
f (x) − α ≤ p (x − x0 )
26
due to homogeneity of p. However, (using sublinearity of p) we have that ∀x, y ∈ D (f ) :

f (x) + f (y) ≤ p (x + y)
≤ p (x + x0 ) + p (y − x0 )
f (y) − p (y − x0 ) ≤ p (x + x0 ) − f (x) .
We can thus choose α in between the two sides of the inequality, which satisfies (∗).
Therefore, we have f˜ ∈ P. Note that f ≤ f˜ 6= f, in contradiction to maximality of f.
This completes the proof.
Corollary 4.3. Let (X, k·k) be a normed K-vector space (K ∈ {R, C}) and let G be a subspace of
X. Let
g : G → K linear,
kgkG0 := sup |g (x)| .
x∈G
kxkX ≤1
Then there exists a continuation f ∈ X 0 of g such that

kf kX 0 = kgkG0 .
Proof. Exercise.
Corollary 4.4. Let X be a K-vector space. Then ∀x0 ∈ X : ∃f0 ∈ X 0 :
2
kf0 kX 0 = kx0 kX and f0 (x0 ) = kx0 k .
Proof. Take
G := K · x0 := span ({x0 }) ,
2
g (tx0 ) := t kx0 k .
Then
kgkG0 = kx0 k .
With Corollary 4.3 the existence of f0 follows.
Corollary 4.5. Let X be a K-vector space, x, y ∈ X, x 6= y. Then there exists f ∈ X 0 such that
f (x) 6= f (y) .
Proof. Corollary 4.4 with x0 = x − y.
Corollary 4.6. Let X be a normed K-vector space. Then for all x ∈ X we have
kxk = sup |f (x)| = max0 |f (x)| .
f ∈X 0 f ∈X
kf kX 0 ≤1 kf kX 0 ≤1
Proof. Let x 6= 0. We immediately have

sup |f (x)| ≤ 1 · kxk
f ∈X 0
kf kX 0 ≤1
Using Corollary 4.4 there exists f0 ∈ X 0 such that

kf0 kX 0 = kxkX ,
2
f0 (x) = kxkX .
Now let
1
f1 := f0 ,
kxkX
and we have kf1 kX 0 = 1 and f1 (x) = kxkX .
27
4 Hahn-Banach theorem and some consequences 4.2 Separation of convex sets
4.2 Separation of convex sets

In the following, let X be a normed R-vector space (complex at the end of the section).
Definition 4.7. An affine hyperplane is a set of the form
H := {x ∈ X | f (x) = α} ,
where 0 6≡ f : X → R is a linear (not necessarily continuous) map.

H is called hyperplane of the equation
[f = α] .
Proposition 4.8. The hyperplane of the equation [f = α] is closed iff f is continuous.

Proof. Exercise.
Definition 4.9. Consider A, B ⊂ X. The hyperplane H of the equation [f = α] separates A from
B if we have
f (x) ≤ α ≤ f (y) ∀x ∈ A, y ∈ B.
H strictly separates A from B if there exists ε > 0 :
f (x) + ε ≤ α ≤ f (y) − ε ∀x ∈ A, y ∈ B.
Theorem 4.10 (First separation theorem). Consider A, B ⊂ X convex, not empty and disjoint
and let A be open. Then there exists a closed hyperplane that separates A from B.
Lemma 4.11 (Minkowsky Functional). Let C ⊂ X be convex and open with 0 ∈ C. For all x ∈ X,
we define
p (x) := inf α > 0 | α−1 · x ∈ C .

Then p is positively homogeneous of degree 1 and sublinear with respect to vector addition. Fur-
thermore, we have
1. ∃M > 0 : 0 ≤ p (x) ≤ M kxk ∀x ∈ X.
2. C = {x ∈ X : p (x) < 1} .
p is called Minkowsky functional or gauge of C.

Proof. Homogeneity is clear by definition.
1. Let r > 0 : Br (0) ⊂ C. ⇒
1
p (x) ≤ kxk .
r
Property 1. follows.
2. Let x ∈ C, thus ∃ε > 0 : (1 + ε) x ∈ C. With that we have

1
p (x) ≤ < 1.
1+ε
Let p (x) < 1. Then ∃0 < α < 1 :
α−1 x ∈ C.
We thus have
x = α α−1 x + (1 − α) |{z}

0 ∈ C. (convex combination)
| {z }
∈C
∈C
So 2. follows.
28
Sublinearity: Let x, y ∈ X, ε > 0. With homogeneity and 1., we have

x y
∈ C, ∈ C.
p (x) + ε p (y) + ε
Due to convexity, we have
tx (1 − t) y
+ ∈C ∀t ∈ [0, 1] .
p (x) + ε p (y) + ε
Take
p (x) + ε
t :=
p (x) + p (y) + 2ε
and we get
x+y
∈ C.
p (x) + p (y) + 2ε
By 2.
x+y
p < 1.
p (x) + p (y) + 2ε
So (by homogeneity) we have
p (x + y) ≤ p (x) + p (y) + 2ε ∀ε > 0.
The claim follows by arbitrary choice of ε (ε → 0).
Lemma 4.12 (Separation of a point and a convex set). Take C ⊂ X open, non-empty, and convex,
take x0 ∈ X\C. Then there exists f ∈ X 0 :
f (x) ≤ f (x0 ) ∀x ∈ C.
In particular, the hyperplane [f = f (x0 )] separates the set C from {x0 } .
Proof. W.l.o.g. we can assume 0 ∈ C. Consider p (x) , the Minkowsky functional of C. Take
G := Rx0 and g (tx0 ) = t.
Then, we have
g (x) ≤ p (x) ∀x ∈ G.
0
By Hahn-Banach theorem 4.1, there exists f ∈ X with the required properties:
f (x0 ) = 1 and f (x) < 1 ∀x ∈ C.
Proof (Theorem 4.10). Take
[ [ [
C := (A − y) = {x − y} .
y∈B y∈B x∈A
By definition, C is open and 0 ∈

/ C (since by assumption A ∩ B = ∅).
By Lemma 4.12, there exists f ∈ X 0 :
f (z) ≤ 0 ∀z ∈ C.
Therefore,
f (x) ≤ f (y) ∀x ∈ A, y ∈ B,
since
z∈C ⇒ z = x − y, x ∈ A, y ∈ B,
and thus
0 ≥ f (z) = f (x) − f (y) .
With α ∈ R such that
sup f (x) ≤ α ≤ inf f (y) ,
x∈A y∈B
the theorem follows.
29
Theorem 4.13 (Second separation theorem). Take A, B ⊂ X convex, non-empty and disjoint.
Let A be closed and B compact. Then there exists a closed hyperplane that strictly separates A
from B.
Proof. Fix ε > 0. Take
[ [
Aε := A + Bε (0) := {x + y} ,
x∈A y∈Bε (0)
Bε := B + Bε (0) .
With that, Aε , Bε are open, convex, and non-empty.

Claim. For sufficiently small ε, we have Aε ∩ Bε = ∅.
Proof. Assume that for all ε : Aε ∩ Bε 6= ∅. Then there exists (εn )n∈N :
εn → 0, εn > 0,
with the property that there exist (xn )n∈N ∈ AN , (yn )n∈N ∈ B N , such that
kxn − yn k < 2εn .
Using compactness of B, there exists a subsequence
ynk → y ∈ B.
But
∀ε > 0 : ∃x ∈ A : y ∈ Bε (x) .
Thus, we have y ∈ A = A. This is a contradiction to A ∩ B = ∅.
Therefore, by Theorem 4.10 we have a closed hyperplane [f = α] separating Aε from Bε . Thus
we have
f (x + εz) ≤ α ≤ f (y + εz) ∀x ∈ A, y ∈ B, z ∈ B1 (0) .
But that gives us
f (x) + ε kf k ≤ α ≤ f (y) − ε kf k ,
and the claim follows.
Remark. • The most common application of this theorem is to take B = {x0 } .
• Let A, B ⊂ X non-empty, disjoint and convex. Without assumptions, it is only generally

possible to separate A from B in the finite-dimensional case.
• With the following Lemma, one can extend the above theorems to the complex setting. By
taking a real closed hyperplane [Re f = α] , α ∈ R.
Lemma 4.14. Take X to be a Banach space over C, A ⊂ X convex, non-empty and open. Take
x0 ∈ X\A. Then there exists f ∈ X 0 :
Re f (x) ≤ Re f (x0 ) ∀x ∈ A.
Corollary 4.15. Let F ⊂ X be a subspace, such that F 6= X. Then there exists 0 6≡ f ∈ X 0 :
f (x) = 0 ∀x ∈ F.
Remark. This is very useful for proving the denseness of subspaces:
[(∀x ∈ F : f (x) = 0) ⇒ f ≡ 0] ⇒ F = X.
30
5 Baire category argument
Proof. Take x0 ∈ X such that x0 ∈ / F . By Theorem 4.13 with A = F , B = {x0 } , there exists
0 6≡ f ∈ X 0 , such that the hyperplane
[(Re) f = α]
separates F and {x0 } strictly. We have
(Re) f (x) < α < Re f (x0 ) ∀x ∈ F.
Since F is a subspace of X, it follows
λ Re f (x) < α ∀λ ∈ R, x ∈ F
⇒ Re f (x) = 0 ∀x ∈ F
⇒ f (x) = 0 ∀x ∈ F
See exercise 10.
5 Baire category argument

Lemma 5.1 (Baire). Let X be a complete metric space and take a sequence (Fn )n∈N of closed
subsets of X. If
Fn◦ = ∅ ∀n ∈ N,
then !◦
[
Fn = ∅.
n∈N
◦
Remark. • A set F is called nowhere dense, if F = ∅.
• A set M is called meagre (or set of category 1 ), if ∃ (Fn )n∈N , Fn nowhere dense and
[
Fn = M.
n∈N
• Non-meagre sets are called fat (or of category 2 ).

• In particular, every complete metric space is fat.
Remark. The most common application of Baire’s Lemma is the following: Take (An )n∈N a se-
quence of sets in a complete metric space X. If
[ ◦
An = X ⇒ ∃n ∈ N : An 6= ∅.
n∈N
Proof. Let On := Fnc . Then On is open and dense in X. Required to prove is that
\
G := On
n∈N
is still dense in X. Take ω a non-empty open set in X. We will show that
ω ∩ G 6= ∅.
Take x0 ∈ ω, r0 > 0 such that

Br0 (x0 ) ⊂ ω.
Now pick x1 ∈ Br0 (x0 ) ∩ O1 and r1 > 0 such that
r0
Br1 (x1 ) ⊂ Br0 (x0 ) ∩ O1 , 0 < r1 <
2
31
5 Baire category argument 5.1 Banach-Steinhaus theorem
(This is possible, since O1 is open and dense.) In the same manner we can inductively construct
(xn )n∈N ∈ X N , (rn )n∈N ∈ RN
>0 such that ∀n ∈ N :
rn
Brn+1 (xn+1 ) ⊂ Brn (xn ) ∩ On+1 , 0 < rn+1 < .
2
Thus (xn )n∈N is a Cauchy sequence and we have
xn → x ∈ X,
and
xn+p ∈ Brn (xn ) ∀p > 0, n ∈ N.
But then for the limit, we get
x ∈ Brn (xn ) ∀n ∈ N.
Thus x ∈ ω ∩ G.
5.1 Banach-Steinhaus theorem

Theorem 5.2 (Banach-Steinhaus). Let X, Y be normed K-vector spaces, X a Banach space. Let
(Ti )i∈I a family of linear bounded maps from X to Y (not necessarily countably many). Assume
that
sup kTi xk < ∞ ∀x ∈ X.
i∈I
Then we have
sup kTi k < ∞.
i∈I
Remark. Theorem 5.2 is also called uniform boundedness principle, since from the pointwise state-
ment, it follows that
kTi xk ≤ C kxk ∀i ∈ I, x ∈ X.
Proof. Let
Xn := {x ∈ X : ∀i ∈ I : kTi xk ≤ n} .
With that, we have Xn closed and by assumption of pointwise boundedness, we have
[
Xn = X.
n∈N
By Lemma 5.1, there exists n0 ∈ N :

Xn◦0 6= ∅.
Now pick x0 ∈ Xn0 , r > 0 :
Br (x0 ) ⊂ Xn0 .
But then we have
kTi (x0 + rz)k ≤ n0 ∀i ∈ I, z ∈ B1 (0) .
It follows
r · kTi zk ≤ n0 + kTi x0 k ∀i ∈ I, z ∈ B1 (0) ,

n0 + kTi x0 k
kTi k ≤ ∀i ∈ I,
r
and the theorem is proven.
Corollary 5.3. Let X be a Banach space, Y a normed space, (Tn )n∈N a sequence of bounded linear
maps Tn : X → Y :
Tn x −−−−→ T x ∀x ∈ X.
n→∞
Then we have
32
5 Baire category argument 5.2 Open mapping and closed graph theorems
1. supn∈N kTn k < ∞.
2. T ∈ L (X, Y ) .
3. kT k ≤ lim inf n→∞ kTn k .
Proof. Exercise.
Corollary 5.4. Let Z be a Banach space, B ⊂ Z. Assume ∀f ∈ Z 0 we have
[
f (B) := {f (b)}
b∈B
is bounded in K. Then B is bounded set.

Proof. We will use Theorem 5.2with
X = Z 0 , Y = K, I = B.
For any b ∈ B, we define

Tb (f ) = f (b) ∀f ∈ Z 0 .
Thus by assumption we have
sup (Tb (f )) < ∞ ∀f ∈ Z 0
b∈B
Banach-Steinhaus 5.2 yields C > 0 :
|f (b)| ≤ C kf k ∀f ∈ Z 0 , b ∈ B.
With Corollary 4.4, it follows that

kbk ≤ C ∀b ∈ B.
Corollary 5.5. Let Z be a Banach space, B 0 ⊂ Z 0 and the sets
[
f (x) be bounded (in K) ∀x ∈ Z.
f ∈B 0
Then B 0 is bounded (in the operator norm k·kL (Z,K) ).
Proof. Use Theorem 5.2 with X = Z, Y = K, I = B 0 . The conclusion is made as in 5.4.
5.2 Open mapping and closed graph theorems

Theorem 5.6 (Open mapping theorem). Let X, Y be Banach spaces, T : X → Y bounded and
linear and surjective. Then there exists c > 0 :
T (B1 (0)) ⊃ Bc (0) .

| {z } | {z }
in X in Y
Remark. It follows that the image of any open set in X is an open set in Y (under surjective maps).
Such a map is called an open map.
Proof. 1. We first show that for T : X → Y surjective, linear, we have ∃c > 0 :
T (B1 (0)) ⊃ B2c (0) . (∗)
Let
Yn := T (Bn (0)).
By surjectivity of T , we have [
Y = Yn .
n∈N
33
Due to Baire’s Lemma, we have some n0 ∈ N :
Yn◦0 6= ∅.
But we have
1
Bn0 (0) = x ∈ X| x ∈ B1 (0) .
n0
Using linearity of T , we have

1
Yn0 = T (Bn0 (0)) = y ∈ Y | y ∈ T (B1 (0)) .
n0
Therefore, we also have ◦

T (B1 (0)) 6= ∅.
Now pick c > 0, y0 ∈ Y :

B4c (y0 ) ⊂ T (B1 (0)).
By linearity of T , we have that not only y0 ∈ T (B1 (0)), but also −y0 ∈ T (B1 (0)). Thus
S
B4c (0) ⊂ B4c (y0 ) + B4c (−y0 ) (A + B := (x,y)∈A×B {x + y})
⊂ T (B1 (0)) + T (B1 (0)).
Linearity and convexity hold
T (B1 (0)) + T (B1 (0)) = T (B2 (0)).
The claim follows by additional rescaling by a factor 21 .

2. Claim. Assume T is a continuous linear operator from X to Y satisfying (∗). Then we have
T (B1 (0)) ⊃ Bc (0)
for the c > 0 from 1.
Proof. Choose any y ∈ Y such that kyk < c. We need to find x ∈ X :
kxk < 1 and T x = y.
By (∗) we know that ∀ε > 0∃z ∈ X :

1
kzk < and ky − T zk < ε.
2
Choosing ε = 2c , we find z1 ∈ X :
1 c
kz1 k < and ky − T z1 k < .
2 2
By the same construction applied to y − T z1 instead of y and ε = 4c , we find z2 ∈ X :
1 c
kz2 k < and k(y − T z1 ) −T z2 k <
4 | {z } 4
∈Bc/2 (0)
| {z }
∈Bc/4 (0)
This way, we can construct a sequence (zn )n :

!
n
1 X c
kzn k < n and y − T zk < n ∀n ∈ N.

2 2
k=1
34
It follows that
n
X
xn := zk
k=1
is a Cauchy sequence. Let
x = lim xn .
n→∞
Clearly kxk < 1 and

ky − T xk = y − T lim xn = lim ky − T xn k = 0. (by continuity of T )

n→∞ n→∞
Thus
y = T x.
Remark. Both completeness of Y and completeness of X are necessary. Counterexamples are in

the exercises.
Corollary 5.7. Let X, Y be Banach spaces, T : X → Y linear, bounded and bijective. Then T −1
is continuous.
Proof. By Theorem 5.6, we have
kT (x)k < c ⇒ kxk < 1.
By homogeneity of the norm, we have
1
kxk < kT (x)k ∀x ∈ X.
c
So T −1 is bounded, thus continuous.
Remark. Let X be a vector space and k·k1 , k·k2 two norms on X, such that X is complete w.r.t
both norms. Assume that
kxk2 ≤ c kxk1 ∀x ∈ X.
By Corollary 5.7 with T = idX , the norms are already equivalent.
Theorem 5.8 (Closed graph theorem). Let X, Y be Banach spaces, T : X → Y linear. Assume
G (T ) := {(x, T (x)) ∈ X × Y } (graph of T )
to be closed in the product norm. Then, T is continuous.
Remark. We also have T continuous ⇒ G (T ) closed (proven by sequence criterion).
Proof. Consider the two norms
kxk1 := kxkX + kT (x)kY ,
kxk2 := kxkX .
Since we assumed that G (T ) is closed, G (T ), as a closed linear subspace of a Banach space, is a
Banach space, so (X, k·k1 ) is a Banach space as well. Obviously
kxk2 ≤ kxk1 ∀x ∈ X.
By the remark to Corollary 5.7, ∃C > 0 :
kT (x)kY ≤ C · kxkX ∀x ∈ X.
Thus, T is bounded.
Remark. Theorem 5.8 sometimes simplifies the proof of continuity for a linear operator T : X → Y :
Instead of showing
xk → x ⇒ T (xk ) → y and y = T (x) ,
we only need to prove
(xk → x and T (xk ) → y) ⇒ y = T (x) .
35
6 The weak topology
6 The weak topology

So in ∞-dimensional spaces, bounded sequences do not necessarily admit a converging subsequence.
What now?
Theorem 6.1. Let X be a separable Banach space and let (fk )k∈N be a sequence in X 0 with
kfk kX 0 ≤ 1 ∀k ∈ N.
Then there exists f ∈ X 0 and a subsequence fkj j∈N :

f − fk0 j (x) −−−→ 0 ∀x ∈ X.
j→∞
| {z }
∈X 0
0
Example. For X = `p , we have X 0 = `p See Exercise 31. Therefore, immediately any bounded
sequence in lp admits a weak(∗)ly convergent subsequence.
Proof. Let (xn )n∈N be a sequence with dense span in X. We have (fk , (xn ))k∈N bounded K, for
all n ∈ N. Using a diagonal sequence argument, we have (fkj )j∈N , such that for all n ∈ N we have
lim fkj (xn )

j→∞
exists in K.
But thus we have for any y ∈ Y = span({xn }n∈N ) existence of the limit
f (y) = lim fkj (y),

j→∞
for which linearity follows immediately. Since |f (y)| ≤ kyk we have uniform continuity of f on Y .
Thus we can find a unique continuation of f onto all of X 0 (since Y dense in X).
We also call this continuation f , noting that it is linear. It follows that kf k ≤ 1 und
∀x ∈ X, y ∈ Y : |(f − fky )[x]| ≤ |(f − fkj )[x − y]| + |(f − fkj )[y]| ≤ 2kx − yk + |(f − fkj )[y]|
| {z }
→0 (j→∞)
Since Y is dense in X, we can choose kx − yk arbitrarily small and the theorem follows.
Question. Can we generalise this?
6.1 The weak topology σ (X, X 0 )

Let X be a set, fi : X → Yi maps, Yi topological spaces for i ∈ I.
Goal. Find the coarsest topology T on X such that all functions fi are continuous.
Remark. There is such a topology, since with the discrete topology on X all functions are continu-
ous. However, this is not necessarily the coarsest.

Let now wij be the open sets in Yi . In order for all fi to be continuous, we must have
j∈Ji

fi−1 wij ∈ T ∀i ∈ I, j ∈ Ji .
So we need n o
fi−1 wij | i ∈ I, j ∈ Ji ⊂ T .
Proposition 6.2. Let S be a family of subsets of X containing ∅ and X. Let Φ be the set of
subsets of X which can be written as finite intersections of sets in S, i.e.
 
 k
\ 
k
Φ = ϕ ⊂ X | ∃k ∈ N, (sj )j=1 , sj ∈ S, ϕ = sj .
 
j=1
36
6 The weak topology 6.1 The weak topology σ (X, X 0 )
Furthermore, let Ψ be the set of all arbitrary unions of sets in Φ, i.e.

( )
[
Ψ = ψ ∈ X | ∃ (φλ )λ∈Λ , ψ = φλ , φλ ∈ Φ .
λ∈Λ
Then Ψ is the coarsest topology on X that contains all sets in S.

Proof. We have to show that
1. Ψ is a topology on X, that is, it is stable under finite intersections and arbitrary unions and
contains ∅, X.
2. Ψ contains all sets in S.
3. Ψ0 ⊂ Ψ, Ψ0 is a topology on X, S ⊂ Ψ0 , then Ψ0 = Ψ.
1. was Proposition 1.8 (the only thing to show is stability under finite intersections, which follows
as noted below), 2. and 3. are fairly obvious.
Take ψ1 , ψ2 ∈ Ψ. We have
(Aλ )λ∈Λ , (Bκ )κ∈K , Aλ , Bκ ∈ Φ ∀λ, κ
such that [ [
ψ1 = Aλ , ψ2 = Bκ .
λ∈Λ κ∈K
But then we have

! !
[ [
ψ1 ∩ ψ2 = Aλ ∩ Bκ
λ κ
!
[ [
= Aλ ∩ Bκ
λ κ
[[
= (Aλ ∩ Bκ ) ∈ Ψ.
λ κ
Corollary 6.3. Taking n o

S = fi−1 wij | i ∈ I, j ∈ Ji
and Ψ as in Proposition 6.2, we get the required coarsest topology that makes all fi continuous.
Proposition 6.4. Let (X, T ) be a topological space, such that T is the coarsest topology such that
fi : X → Yi is continuous ∀i (fi , Yi as above). Then we have
T
xk −
→x ⇔ fi (xk ) → fi (x) ∀i ∈ I.
Proof. “⇒” True due to continuity of the fi , i ∈ I.

“⇐” Let U ∈ T : x ∈ U. We have
[ K(λ)
\
−1
U= fi(κ,λ) (Vκ,λ ) , Vκ,λ ⊂ Yi(κ,λ) open.
λ∈Λ κ=1
There exists λ0 ∈ Λ :
K(λ0 )
\
−1
x∈ fi(κ,λ 0)
(Vκ,λ0 ) .
κ=1
By assumption, we have
fi(κ,λ0 ) (xk ) → fi(κ,λ0 ) (x) ∀κ ∈ {1, ..., K (λ0 )}
37
⇒ ∃N ∈ N :
fi(κ,λ0 ) (xn ) ∈ Vκ,λ0 ∀n ≥ N, κ ∈ {1, ..., K (λ0 )} .
⇒
K(λ0 )
\
−1
xn ∈ fi(κ,λ 0)
(Vκ,λ0 ) ⊂ U ∀n ≥ N.
κ=1
⇒
T
xn −
→ x.
Proposition 6.5. Let Z be a topological space, ψ : Z → (X, T ) , T as above. Then ψ is continuous,
iff fi ◦ ψ is continuous for all i ∈ I.
Proof. “⇒” clear.
“⇐” Let U ∈ T . Then [\
U= fi−1 (Vi,k ) , Vi,k ∈ Yi .
arb. fin.
Thus we have
[\
ψ −1 (U ) = ψ −1 fi−1 (Vi,k )

a f
−1
[\
= (fi ◦ ψ) (Vi,k )
a f
open in Z.
Definition 6.6. Let X be a topological space.
1. N ⊂ X is called a neighbourhood of x0 ∈ X, if
∃U ∈ T : x0 ∈ U ⊂ N.
2. A family W of open sets in X is called a neighbourhood basis of x0 , if every N ∈ W is a

neighbourhood of x0 and we have: M neighbourhood of x0 ⇒
∃N ∈ W : N ⊂ M.
Definition 6.7. Let X be a Banach space. The weak topology σ (X, X 0 ) is the coarsest topology
that renders all maps in the dual space X 0 continuous as maps from X to K with its usual topology.
Remark. We immediately see that if U is open w.r.t. σ (X, X 0 ) , then U is open w.r.t. the topology
induced by k·kX .
Proposition 6.8. The space (X, σ (X, X 0 )) is a Hausdorff space.
Proof. Let x1 , x2 be in X, x1 6= x2 . By the Hahn-Banach theorem 4.1, there exists f ∈ X 0 :
f (x1 ) 6= f (x2 ) .
So we have U1 , U2 open in K :
U1 ∩ U2 = ∅ and f (x1 ) ∈ U1 , f (x2 ) ∈ U2 .
As preimages of disjoint sets are disjoint, we have
f −1 (U1 ) ∩ f −1 (U2 ) = ∅.
Proposition 6.9. Let x0 ∈ X. A neighbourhood basis w.r.t. σ (X, X 0 ) of x0 can be constructed of

sets of the form
V = {x ∈ X | |fi (x − x0 )| < ε ∀i ∈ I} ,
where I is a finite index set and
fi ∈ X 0 ∀i ∈ I, ε > 0.
38
Proof. It’s clear that all sets of the form of V are open and contain x0 . Let N be a neighbourhood
of x0 w.r.t. σ (X, X 0 ) . Let U ⊂ N, x0 ∈ U, U ∈ σ (X, X 0 ) . We have
[ κ(λ)
\
−1
U= fκ,λ (Vκ,λ ) , fκ,λ ∈ X 0 , Vκ,λ ⊂ K open.
λ∈Λ κ=1
So there exists λ0 ∈ Λ :
κ(λ0 )
\
−1
x0 ∈ fκ,λ 0
(Vκ,λ0 ) .
κ=1
Let now
yκ,λ0 = fκ,λ0 (x0 ) .
We have ε > 0 :
Bε (yκ,λ0 ) ⊂ Vκ,λ0 .
Now consider
κ(λ0 )
\
−1
V := fκ,λ0
(Bε (yκ,λ0 )) = {x ∈ X | |fi (x − x0 )| < ε} .
κ=1
Thus V is of the required form and by construction we have V ⊂ U.

Notation. Let (xn )n∈N be a sequence in X. By
xn * x
we denote convergence of xn to x w.r.t. the weak topology. If there is the danger of ambiguity, we
write
weakly
xn −−−−* x
for convergence w.r.t. σ (X, X 0 ) and
strongly
xn −−−−−→ x
for convergence w.r.t. k·kX . We call the topology induced by k·kX the strong topology.
Proposition 6.10. Let (xn )n∈N be a sequence in a Banach space X and x ∈ X. We have
1. xn * x ⇔ f (xn ) → f (x) ∀f ∈ X 0 .
strongly weakly
2. xn −−−−−→ x ⇒ xn −−−−* x.
3. xn * x ⇒ ∃C > 0 : kxn kX < C ∀n ∈ N and kxk ≤ lim inf n→∞ kxn k .
0
X
4. xn * x, fn −−→ f ⇒ fn (xn ) → f (x) .
Proof. 1. Proposition 6.4
2. Observe
|f (xn ) − f (x)| ≤ kf kX 0 kxn − xk ,
then use “if” from 1. to get the result.
3. By Corollary 5.4, we only need to check that
{f (xn )}n∈N is bounded ∀f.
This is true by assumption of convergence of f (xn ) . For the second statement note
|f (xn )| ≤ kf kX 0 kxn kX and f (xn ) → f (x) ∀f ∈ X 0 .
Thus we have
|f (x)| ≤ kf kX 0 lim inf kxn k .
n→∞
39
4. By the triangle inequality we have
|fn (xn ) − f (x)| ≤ |fn (xn ) − f (xn )| + |f (xn ) − f (x)|

≤ kfn − f k kxn k + |f (xn − x)| .
The first term vanishes due to 3. and strong convergence of fn , the second term does so by
assumption.
Theorem 6.11. In finite dimensional Banach spaces the strong and weak topologies agree.
Proof. Exercise.
Remark. However, in ∞-dimensional Banach spaces, they never agree.
Example. S := {x ∈ X | kxk = 1} , X ∞-dimensional Banach space. We will show
σ (X,X 0 ) k·kX
S ⊃ B1 (0) =: BX =: B.
Let x0 ∈ B, U neighbourhood of x0 w.r.t. to σ (X, X 0 ), then U ∩ S 6= ∅ :

By Proposition 6.9, we can assume
V = {x ∈ X | |fi (x − x0 )| < ε ∀i ∈ {1, ..., n}} , fi ∈ X 0 , n ∈ N, ε > 0.
Now let 0 6= y0 ∈ X :
fi (y0 ) = 0 ∀i ∈ {1, ..., n} .
Such a y0 exists: Consider
φ : X → Kn , φ (x) = (fi (x))i∈{1,...,n} .
If there were no such y0 , φ would only vanish on the origin and thus be injective, so X could be
at most n-dimensional.
Let now
g (t) := kx0 + ty0 k ,
which is continuous and
g (0) = kx0 k ≤ 1 and g (t) −−−→ ∞.
t→∞
Therefore, there exists t0 ∈ R : g (t0 ) = 1 and thus
kx0 + t0 y0 kX = 1 ; x0 + t0 y0 ∈ S,
and
x0 + t0 y0 ∈ V.
Example. B1 (0) is not open w.r.t. σ (X, X 0 ), if X is ∞-dimensional. Indeed, we even have
◦σ X,X 0 )
(B1 (0)) ( = ∅.
Proof works exactly as above.

Remark. • The weak topology on an ∞-dimensional space is never metrisable (i.e. not induced
by any metric).
• Two metrics on X that induce the same converging sequences also induce the same topo-
logy. However this is not necessarily true if only the topologies induce the same converging
sequences.
Example. X = l1 . We have
0
k·kl1 σ (l1 ,(l1 )
)
xn −−−→ x ⇔ xn −−−−−−−* x.
Luckily, such examples are “rare”.
40
Theorem 6.12. Let C ⊂ X convex. Then C is strongly closed, if and only if it is weakly closed.
Remark. Together with the example above, this shows that

σ (X,X 0 ) k·kX
S = B1 (0) .
Proof. “⇐” We always have
U weakly open ⇒ U strongly open.
Thus, if C is weakly closed, it is strongly closed.

“⇒” / C. We need to show ∃U ∈ σ (X, X 0 ) :
Let C be strongly closed and convex, x0 ∈
x0 ∈ U and U ∩ C = ∅.
By the second separation theorem 4.13, there exists f ∈ X 0 , α ∈ R :
Re f (x0 ) < α < Re f (x) ∀x ∈ C.
Let now
U := {x ∈ X | Re f (x) < α} .
Then U ∈ σ (X, X 0 ) :
x0 ∈ U and U ∩ C = ∅.
Lemma 6.13 (Mazur). Let (xn )n∈N be a sequence in X,
weakly
xn −−−−* x.
n
Then there exists a sequence (yn )n∈N of convex combinations of (xj )j=1 , i.e.
n
X n
X
yn = λn,j xj , λn,j = 1, λn,j ≥ 0 ∀n, j,
j=1 j=1
such that
strongly
yn −−−−−→ x.
Proof. Exercise.
Remark. In particular, we have
x ∈ conv {xn }n∈N . (convex hull)
Theorem 6.14. Let X, Y be Banach spaces, T : X → Y linear. Then T is continuous w.r.t.

the strong topology on both spaces, if and only if it is continuous w.r.t. the weak topology on both
spaces.
Proof. “⇒” By Proposition 6.5, it suffices to show ∀f ∈ Y 0 :
F : x 7→ f (T x)
is continuous w.r.t. σ (X, X 0 ) as a map to K. This, however, is clear, since
f (T x) ∈ X 0 .
“⇐” The graph G (T ) of T is closed in X × Y w.r.t. σ (X, X 0 ) ⊗ σ (Y, Y 0 ), since, by assump-

tion, T is continuous. By Theorem 6.12, G (T ) is also strongly closed and the closed
graph theorem 5.8 yields the result.
Remark. In general, this does not hold for non-linear functions.
41
6 The weak topology 6.2 The weak* topology σ(X 0 , X)
6.2 The weak* topology σ(X 0 , X)

Let X be a Banach space, X 0 the dual space of X endowed with its usual norm
kf kX 0 = sup |f (x)| .
x∈X
kxk≤1
Let further X 00 be the bidual of X, i.e. the dual space of X 0 endowed with the norm
kξkX 00 = sup |ξ (f )| .
f ∈X 0
kf k≤1
Definition 6.15. The canonical injection J : X → X 00 is given by
x 7→ J (x) , J (x) (f ) := f (x) .
Remark. 1. For fixed x ∈ X, the map f 7→ f (x) is continuous as a map from X 0 to K and also
linear, so J (x) is indeed a continuous linear form on X 0 .
2. J is isometric, since
kJ (x)kX 00 = sup |J (x) (f )|

kf k≤1
= sup |f (x)|
kf k≤1
= kxkX (Hahn-Banach 4.1)
This also implies injectivity.

3. There are examples where J is not surjective (see exercise). One can, however, always identify
the subspace J (X) of X 00 with X.
On X 0 we have already defined two topologies
1. The strong topology, induced by k·kX 0 ,
2. The weak topology σ (X 0 , X 00 ) .
We now introduce a third topology.

Definition 6.16. The weak* topology σ (X 0 , X) on X 0 is the coarsest topology on X 0 that renders
all maps of the form
ϕX : X 0 → K, f 7→ ϕX (f ) = (Jx) (f ) = f (x)
continuous for all x ∈ X.
Remark. 1. If dim X < ∞, weak, weak* and strong topologies all agree.
2. If J (X) is a strict subspace of X 00 (i.e. J (X) 6= X 00 ), then the weak topology σ (X 0 , X 00 ) is

strictly finer than the weak* topology σ (X 0 , X) .
Example. Let ξ ∈ X 00 \J (X) . Then
H := {f ∈ X 0 | ξ (f ) = 0}
is closed w.r.t. σ (X 0 , X 00 ), but not w.r.t. σ (X 0 , X) .

Note. convex, strongly closed ⇒ weakly closed, however convex, strongly closed 6⇒ weakly* closed.
Proposition 6.17. The space (X 0 , σ (X 0 , X)) is Hausdorff.
42
Proof. Let f1 , f2 ∈ X 0 , f1 6≡ f2 . ⇒ ∃x ∈ X :
f1 (x) 6= f2 (x) .
Without loss of generality, for some α ∈ R, we have
Re f1 (x) < α < Re f2 (x) .
Now let
U1 := {f ∈ X 0 | Re f (x) < α} ,
U2 := {f ∈ X 0 | Re f (x) > α} .
f1 ∈ U1 , f2 ∈ U2 U1 , U2 open, U1 ∩ U2 = ∅.
Proposition 6.18. Let f0 ∈ X 0 . A neighbourhood basis w.r.t. the weak* topology is given by sets
of the form
V = {f ∈ X 0 | |(f − f0 ) (xj )| ≤ ε ∀j ∈ J} , |J| < ∞, xj ∈ X∀j, ε > 0.
Proof. Same as the proof of Proposition 6.9.
Notation. Let (fn )n∈N be a sequence in X 0 . We write
∗
fn * f,
if fn converges to f ∈ X 0 w.r.t. the weak* topology.
Proposition 6.19. Let (fn )n∈N be a sequence in X 0 . We have
∗
1. fn * f ⇔ fn (x) → f (x) ∀x ∈ X.
k·k
X 0 ∗
2. fn −−−→f ⇒ fn * f.
∗
3. fn * f ⇒ fn * f.
∗
4. fn * f ⇒ (kfn k)n∈N is bounded in R and kf k ≤ lim inf n→∞ kfn k .
∗ strongly
5. fn * f, xn −−−−−→ x ⇒ fn (xn ) → f (x) .
Proof. As in Proposition 6.10. 3. follows from J (X) being a subspace of X 00 (σ (X 0 , X) is certainly
not finer than σ (X 0 , X 00 )).
∗
Remark. A counterexample to fn * f, xn * x ⇒ fn (xn ) → f (x) can be found in the exercises (in
0 00
l2 , where l2 = l2 = l2 ).
Now, we look at the fundamental reason, why we deal with weak* topologies.
Theorem 6.20 (Banach-Alaoglu). The set
{f ∈ X 0 | kf kX 0 ≤ 1} = B1 (0) ⊂ X 0
is compact w.r.t. the weak* topology σ (X 0 , X) .
Remark. The proof uses Tychonov’s theorem, that says, that the product space of compact topo-
logical spaces (even uncountably many) is compact w.r.t. the product space topology.
Definition 6.21 (Product space topology). Let (Yi )i∈I be a family of topological spaces and let
Y
Y := Yi := (yi )i∈I | yi ∈ Yi ∀i ∈ I .
i∈I
The product topology on Y is the coarsest topology that renders all maps of the form

ϕi : Y → Yi , ϕi (y) = ϕi (yi )i∈I = yi (projections)
continuous.
43
Remark. For finite products (|I| < ∞), this topology agrees with the box topology. This is not the
case in ∞-dimensional spaces. In particular, the following theorem is untrue for the box topology.
Theorem 6.22 (Tychonov). Let
Y
X= Xi , Xi compact ∀i ∈ I.
i∈I
Then X is compact w.r.t. the product topology.

Remark. We will prove Tychonov’s theorem by means of Alexandrov’s subbase theorem.
Theorem 6.23 (Alexandrov). Let (X, T ) be a topological space, B a subbasis of T . If every cover
of X by sets in B admits a finite subcover, X is compact.
Proof. We argue by contradiction.
• Assume X is not compact, however every cover by sets in B admits a finite subcover.
• Let P be the set of all open covers of X that do not admit a finite subcover.
• We endow P with the partial order of set inclusion.
– P is not empty.
– P is partially ordered.
Let C be a chain in P (i.e. a totally ordered subset of P ). We have an upper bound of C by

taking [ [
S := Vj = {U | U ∈ Vj } .
Vj ∈C Vj ∈C
Claim. S ∈ P.
Proof. Assume U1 , ..., Un is an open subcover of X by sets in S. We have
∀i ∈ {1, ..., n}∃Vji ∈ C : Ui ∈ Vji .
However, {Vji }i=1...n ⊂ C is totally ordered, thus there is a maximal element Vj0 . We have
Ui ∈ Vj0 ∀i = 1...n.
But then {Ui }i=1...n is a finite subcover by sets in Vi0 . This contradicts Vi0 ∈ P.
By Zorn’s Lemma, we thus have a maximal element in P , i.e. ∃M ∈ P :
A ∈ P, M ⊂ A ⇒ A = M.
This maximal element has the following properties

1. U ∈
/ M open ⇒ M ∪ {U } ∈
/ P ⇒ M ∪ {U } admits a finite subcover that must contain U .
n
Thus there exist {Uj }j=1 in M :
n
[
X= Uj ∪ U.
j=1
2. For U open
n
[
X= Uj ∪ U ⇒ U∈
/ M.
j=1
3. If U1 , ..., Un ∈
/ M , then
n
\
Uj ∈
/ M.
j=1
44
Proof. ∀j ∈ {1, ..., n}∃Vj,k ∈ M, k = 1...lj :

lj
[
Uj ∪ Vj,k = X. (by 1.)
k=1
Let x ∈ X. Then either x ∈ Vj,k for some j, k or

x ∈ Uj ∀j ∈ {1, ..., n}.
With that, we have    
n lj
n [
\ [
X= Uj  ∪  Vj,k 
j=1 j=1 k=1
| {z }
finite
and by 2.
n
\
Uj ∈
/ M.
j=1
4. U ∈
/ M, U ⊂ W open ⇒ W ∈
/ M. (follows from 1. and 2.)
We will now apply the subbasis.
Claim. B ∩ M is a cover of X.
Proof. Let x ∈ X. We want to show:
∃A ∈ B ∩ M : x ∈ A.
Clearly, there exists U ∈ M : x ∈ U. U is open, so we can write it as
[\
U= Bj , Bj ∈ B.
arb fin
n
Thus there exists {Bj }j=1 in B :
n
\
x∈ Bj ⊂ U.
j=1
T 4.
We now necessarily have a j0 : Bj0 ∈ M (otherwise, by 3. j Bj ∈ ⇒U∈
/M = / M ). But then,
x ∈ Bj0 ∈ B ∩ M.
By assumption, B ∩ M ⊂ B contains a finite subcover of X. This subcover is also a finite
subcover to M , contradicting M ∈ P . Hence P is empty and the proof is complete.
Proof (Tychonov 6.22). Let Xi , i ∈ I be compact topological spaces,
Y
X= Xi .
i∈I
We need to prove that X is compact w.r.t. the product topology.

Let thus Γ be an open cover of X. By Theorem 6.23 we may assume Γ only consists of sets of
an arbitrary subbasis of the product topology. Such a subbasis is given by
S := U ⊂ X : U = ϕ−1

i (Ui ) , Ui ⊂ Xi open, i ∈ I . (ϕi projection on Xi )
Let Si be the subset of S generated by a particular ϕi :
Si := U ⊂ X | U = ϕ−1

i (V ) , V ⊂ Xi open , i ∈ I.
We have [
S= Si .
i∈I
45
Claim. There exists i0 ∈ I : Γ ∩ Si0 is a cover of X.

Proof. Assume the contrary, namely
∀i ∈ I : ∃yi ∈ X : yi ∈
/ A ∀A ∈ Γ ∩ Si .
Take
xi := ϕi (yi ) ∈
/ ϕi (A) ∀A ∈ Γ ∩ Si . (note that ϕ−1
i ϕi (A) = A)
Let now x := (xi )i∈I . We have x ∈ X, but
[
x∈
/ A ∀A ∈ Γ ∩ Si = Γ ∩ S = Γ,
i∈I
in contradiction to Γ being a cover of X.

We now have
Γ ∩ Si0 = ϕ−1

i0 (Vj ) ⊂ X | j ∈ J
for some sets
Vj ⊂ Xi0 open ∀j ∈ J.
Since Γ ∩ Si0 covers X, we have [
Vj = Xi0 .
j∈J
Xi0 is compact, thus we can select a finite Subset

n
[
n
{Vk }k=1 ⊂ {Vj }j∈J : Vk = Xi0 .
k=1
Therefore,
{ϕi0 (Vk ) ⊂ X | k ∈ {1, ..., n}}
is a finite subcover of X to Γ.
Proof (Banach-Alaoglu 6.20). Let X be a Banach space. Consider the product space
Y = KX = ω = (ωx )x∈X | ωx ∈ K

with the product topology and X 0 with the weak* topology σ (X 0 , X) . Consider further the map
Φ : X 0 → Y, f 7→ (f (x))x∈X .
Claim. Φ is a homeomorphism from X 0 to Φ (X 0 ) (bijective, continuous, with continuous inverse).
Proof. 1. Continuity of Φ is follows from continuity of
f 7→ (Φ (f ))x = f (x)
for any x ∈ X by Proposition 6.5.
2. Injectivity: Let f1 , f2 ∈ X 0 , f1 6= f2 .
⇒ ∃x ∈ X : f1 (x) 6= f2 (x)
⇒ (Φ (f1 ))x 6= (Φ (f2 ))x
⇒ Φ (f1 ) 6= Φ (f2 ) .
3. Continuity of the inverse: Again, by 6.5 it suffices to show that

ω 7→ Φ−1 (ω) (x)
is continuous for any x ∈ X. But
ω 7→ Φ−1 (ω) (x) = ωx
is continuous by definition of the product topology.
46
6 The weak topology 6.3 Reflexive spaces and separable spaces
We have
Φ B1 (0) = K , B1 (0) ⊂ X 0 ,
where
K := {ω ∈ Y | |ωx | ≤ kxk , ωx+y = ωx + ωy , ωλx = λωx ∀x, y ∈ X, λ ∈ K} .
With that, we have K = K1 ∩ K2 , where
Y
K1 := [− kxk , kxk] , (compact by Theorem 6.22)
x∈X
   
\ \
K2 :=  Ax,y  ∩  Bx,λ  ,
x,y∈X x∈X,λ∈K
Ax,y := {ω ∈ Y | ωx + ωy − ωx+y = 0} , x, y ∈ X,
| {z }
continuous function in ω
Bx,λ := {ω ∈ Y | ωλx − λωx = 0} , x ∈ X, λ ∈ K.
| {z }
continuous in ω
All Ax,y , Bx,λ are closed as continuous preimages of {0}, and thus K2 is closed as an intersection of
closed sets. Therefore, K is compact as an intersection of a compact and a closed set and finally,
X 0 ⊃ B1 (0) = Φ−1 (K )
is compact as a continuous image of a compact set.
6.3 Reflexive spaces and separable spaces

For this section denote by BX the strongly closed unit ball in a Banach space X.
Definition 6.24. Let X be a Banach space and let J be the canonical injection in the bidual
space X 00
J : X → X 00 , x 7→ J (x) : X 0 → K, f 7→ f (x) .
X is called reflexive, if
J (X) = X 00 .
In reflexive spaces, we can thus identify X with X 00 .
Remark. It is necessary to use the canonical injection, since there are Banach spaces X, that are
not reflexive, however do admit a surjective isometry X → X 00 .
Lemma 6.25 (Helly). Let X be a Banach space, f1 , ..., fn ∈ X 0 , α1 , ..., αn ∈ K. Then the following
are equivalent:
1. ∀ε > 0 : ∃xε ∈ X, kxε k ≤ 1 :
|fi (xε ) − αi | < ε.
Pn Pn
2. | i=1 βi αi | ≤ k i=1 βi fi kX 0 ∀β1 , ..., βn ∈ K.
Proof. Exercise.

Lemma 6.26 (Goldstine). Let X be a Banach space. Then J BX is dense in BX 00 w.r.t.
σ (X 00 , X 0 ) (weak* topology on X 00 ).
Proof. Let ξ ∈ BX 00 and V be a neighbourhood of ξ w.r.t. σ (X 00 , X 0 ) . We have to show that

J BX ∩ V 6= ∅.
By construction of the neighbourhood basis in σ (X 00 , X 0 ) we can assume that V is of the type
V := {η ∈ X 00 | |(η − ξ) (fi )| < ε ∀f1 , ..., fn ∈ X 0 } , ε > 0.
47
We are thus looking for some x ∈ BX :
|fi (x) − ξ (fi )| < ε ∀i ∈ {1, ..., n} .
Let
αi := ξ (fi ) ∈ K.
Since kξk ≤ 1, we have ∀β1 , ..., βn ∈ K :
! n
X n n
X X
βi αi = ξ β i fi ≤ βi fi .

i=1 i=1 i=1 X0
By Lemma 6.25 there is xε ∈ BX :
|fi (xε ) − αi | < ε ∀i ∈ {1, ..., n} .
So
J (xε ) ∈ J BX ∩ V.
Theorem 6.27 (Kakutani). Let X be a Banach space. X is reflexive, if and only if BX is compact
w.r.t. the weak topology σ (X, X 0 ) .
Proof. “⇒” By reflexivity
J BX = BX 00 = {ξ ∈ X 00 | kξkX 00 ≤ 1} .

Thus J BX is compact w.r.t. σ (X 00 , X 0 ) (by Banach-Alaoglu 6.20).

Hence it is enough to show that
J −1 : (X 00 , σ (X 00 , X 0 )) → (X, σ (X, X 0 ))
is continuous. However, we have
f ◦ J −1 : (X 00 , σ (X 00 , X 0 )) → K, f J −1 (ξ) = ξ (f )

is continuous ∀f ∈ X 0 . Thus J −1 is continuous and
J −1 BX 00 = BX

is compact.
“⇐” Let BX be compact w.r.t. σ (X, X 00 ) . By Theorem 6.14 (strong-strong continuity ⇔
weak-weak continuity), J is continuous as a map
(X, σ (X, X 0 )) → (X 00 , σ (X 00 , X 000 )) .
Since σ (X 00 , X 000 ) is finer than σ (X 00 , X 0 ), it is also continuous as a map
(X, σ (X, X 0 )) → (X 00 , σ (X 00 , X 0 )) .
(There are “no more” elements in X 0 than in X 000 .)

Therefore J BX is compact w.r.t. σ (X 00 , X 0 ) and dense by Lemma 6.26. Compact

sets are closed in Hausdorff spaces, so we have

BX 00 = J BX = J BX .
By linearity of J, we further have
J (X) = X 00 .
48
1. J BX is always strongly closed in X 00 .

Remark.

In general, however, J BX is not strongly dense in BX 00 .
2. Finite dimensional vector spaces are reflexive.
Lemma 6.28. Let X be a reflexive Banach space, M ⊂ X a closed subspace. Then M is reflexive
w.r.t. its induced norm.
Proof. On M , we have defined two weak topologies:

1. The topology σ (M, M 0 ) .
2. The subspace topology of σ (X, X 0 ) .
By restriction or extension of continuous linear functionals, those two topologies agree.

However, we have BM is compact w.r.t. σ (M, M 0 ) as a weakly closed subset of a compact set.
Corollary 6.29. Let X be a Banach space. X is reflexive, if and only if X 0 is reflexive.
Proof. “⇒” By Banach-Alaoglu 6.20, BX 0 is compact w.r.t. σ (X 0 , X), however, by reflex-
ivity of X :
σ (X 0 , X) = σ (X 0 , X 00 ) .
Thus X 0 is reflexive by Theorem 6.27.
“⇐” The same argument as above holds X 00 is reflexive. By 6.28, J (X) is reflexive as a
closed subspace of X 00 . f , however, is certainly a surjective isometry
X → J (X) .
Therefore, X is reflexive, since
(T : X → Y surjective Isometry) ⇒ (X reflexive ⇔ Y reflexive) .
Corollary 6.30. Let X be a reflexive Banach space, K ⊂ X strongly closed, convex and bounded.
Then K is compact w.r.t. σ (X, X 0 ) .
Proof. ∃m > 0 :K ⊂ m · BX and K is weakly closed. Therefore, K is compact as a closed subset

of a compact set.
Question. What are we going to use this for? We will prove:
Let φ be a convex and strongly lower semicontinuous function, i.e.
xi → x ⇒ lim inf φ (xi ) ≥ φ (x) .
Then φ admits a minimum on any convex bounded set.

Definition 6.31. Let X be a topological space, A ⊂ X. A function
φ : A → (−∞, ∞]
is called lower semicontinuous, if
{y ∈ X | φ (y) > α} ⊂ X open ∀α ∈ R.
Note. On metric spaces, this is the same as sequential lower semicontinuity:
(xi → x in X) ⇒ lim inf φ (xj ) ≥ φ (x) .

j→∞
49
Theorem 6.32. Let X be a reflexive Banach space, A ⊂ X convex, closed, not empty, and let
φ : A → (−∞, ∞]
be a strongly lower semicontinuous function that is convex and 6≡ ∞.

In case A is not bounded, assume in addition that
lim φ (x) = ∞.
kxk→∞,x∈A
Then φ attains its minimum on A, i.e. ∃x0 ∈ A :
φ (x0 ) ≤ φ (x) ∀x ∈ A.
Proof. Let
Ã := {x ∈ A | φ (x) ≤ λ0 } , λ0 = φ (a) < ∞, a ∈ A.
Ã is thus bounded by growth, convex by convexity and closed by strong lower semicontinuity of φ.
Claim. φ attains its minimum on Ã.
Note. If we prove this claim, we are done, since
φ (x) > φ (x0 ) ∀x ∈ Ã.
Proof. Take a sequence (xk )k∈N in Ã :

1
φ (xk ) ≤ inf φ (x) + .
x∈Ã k
By compactness of Ã (by Corollary 6.30), this sequence admits an accumulation point x0 ∈ Ã.
We continue to argue by contradiction. Assume thus ∃c > 0 :
φ (x0 ) ≥ inf φ (x) + c.

x∈Ã
The set
c
U := y ∈ Ã | φ (y) > inf φ (x) +
x∈Ã 2
is weakly open, since its complement is convex and strongly closed, thus weakly closed.
We have x0 ∈ U , thus U is an open neighbourhood of x0 . However, only finitely many elements
of the sequence can be in U .
Example. u in a Sobolev space W is a minimizer of

ˆ ˆ
2
φ (a) = |∇u| − f · u, D ⊂ W bounded
D D
⇔ −∆u = f.
Now we turn our attention to separable spaces, which will enable us to extract converging
subsequences, not just find accumulation points.
Lemma 6.33. Subsets of separable metric spaces are separable.
Proof. Take (xn )n∈N dense in X (X separable metric space). Take ∅ =
6 Y ⊂ X, 0 < rn → 0. We
can then pick
ym,n ∈ Brn (xm ) ∩ Y.
This set is countable and dense in Y .
Theorem 6.34. Let X be a Banach space. We have
X 0 separable ⇒ X separable.
50
Remark. The converse does not necessarily hold. e.g. L1 (next week).
Proof. Let (fn )n∈N be a dense sequence in X 0 . Since
kfn kX 0 = sup |fn (x)| ,

kxkX ≤1,x∈X
there exist xn ∈ X :
1
kxn kX = 1, kfn k ∀n.
fn (xn ) ≥
2
Let L0 be the set of all rational, finite linear combinations of elements in {xn }n∈N . This set is
countable and clearly dense in
L := span {xn }n∈N
the set of real finite linear combinations of the xn .
Claim. L is dense in X.
Proof. Let f ∈ X 0 :
f (x) = 0, ∀x ∈ L.
Therefore, we have ∀ε > 0 : ∃ {fn }n dense, countable subset:
kf − fn k < ε
and we have
1
kfn k ≤ fn (xn ) = (fn − f ) (xn ) + f (xn ) .
2 | {z } | {z }
≤ε =0
⇒ kf k < 3ε ∀ε > 0.
⇒ kf k = 0.
We have thus shown that ∀f ∈ X 0 , f (x) = 0 ∀x ∈ L :
f ≡ 0.
The statement of Corollary 4.15 says exactly that L must be dense in X.

In the complex case, as usual we consider the characterization of complex linear functionals as
f (x) = g (x) − ig (ix)
with g : X → R linear.
Corollary 6.35. Let X be a reflexive, separable Banach space. Then X 0 is separable.
6.34
Proof. X separable, reflexive. ⇒ X 00 separable ==⇒ X 0 separable.
Theorem 6.36. Let X be a Banach space. X is separable, if and only if the weak* topology on BX 0
is metrisable (that is, there exists a metric d : BX 0 × BX 0 → R that induces the weak* topology).
Remark. The weak* topology is never metrisable on the whole space if X is ∞-dimensional.
Proof. “⇒” Let (xn )n∈N be dense in BX . For f, g ∈ BX 0 , set
∞
X 1
d (f, g) := |(f − g) (xn )| .
n=1
2n
Claim (1). d is a metric.
Proof. • Positive definiteness follows from the density of xn .
51
• Triangle inequality and symmetry follow from respective properties of |·| as seen
in exercises.
In the following, we will prove that if
V open neighbourhood of x0 w.r.t. σ (X, X 0 ),
then there exists
U ⊂ V open neighbourhood of x0 w.r.t. d (·, ·)
and vice versa.

Question. Why is that enough?
Take a set O open w.r.t. σ (X 0 , X) then ∀f ∈ O there is an open neighbourhood V of
f w.r.t. σ (X 0 , X) in O. The claim provides an open neighbourhood U ⊂ V ⊂ O of f
w.r.t. d (·, ·), so
O = O◦d (interior of Ow.r.t. d)
and vice versa.
Claim (2). Take f0 ∈ BX 0 , U a neighbourhood of f0 w.r.t. σ (X 0 , X) . Then ∃r > 0 :

U := f ∈ BX 0 | d (f0 , f ) < r ⊂ V.
Proof. We can, as usual, take V of the form

V = f ∈ BX 0 | |(f − f0 ) (yi )| < ε ∀i ∈ {1, ..., k} , yi ∈ X, ε > 0.
Without loss of generality, we can assume
kyi k = 1 ∀i.
Now for all i ∈ {1, ..., k} we pick ni ∈ N :

ε
kyi − xni k < .
4
We also pick r > 0 :
ε
2 ni · r < ∀i ∈ {1, ..., k} .
2
For f : d (f, f0 ) < r, we have
1
|(f − f0 ) (xni )| < r ∀i ∈ {1, ..., k} .
2 ni
But then we have
|(f − f0 ) (yi )| = |(f − f0 ) (yi − xni ) + (f − f0 ) (xni )|

| {z } | {z } | {z }
k·k≤2 k·k< 4ε k·k≤2ni r< 2ε
ε ε
< 2 · + = ε.
4 2
Claim (3). Take f0 ∈ BX 0 , r > 0. Then there exists a neighbourhood V of f0 w.r.t.
σ (X 0 , X) :
V ⊂ U := f ∈ BX 0 | d (f, f0 ) < r .
52
Proof. We have
∞
X 1
d (f, f0 ) = |(f − f0 ) (xj )|
j=1
2j
k ∞
X 1 X 1
= j
|(f − f0 ) (x j )| + |(f − f ) (xj ) | .
j | {z 0} |{z}
j=1
2 2
j=k+1
k·k≤2 k·k≤1
| P∞
{z }
2 1
≤ j=k+1 2j =
2k−1
Now take n r o
V := f ∈ BX 0 | |(f − f0 ) (xi )| < ∀i ∈ {1, ..., k} ,
2
where we pick k :
1 r
< .
2k−1 2
Then, for all f ∈ V :
k
X 1 1
d (f, f0 ) ≤ |(f − f0 ) (xj )| + k−1 < r.
j |
j=1
2 {z } 2
< r2
| {z }
< r2
| {z }
< r2
Therefore V ⊂ U.
This proves, that in a separable Banach space the weak* topology on BX 0 is metrisable.
“⇐” Take
1
Un := f ∈ BX 0 | d (f, 0) < , n ∈ N.
n
By assumption, there exists a neighbourhood Vn of the origin w.r.t. σ (X 0 , X) :
V n ⊂ Un .
We can take
Vn := f ∈ BX 0 | |f (x)| < εn ∀x ∈ Φn ,
where Φn ⊂ BX finite. Thus, the Set
[
D := Φn
n∈N
is countable.
Claim. D is dense in BX .
Proof. Assume that

f (x) = 0 ∀x ∈ D.
Then
f ∈ Vn ∀n
\
⇒ f∈ Vn
n∈N
\
⇒ f∈ Un
n∈N
⇒ d (f, 0) = 0
⇒ f ≡ 0.
By corollary 4.15, D is dense in BX , since span D is dense in X.
53
Theorem 6.37. Let X be a Banach space. Then X 0 is separable, if and only if BX is metrisable
w.r.t. σ (X, X 0 ).
Proof. Exactly as in the proof of 6.36:
X 0 separable BX , σ (X, X 0 ) metrisable.

⇒
Remark. The other direction holds as well, but the proof is much harder.
Corollary 6.38. Let X be a separable Banach space, (fn )n∈N a bounded sequence in X 0 . Then
there exists a weakly* convergent subsequence of (fn )n∈N .
Proof. Since K · BX 0 is compact and metric, there exists a convergent subsequence of
{fn }n∈N ⊂ K · BX 0 .
Remark. See the exercises for an example where X is not separable and a bounded sequence in X 0
does not admit a convergent subsequence.
Theorem 6.39. Let X be a reflexive Banach space and let (xn )n∈N be a bounded sequence in X.
Then there exists a weakly converging subsequence.
Proof. Let
M0 := span {xn }n∈N and M := M0 .
M is separable and reflexive as a closed subspace of a reflexive space. ⇒ M 00 is separable ⇒ M 0
separable. Therefore
BM , σ (M, M 0 )

is metrisable and weakly compact as the closed unit ball in a reflexive space. Thus, there exists a
subsequence of (xn )n∈N that converges w.r.t. σ (M, M 0 ). ∀f ∈ X 0 : ∃f˜ ∈ M 0 :
f |M = f˜.
Therefore, the subsequence also converges w.r.t. σ (X, X 0 ) .

We even have the converse statement:
Theorem 6.40 (Eberlein-Shmulian). Let X be a Banach space, such that every sequence admits
a weakly converging subsequence. Then X is reflexive.
Proof. See Rudin.

Definition 6.41. A Banach space is called uniformly convex, if ∀ε > 0 : ∃δ > 0 :
kx + yk
∀x, y ∈ X, kxk ≤ 1, kyk ≤ 1, kx − yk > ε : < 1 − δ.
2

Example. X = R2 , k·kp ,
(
p p 1/p
(|x1 | + |x2 | ) , 1 ≤ p < ∞
kxkp = , x = (x1 , x2 ) ∈ X.
max {|x1 | , |x2 |} , p = ∞
The unit balls X are uniformly convex ∀1 < p < ∞.

Theorem 6.42 (Milman-Pettis). Every uniformly convex Banach space is reflexive.
54
Proof. Let ξ ∈ X 00 . W.l.o.g. assume kξkX 00 = 1. We show

ξ ∈ J BX .

Since J BX strongly closed, it is enough to show, that ∀ε > 0 : ∃x ∈ BX :
kξ − J (x)kX 00 < ε.
Let now ε > 0 and fix δ to be the constant from uniform convexity. Let f ∈ X 0 :
δ
|ξ (f )| > 1 −
2
(this f exists since
kξkX 00 = sup |ξ (f )| = 1.)
kf k≤1
f ∈X 0
Let
00 δ
V := η ∈ X | |(η − ξ) (f )| <
2
00 0

be a neighbourhood of ξ w.r.t. σ (X , X ) . Since J BX is weakly* dense in BX 00 (by Goldstine),
we have
V ∩ J BX 6= ∅.
Now let x ∈ BX :
J (x) ∈ V.
Claim. kξ − J (x)kX 00 < ε.
Proof. Assume for a contradiction that
c
ξ ∈ Bε (J (x)) =: W.
With that, W is an open neighbourhood of ξ w.r.t. σ (X 00 , X 0 ), since εBX 00 is weakly* compact

and thus closed. Therefore, we have

(V ∩ W ) ∩ J BX 6= ∅.
Now we pick x̂ ∈ BX :
J (x̂) ∈ V ∩ W.
We have
δ δ
|f (x) − ξ (f )| < , |f (x̂) − ξ (f )| <
2 2
and thus
2 |ξ (f )| ≤ |f (x + x̂)| + δ
≤ kx̂ + xk + δ.
Since |ξ (f )| > 1 − 2δ , we have

2 |ξ (f )| > 2 − δ,
thus
kx + x̂k + δ > 2 − δ
and therefore
x + x̂
2 > 1 − δ.

Due to uniform convexity we now have
kx − x̂k ≤ ε.
However, kx − x̂k > ε, since J (x̂) ∈ W . This is a contradiction.
55
7 Lebesgue-Spaces Part II
Theorem 6.43. Let X be a uniformly convex Banach space, (xn )n∈N a sequence in X :
xn * x and lim sup kxn k ≤ kxk .

n→∞
Then we have
xn → x.
Proof. Let kxk =
6 0 (otherwise the statement is trivial). Take
λn := max {kxn k , kxk} ,
so
λn → kxk .
Now let
−1
yn := λ−1
n xn , y := kxk x.
We have
yn + y
kyk ≤ lim inf
.
n→∞ 2
However, kyk = 1, kyn k ≤ 1, so we have

yn + y
2 → 1.

Using uniform convexity, it follows that
ky − yn k → 0.
7 Lebesgue-Spaces Part II
7.1 The Dual of Lp
7.1.1 Case 1 < p < ∞
Theorem 7.1. Lp is reflexive for 1 < p < ∞.
Proof. 1. 2 ≤ p < ∞. Then Lp is a strictly convex space.
(This is a consequence of Carlson’s first inequality and will be proven as an exercise.)
Thus by the theorem of Milman and Pettis 6.42, Lp is reflexive for 2 ≤ p < ∞.
2. 1 < p < 2. First define for 1 < p < ∞ :
0 ˆ
p p0 p0
T :L → L , u 7→ T u : L → K, f 7→ u · f.
0
This map is well defined, since T u is a linear functional on Lp and continuous by Hölder’s
inequality 2.4.
Claim.
kT uk(Lp0 )0 = kukLp ∀u ∈ Lp .
Proof. By Hölder’s inequality, we have

0
|(T u) (f )| ≤ kukLp kf kLp0 ∀f ∈ Lp .
Therefore,
kT uk(Lp0 )0 ≤ kukLp .
56
7 Lebesgue-Spaces Part II 7.1 The Dual of Lp
For the other direction, set

(
p−2
|u (x)| u (x), u (x) 6= 0
f0 (x) := .
0, else
A short calculation shows

p−1 0
kf0 kLp0 = kukLp < ∞ ⇒ f0 ∈ Lp ,
ˆ
p−2 p
(T u) (f0 ) = |u| · |{z}
u · u = kukLp .
=|u|2
Therefore,
|(T u) (f0 )|
kT uk(Lp0 )0 ≥ = kukLp .
kf0 kLp0
We have thus shown that 0 0

T : Lp → Lp
0 0
is an isometry. Therefore, T (Lp ) is a closed subset of Lp for all 1 < p < ∞.
0
For 1 < p ≤ 2, we have p0 ≥ 2 and thus by 1. Lp is reflexive.
0 0
By Corollary 6.29, Lp is reflexive and by Lemma 6.28, T (Lp ) is reflexive.
Since
T : Lp → T (Lp )
is a surjective isometry, Lp is also reflexive.
X 00 Y 00
J ↑↓ T̃ J ↑↓ T J −1 T̃ −1
X Y
T
0
Notation. For the dual pairing, f ∈ (Lp ) , g ∈ Lp , we often write
f (g) = hf, gi .
This is very common on Lp or Sobolev spaces and somewhat common in general.

0
Theorem 7.2 (Riesz representation theorem). Let 1 < p < ∞ and φ ∈ (Lp ) . Then there exists a
0
unique function u ∈ Lp : ˆ
hφ, f i = φ (f ) = u · f ∀f ∈ Lp .
Furthermore,
kukLp0 = kφk(Lp )0 .
0
Remark. The elements of the abstract space (Lp ) can be uniquely identified with a concrete
0
function in Lp . We systematically make this identification
0 0
(Lp ) = Lp .
0 0
Proof. We consider the operator T : Lp → (Lp ) defined by
ˆ
0
hT u, f i = u · f ∀u ∈ Lp , f ∈ Lp .
The argument in the proof of Theorem 7.1 shows that
kT uk(Lp )0 = kukLp0 .
57
It remains to show that T is surjective. Take

0
E = T Lp .
0 00
Since E is a closed subspace of (Lp ) , we only have to show that E is dense. Consider h ∈ (Lp )
such that
hh, T ui = 0 ∀u ∈ Lp (∗)
⇔ hh, vi = 0 ∀v ∈ E.
Claim. h ≡ 0.
Proof. Since Lp is reflexive, we can simply take h ∈ Lp and then (∗) implies that
ˆ
0
u · h = 0 ∀u ∈ Lp .
By the choice
p−2
u = |h| h,
we see h = 0 and the claim is proven.
By Corollary 4.15, for a subspace E :
(hh, vi = 0 ∀v ∈ E ⇒ h = 0) ⇒ E is dense.
So the Theorem is proven.
7.1.2 The space L1 (µ)

For σ-additive µ.
0
Theorem 7.3 (Riesz representation theorem). Take φ ∈ L1 . Then there exists a unique function
u ∈ L∞ : ˆ
φ (f ) = u·f ∀f ∈ L1 .
Moreover,
kukL∞ = kφk(L1 )0 .
0
Remark. This allows us to identify L1 with L∞ .

Proof. Let (Ωn )n∈N be a sequence of measurable sets in Ω such that:

∞
[
Ω= Ωn and µ (Ωn ) < ∞ ∀n.
n=1
Set χn := χΩn .
Claim (1). If such a u exists, it is unique.
Proof (1). Suppose, we have u1 , u2 ∈ L∞ :
ˆ ˆ
u1 · f = u2 · f ∀f ∈ L1 .
2
Then ˆ
(u1 − u2 ) f = 0 ∀f ∈ L1 .
| {z }
=:ũ
Choosing
f = χn · sgn u,
we see ∀n : ˆ
|ũ| = 0 ⇒ ũ = 0 on Ωn .
Ωn
⇒ ũ = 0 on Ω.
58
Now for existence: Consider a function θ ∈ L2 (Ω) :
θ (x) ≥ εn > 0 ∀x ∈ Ωn .
(It is clear, that such a function exists, take eg.
θ := α1 on Ω1 ,
θ := αn on Ωn \Ωn−1 ∀n > 1
and pick αn such that θ remains square integrable.) The map
L2 → K, f 7→ φ (θ · f )
is a continuous linear functional on L2 , since by Hölder’s inequality, θ · f ∈ L1 :
kθ · f kL1 ≤ kθkL2 kf kL2
and φ is a continuous linear functional on L1 . By Theorem 7.2, there exists a function v ∈ L2 :

ˆ
φ (θ · f ) = v · f ∀f ∈ L2 (Ω) . (7.1)
Set
v (x)
u (x) := . (θ > 0on Ω)
θ (x)
u is measurable and
u · χn ∈ L1 .
Claim (2). u is the sought after function.
Proof (2). Note ˆ
φ (χn g) = uχn g ∀g ∈ L∞ (Ω) , n. (7.2)
This follows from picking

χn g
f=
θ
2
in (7.1). Further note f ∈ L , since it is bounded on Ωn on zero otherwise.
Claim (2a).
kukL∞ ≤ kφk(L1 )0 . (7.3)
Proof (2a). Fix any constant C > kφk(L1 )0 and set
A := {x ∈ Ω | |u (x)| > C} .
Further choose
g = χA · sgn u
in (7.2) and get ˆ
|u| ≤ kφk(L1 )0 · µ (A ∩ Ωn ) .
A∩Ωn
Therefore, we have
C · µ (A ∩ Ωn ) ≤ kφk(L1 )0 · µ (A ∩ Ωn ) .
Since C > kφk(L1 )0 , this can only hold, if
µ (A ∩ Ωn ) = 0.
By arbitrary choice of n, X
µ (A) = µ (A ∩ Ωn ) = 0.
n∈N
59
Claim (2b). We have ˆ

φ (h) = u · h ∀h ∈ L1 (Ω) . (7.4)
Proof (2b). Choose

gn := Tn h ∈ L∞
and apply (7.2). Note
L1 (Ω)
χn Tn h −−−−→ h
and take limits on both sides of (7.2) to prove Claim (2b).
Claim (2c). We have
kukL∞ = kφk(L1 )0 .
Proof (2c). By (7.4), we immediately have
|φ (h)| ≤ kukL∞ · khkL1 .
Therefore,
kφk(L1 )0 ≤ kukL∞ .
Together with (7.3), we get Claim (2c).
Those claims prove Claim (2).
The theorem follows from the claims.

Remark. 1. L1 (Rn ) is not reflexive. Let εn → 0 and take
1
fn = χBεn (0) · .
µ (Bεn (0))
We have
kfn kL1 = 1.
If L1 (Rn ) were reflexive, we could extract a weakly converging subsequence with limit f , i.e.
ˆ ˆ
fn · u → f · u ∀u ∈ L∞ .
Since ˆ
fn · u → 0 ∀u ∈ L∞ , 0 ∈
/ supp (u) ,
in particular for
u = χRn \Bεk (0) ∀k ∈ N,
we have,
f = 0 a.e.
However ˆ ˆ
fn · 1 = 1 → 1 6= 0 · 1.
2. Indeed, L1 (µ) is never reflexive, unless µ consists of finitely many atoms (and L1 (µ) is thus
finite dimensional).
60
7.1.3 Study of L∞
0
We already know by 7.3, that L∞ = L1 is the dual of a separable space. Therefore, we have
1. BL∞ is weakly* compact. (Banach-Alaoglu)
2. If Ω ⊂ Rn measurable, (fn )n∈N bounded in L∞ , by Corollary 6.38 and 2.9, there exists a
subsequence (fnk )k∈N and f ∈ L∞ :
∗
fnk −−−−* f.
k→∞
3. However, L∞ is not reflexive (otherwise, L1 would be reflexive by Corollary 6.29) as long as

µ does not consist of finitely many atoms.
Hence, the dual space of L∞ is not equal to L1 , i.e. there are continuous linear functionals
on L∞ that can not be expressed as an integral with an L1 function.
Example. Let
φ0 : Cc (Rn ) → R, φ0 (f ) := f (0) .
φ0 is a continuous linear functional on Cc (R ). Cc (Rn ) is a linear subspace of L∞ (Rn ). By
n
Hahn-Banach 4.1, we can extend it to L∞ and have

φ (f ) = f (0) ∀f ∈ Cc (Rn ) .
Claim. There is no u ∈ L1 (Rn ) :
ˆ
φ (f ) = u·f ∀f ∈ L∞ .
Proof. By contradiction: Note

ˆ
uf = 0 ∀f ∈ Cc (Rn ) , f (0) = 0.
This implies
u = 0 a.e. on Rn .
Thus
φ (f ) = 0 ∀f ∈ L∞ (Rn ) ,
which is a contradiction.
Example. Take Ω = (0, 1) , µ = Lebesgue measure. Then L∞ (µ) is not separable.

Proof. Take
ut := χ(0,t) , t ∈ (0, 1) .
We have
t 6= t0 ⇒ kut − ut0 kL∞ ≥ 1.
Assume now that there were a countable dense set (vj )j∈N in L∞ . Let
Ot := B1/2 (ut ) , t ∈ (0, 1) .
⇒ ∀t ∈ (0, 1) : ∃vj ∈ Ot . This produces a map
t 7→ j (t) .
0
This map is injective: Assume j (t) = j (t ). ⇒
vj(t) = vj(t0 ) ∈ Ot ∩ Ot0 .
0
However, Ot ∩ Ot0 = ∅, unless t = t . So (0, 1) has to be countable, which is a contradiction.
Conclusion. For Ω ⊂ Rn open, µ Lebesgue measure, 1 < p < ∞, p1 + 1
p0 = 1:
Reflexivity Separability Dual
0
Lp X X Lp
.
L1 × X L∞
L∞ × × ) L1
61
8 Hilbert spaces 7.2 Weak convergence in Lp (µ)
7.2 Weak convergence in Lp (µ)

Corollary 7.4. 1. Let (fk )k∈N be a sequence in Lp , 1 ≤ p < ∞. Then
ˆ
Lp 0
fk −−* f ⇔ (fk − f ) · g → 0 ∀g ∈ Lp .
Ω
∞
2. Let (fk )k∈N be a sequence in L . Then
ˆ
∗ in L∞
fk −−−−−* 0 ⇔ (fk − f ) g → 0 ∀g ∈ L1 .
Ω
Proof. The theorem follows immediately from Riesz representation 7.2/7.3 and characterisation of
weak* convergence in Proposition 6.19.
Corollary 7.5. Let (fk )k∈N be a sequence in Lp (µ) :
sup kfk kLp < ∞.
k∈N

1. 1 < p < ∞. There exists a subsequence fkj j∈N
:
Lp
fkj −−−* f.
j→∞

2. p = ∞. There exists a subsequence fkj j∈N
:
L∞ ∗
fkj −−−* f.
j→∞
Proof. Follows from Corollaries to metrisability of weak(*) convergence.

Example. A typical weakly convergent sequence looks like this:
(
˜ −1, x ∈ (0, 1/2)
f (x) := periodically extended to R.
1, x ∈ (1/2, 1)
We have  p
−L ((0,1))
˜ −−−−−* 0, 1≤p<∞
fk := f (k·) L∞ ((0,1)) ∗ .
−−−−−−−* 0
Idea of the proof:

´1
1. Show 0 fk · ϕ → 0 ∀ϕ ∈ Cc (0, 1).
2. Use density of Cc (0, 1) in Lp (0, 1) (be careful with L∞ ).
8 Hilbert spaces
In the following, let H denote a Hilbert space. Thus, we have
1. Cauchy-Schwarz inequality:
|(x, y)H | ≤ kxkH kykH ∀x, y ∈ H.
2. Parallelogram identity
1 2 1 2 2 2
kx + ykH + kx − ykH = kxkH + kykH ∀x, y ∈ H.
2 2
Example. • L2 (Ω) with ˆ
(x, y)L2 = xy ∗ .
Ω
• H 1 (Ω) with ˆ ˆ
∗
(x, y)H 1 = xy + DxDy ∗ .
Ω Ω
62
8 Hilbert spaces 8.1 Projection onto convex sets
8.1 Projection onto convex sets

Proposition 8.1. H is uniformly convex, thus reflexive.
Proof. Follows from the Parallelogram identity: Take ε > 0, x, y ∈ H : kxk < 1, kyk < 1, kx − yk >
ε⇒
1 2 ε2
kx + yk < 1 − .
4 4
Thus
2 1/2

x + y
< 1 − δ, δ := 1 − 1 − ε > 0.
2 4
Theorem 8.2 (Projection). Take K ⊂ H closed, not empty, convex. Further take f ∈ H. Then
there exists a unique u ∈ K :
ku − f k = min kv − f k
v∈K
and u is characterised as the only u ∈ K :
Re (f − u, v − u) ≤ 0 ∀v ∈ K.
Notation. We write
u = PK f.
Proof. 1. Existence:
ϕ : K → R, v 7→ kf − vk
is strongly continuous, convex and
ϕ (v) −−−−−→ ∞.
kvk→∞
By Theorem 6.32, there exists a minimiser u of ϕ.

2. Characterisation: Exercise.
3. Uniqueness: Exercise as well.
Proposition 8.3. For K ⊂ H not empty, closed, convex, PK : H → H is continuous, in particular

kPk f1 − Pk f2 k ≤ kf1 − f2 k ∀f1 , f2 ∈ H.
Proof. Take uj := PK fj , j ∈ [2] . By Theorem 8.2, we have
Re (f1 − u1 , v − u1 ) ≤ 0 ∀v ∈ K,
Re (f2 − u2 , v − u2 ) ≤ 0 ∀v ∈ K.
Take v = u2 in the first and v = u1 in the second inequality.
Re (f1 − u1 , u2 − u1 ) ≤ 0
Re (f2 − u2 , u1 − u2 ) ≤ 0
⇒ Re (f1 − f2 + u2 − u1 , u2 − u1 ) ≤ 0
2
Re (f1 − f2 , u2 − u1 ) + ku2 − u1 k ≤ 0
2
ku2 − u1 k ≤ − Re (f1 − f2 , u2 − u1 )
≤ kf2 − f1 k · ku2 − u1 k (CS inequality)
Corollary 8.4. Let M ⊂ H be a closed subspace. Then
u = PM f
is characterised by u ∈ M :
(f − u, v) = 0 ∀v ∈ M.
Proof. Exercise.
63
8 Hilbert spaces 8.2 Dual spaces of Hilbert spaces and the theorem of Lax and Milgram
8.2 Dual spaces of Hilbert spaces and the theorem of Lax and Milgram
Theorem 8.5 (Riesz-Fréchet). Take ϕ ∈ H 0 . Then there exists a unique f ∈ H :
ϕ (u) = (f, u)H ∀u ∈ H.
Proof. Take
T : H → H 0, T f : H → R, (T f ) (u) := (f, u)H .
We certainly have
kT f kH 0 = kf kH . (CS inequality and u = f as a test function)
Thus T is a linear isometry and T (H) is a closed subspace of H 0 .
It remains to show density of T (H) in H 0 . Let h̃ ∈ H 00 :
h̃ (ϕ) = 0 ∀ϕ ∈ T (H) .
By reflexivity, we can represent h̃ by h ∈ H :
ϕ (h) = h̃ (ϕ) = 0 ∀ϕ ∈ T (H) .
Thus
(T f ) (h) = 0 ∀f ∈ H
(f, h)H = 0 ∀f ∈ H
h = 0.
It follows immediately that h̃ = 0 and thus T (H) is dense in H 0 . By closedness,

T (H) = H 0
and T is a linear, isometric bijection.
Definition 8.6. A sesquilinear map
a:H ×H →K
is called
• continuous, if ∃C > 0 :
|a (u, v)| ≤ C kuk kvk ∀u, v ∈ H.
• coercive, if ∃α > 0 :
2
|a (v, v)| ≥ α kvk ∀v ∈ H.
Theorem 8.7 (Lax-Milgram, real version). Let H be a real Hilbert space, a : H × H → R a

continuous, coercive, bilinear form and f ∈ H 0 . Then there exists a unique u ∈ H :
a (u, v) = f (v) ∀v ∈ H.
Furthermore, if a is symmetric, u is characterised by u ∈ H :

1 1
a (u, u) − f (u) = min a (v, v) − f (v) .
2 v∈H 2
Proof. Fix v ∈ H. The map
u 7→ a (u, v)
is a continuous linear functional an H. By Theorem 8.5, there exists a unique w ∈ H :
a (u, v) = (w, v)H .
Doing that, we created a map A : H → H given by
A : u 7→ Au := w.
64
8 Hilbert spaces 8.2 Dual spaces of Hilbert spaces and the theorem of Lax and Milgram
Claim (1). A is linear and continuous.
1. Linearity follows immediately by bilinearity of a (·, ·). We also have

2
kAukH = (Au, Au)H = a (u, Au) ≤ C kukH kAukH . (acontinuous)
Dividing by kAukH yields continuity.

Claim (2). A is injective and A (H) is a closed subspace of H.
2.
2
α kukH ≤ a (u, u) = (u, Au)H ≤ kukH kAukH
⇒ α kukH ≤ kAukH
and injectivity follows. Closedness of A (H) follows by the sequence criterion.
Claim (3). A (H) = H.

Proof. Due to closedness, it suffices to show density. Consider thus h̃ ∈ H 0 :
h̃ (v) = 0 ∀v ∈ A (H) .
By Riesz-Fréchet 8.5, equivalently take h ∈ H :
(v, h)H = 0 ∀v ∈ A (H) .
We have
2
0 = (Ah, h)H = a (h, h) ≥ α khkH .
⇒ khkH = 0 and the claim follows.
A second application of Riesz-Fréchet 8.5 yields the existence of a unique w ∈ H :
f (v) = (v, w)H ∀v ∈ H.
Now pick u ∈ H :
Au = w.
With this choice, we have
a (u, v) = (Au, v)H = (w, v)H = f (v) ∀v ∈ H.
Existence is thus proven.

Uniqueness follows from coercivity:
a (u1 , v) = f (v) = a (u2 , v) ∀v ∈ H

⇒ a (u1 − u2 , v) = 0 ∀v ∈ H
⇒ a (u1 − u2 , u1 − u2 ) = 0
⇒ ku1 − u2 kH = 0.
Let now a be symmetrical. Then (H, a (·, ·)) is also a Hilbert space and the norm from the a-scalar
product is equivalent to the original norm. Using Riesz-Fréchet 8.5, we can thus represent the
continuous linear form f by an element g ∈ H :
a (g, v) = f (v) ∀v ∈ H.
By a (u, v) = f (v), it follows that

a (u − g, v) = 0 ∀v.
65
8 Hilbert spaces 8.3 Orthonormal basis in Hilbert spaces
This however, is nothing but the projection of g onto the whole space H w.r.t. the new a-scalar
product. Therefore, u solves the minimisation problem
p
min a (g − w, g − w),
w∈H
which is equivalent to minimising
w 7→ a (g − w, g − w) = a (w, w) − 2a (g, w) + a (g, g),

| {z }
constant
and therefore equivalent to minimzing
w 7→ a (w, w) − 2 (g, w) .
Remark. We see, that in the symmetic case, the proof reduces to a single application of Riesz-
Fréchet 8.5.
Theorem 8.8 (Complex version of Lax-Milgram). Assume A ∈ L (H) , α > 0 satisfy
2
|(Au, u)| ≥ α kuk ∀u ∈ H.
Then A is bijective.
Proof. Follows from exactly as the proof of the claims in the real version.
8.3 Orthonormal basis in Hilbert spaces

Definition 8.9. Let (En )n∈N be a sequence of closed subspaces of H. We say that H is the Hilbert
sum of the En and write M
H= En ,
n
if we have
1. The En are pairwise orthogonal, i.e.
(u, v)H = 0 ∀u ∈ En , v ∈ Em , n 6= m.
2. span {En }n∈N is dense in H.

Lemma 8.10. Let (vn )n∈N be a sequence in H :
(vn , vm ) = 0 ∀n 6= m
and assume that

∞
X 2
kvk kH < ∞.
k=1
Define
n
X
Sn := vn .
k=1
Then
S := lim Sn
n→∞
exists and
∞
X
2 2
kSkH = kvk kH .
k=1
66
Proof. For m > n, we have

m
X
2 2
kSm − Sn kH = kvn k + {zterms} .
mixed
|
k=n+1 =0 by orthogonality
Therefore Sm is a Cauchy sequence and converges to S ∈ H. Furthermore, for alln ∈ N :

n
X
2 2
kSn k = kvk k
k=1
and the claim follows in the limit.

L
Theorem 8.11 (Bessel-Parseval identity). Take H = n En , u ∈ H and consider
un := PEn u.
Let
n
X
Sn := un .
k=1
Then we have
∞
X 2 2
lim Sn = u, kuk k = kuk .
n→∞
k=1
Proof. Let un = PEn u. By the projection theorem 8.2, we have
(u − un , v)H = 0 ∀v ∈ En .
In particular,
2
(un , u)H = (u − un , un )H + (un , un )H = kun kH .
Thus,
n
X 2
(u, Sn ) = kuk k .
k=1
At the same time, we have

n
X 2 2
kuk kH = kSn kH (by Lemma 8.10)
k=1
and therefore,
2
(u, Sn )H = kSn kH .
Using Cauchy-Schwarz inequality on the left, we get
2
kukH kSn kH ≥ kSn kH
and thus
kSn kH ≤ kukH .
This yields
n
X 2 2 2
kuk kH = kSn kH ≤ kukH .
k=1
Using Lemma 8.10, we get

S = lim Sn .
n→∞
Claim. Let F := span En . Then S = PF u.
67
Proof. Note that we have

X
u − Sn = (u − um ) − uk ∀m ≤ n.
k≤n
k6=m
It follows
(u − Sn , v)H = 0 ∀v ∈ Em , m ≤ n.
Taking n → ∞, we get
(u − S, v)H = 0 ∀v ∈ Em , m ∈ N
⇒ (u − S, v)H = 0 ∀v ∈ span Em = F
⇒ (u − S, v)H = 0 ∀v ∈ F (continuity.)

By density of span En in H it follows that
S = PH u = u.
The Bessel-Parseval identity follows with

n
X 2 2
kuj kH = kSn kH
j=1
in the limit.
Definition 8.12. A sequence (en )n∈N in H is called Hilbert basis (or ONB) of H, if
1. ken kH = 1, (en , em )H = 0 ∀n 6= m.
2. span {en }n∈N is dense in H.
Corollary 8.13. Let (en )n∈N be a Hilbert basis of H, u ∈ H. Then
∞
X n
X
u= (u, ek )H ek := lim (u, ek )H ek
n→∞
k=1 k=1
and
∞
X
2 2
kukH = |(u, ek )H | .
k=1
2
Conversely, for (αk )k∈N ∈ l :
∞
X
αk ek =: u, (u, en )H = αn .
k=1
Proof. M
H= En , En := Ken .
n
We thus have
PEn u = (u, en ) en .
The claim follows with Theorem 8.11 and Lemma 8.10.
Remark. We don’t necessarily have absolute convergence. Find an example, such that
∞
X
|(u, ek )| = ∞.
k=1
1

E.g. H = l2 , u = k k∈N , en as usual.
68
9 Some theory of compact operators
Theorem 8.14. Every separable Hilbert space admits a Hilbert basis.

Proof. Take (vn )n∈N dense (countable) in H and
n
Fn = span {vk }k=1 .
(Fn )n is a sequence of finite dimensional, thus closed subspaces of H. We have

∞
[
Fk ⊃ Fj ∀k ≥ j and Fn = H.
k=1
We now iteratively pick
e1 ∈ F1 : ke1 kH = 1,
e2 ∈ Fj , j > 1 : dim Fj = 2, ke2 kH = 1, (e1 , e2 )H = 0,
ek ∈ Fjk , jk > jk−1 :dim Fjk = k, keK kH = 1, (ei , ek )H = 0 ∀i < k.
Remark. It follows that every separable ∞-dimensional Hilbert space is isomorphic to l2 .
9 Some theory of compact operators

In the following, X, Y are Banach spaces and BX is the unit ball in X.
9.1 Compact Operators and the adjoint Operator

Definition 9.1. A bounded linear operator T ∈ L (X, Y ) is called compact, if T BX is compact

in Y (w.r.t. strong topology).

The set of compact operators from X to Y is called K (X, Y ) . We write K (X) := K (X, X).
Theorem 9.2. K (X, Y ) is a closed subspace of L (X, Y ) (w.r.t. operator norm
kAkL (X,Y ) = sup kAxkY ).

x∈X
kxkX ≤1
Proof. Linearity is obvious. To show closedness, consider (Tn )n∈N , T ∈ L (X, Y ) :
Tn ∈ K (X, Y ) ∀n, and kTn − T kL (X,Y ) → 0.
Fix n ∈ N :
ε
kTn − T kL (X,Y ) < .
2
BX by 2ε -balls:

By compactness of Tn , there exists a finite cover of Tn
N
[
Tn BX ⊂ Bε/2 (yj ) , yj ∈ Y.
j=1
It follows
N
[
T BX ⊂ Bε (yi ) , yi ∈ Y.
j=1

Thus T BX is precompact and T BX is compact.
Definition 9.3. The range of an operator T ∈ L (X, Y ) is
R (T ) := {y ∈ Y | ∃x ∈ X : T (x) = y} .
The null space of T is

N (T ) := {x ∈ X | T (x) = 0} .
69
9 Some theory of compact operators 9.1 Compact Operators and the adjoint Operator
Definition 9.4. An operator T ∈ L (X, Y ) has finite range, if
dim R (T ) < ∞.
Remark. Every finite range operator is compact.

Corollary 9.5. The limit of a sequence of finite range operators is compact.
Proof. The space of compact operators is closed.
Proposition 9.6. Let X, Y, Z be Banach spaces, T ∈ K (Y, Z) , S ∈ L (X, Y ). Then
T ◦ S ∈ K (X, Z) .
Take U ∈ L (Y, Z) , V ∈ K (X, Y ), Then
U ◦ V ∈ K (X, Z) .
Proof. Exercise.
Definition 9.7. Let T ∈ L (X, Y ). The adjoint operator T ∗ ∈ L (Y 0 , X 0 ) is defined by
(T ∗ f ) (x) := f (T x) ∀x ∈ X, f ∈ Y 0 .
Remark. Take v ∈ Y 0 . We define ∀u ∈ X :
g (u) := v (T u) ⇒ |g (u)| ≤ C kukX .
Write T ∗ v = g. It follows, that
kT ∗ kL (Y 0 ,X 0 ) = kT kL (X,Y ) .
Remark. Hilbertspaces.
Theorem 9.8 (Schauder).
T ∈ K (X, Y ) ⇔ T ∗ ∈ K (Y 0 , X 0 ) .
Proof. “⇒” Consider (vn )n∈N a sequence in BY 0 . We show that (T ∗ vn )n∈N admits a con-
verging subsequence. Take

S := T BX .
S is a compact metric space. Define
H := {ϕn : S → K, x 7→ vn (x)}n∈N ⊂ C (S) .
These functions are not only continuous, but uniformly equicontinuous.

By Arzelà-Ascoli, it follows ∃ϕ : S → K, (ϕnk )k∈N :
unif. C (S)
ϕnk −−−−−−−→ ϕ.
Thus,
sup |ϕnk (x) − ϕ (x)| −−−−→ 0

x∈S k→∞
⇒ sup |vnk (T u) − ϕ (T u)| −−−−→ 0

u∈BX k→∞
⇒ sup |vnk (T u) − vnl (T u)| −−−−→ 0

u∈BX k,l→∞
∗ ∗
⇒ kT vnk − T vnl kL (Y 0 ,X 0 ) −−−−→ 0
k,l→∞
∗ 0
⇒ T vnk converges in X .
70
J(T BX ) = T ∗∗ BX is precompact in Y 00 and J(Y ) is closed in Y 00 .

“⇐”
X
Remark. It follows that xn −* x, T ∈ K (X, Y ) ⇒
Y
T xn −→ T x.
Definition 9.9 (Anihilator). Let X be a Banach space, M ⊂ X a linear subspace. the anihilator
of M is
M ⊥ := {f ∈ X 0 | f (x) = 0 ∀x ∈ M } .
Let N ⊂ X 0 be a linear subspace. The anihilator of N is
N ⊥ := {x ∈ X | f (x) = 0 ∀f ∈ N } .
Remark. • This simplifies in Hilbert spaces.

• It is clear that M ⊥ , N ⊥ are both closed subspaces.
• By definition N ⊥ ⊂ X and not N ⊥ ⊂ X 00 !
Proposition 9.10. We have
⊥
1. M ⊥ = M ,
⊥
2. N ⊥ ⊃ N.
⊥
Proof. 1. From the definition, it is clear that M ⊥ ⊃ M , since
⊥
x∈M ⇒ f (x) = 0 ∀f ∈ M ⊥ ⇒ x ∈ M⊥ .
⊥
Since M ⊥ is closed, we also have
⊥
M⊥ ⊃ M.
⊥
Assume that ∃x0 ∈ M ⊥ \M .
By the second separation theorem 4.13, there exists f ∈ X 0 , α ∈ R :
Re f (x) < α < Re f (x0 ) ∀x ∈ M .
By−M = M , it follows
Re f (x) = 0 ∀x ∈ M .
Re f (x0 ) > 0.
Thus, by iM ⊂ M ,
f (x) = 0 ∀x ∈ M.
⇒ f ∈ M⊥
⇒ f (x0 ) = 0 .
2. The inclusion follows as in 1.

⊥
Remark. N = N ⊥ only holds in reflexive spaces.
Proposition 9.11. Consider A ∈ L (X). We have
⊥
1. N (A∗ ) = R (A) ,
⊥
2. N (A) = R (A∗ ) .
71
Proof. 1. Take f ∈ N (A∗ ), i.e.
A∗ f (x) = 0 ∀x ∈ X
⇔ f (Ax) = 0 ∀x ∈ X
⇔ f (y) = 0 ∀y ∈ R (A)
⊥
⇔ f ∈ R (A) .
2. Analogous.
Corollary 9.12.
N (A∗ ) = R (A).
Proof. Follows immediately from 9.11 and 9.10.
We state the following theorem only in Hilbert spaces, for a full version in Banach spaces (which
is slightly more technical to prove), see the book by Brezis.
Theorem 9.13 (Fredholm alternative). Let H be a Hilbert space, K ∈ K (H, H) a compact
operator. Then
1. N (id −K) is finite dimensional,
2. R(id −K) is closed,

3. R(id −K) = N (id −K ∗ )⊥ ,
4. N (id −K) = {0} if and only if R(id −K) = H,
5. dim N (id −K) = dim N (id −K ∗ ).

Remark. The theorem concerns the solvability of the equation u − T u = f :
• Either the equation admits a unique solution for all f ∈ X

• or the homogeneous equation u − T u = 0 has n linearly independent solutions.
• u − T u = f is then solvable exactly if f admits n orthogonal solutions, i.e.
⊥
f ∈ N (id −T ∗ ) .
Proof. See Evans, Partial Differential Equations, Appendix D.5.

For information on applications and Sobolev spaces I recommend Chapter 5 in the book by
Evans.
72

Functional Analysis: Prof. Dr. Patrick Dondl Adapted From Lecture Notes by Josias Reppekus (TU-München) 23rd July 2016

Uploaded by

Functional Analysis: Prof. Dr. Patrick Dondl Adapted From Lecture Notes by Josias Reppekus (TU-München) 23rd July 2016

Uploaded by

Functional Analysis

Prof. Dr. Patrick Dondl

3 Continuous linear maps 21

4 Hahn-Banach theorem and some consequences 23

5 Baire category argument 31

6 The weak topology 36

9 Some theory of compact operators 69

• Typical question: X, Y vector spaces; A : X → Y linear continuous. Does A have a continu-

Example. U ⊂ Rn open, with smooth boundary,

A invertible ⇔ The equation

Topics of the class

∃U ∈ T such that A = X\U =: U c .

Let X henceforth be a topological space.

Proposition 1.6. Let A ⊂ X, (X, T ) be a topological space. Then (A, TA ) with

is a topological space. TA is called a subspace topology (relative topology).

2. A mapping f : X → Y is called continuous in x ∈ X, if for any neighborhood NY of f (x)

x = lim xk w.r.t. (with respect to) T .

Proof. Assume xk → x1 , xk → x2 6= x1 . Then

2. TX = {∅, X}, TY Hausdorff. Then we have: f is continuous only if f is constant.

f (x) = inf f (y) for some f : X → [0, ∞).

1. ∃ (xk )k∈N : f (xk ) → inf f. We have f (xk ) < C for k > k0 .

Remark 1.15. i) One topic we left

1. d (x, y) ≥ 0 and d (x, y) = 0 ⇔ x = y.

Then we call (X, d) a metric space, d a metric or a distance.

3. (X, d) metric space, h : Y → X injective,

dY (y1 , y2 ) := d (h (y1 ) , h (y2 ))

is called the pullback metric.

Definition 1.17. Let (X, d) be a metric space,

Br (x) := {y ∈ X | d (x, y) < r} for r > 0, x ∈ X.

We call U ⊂ X open with respect to the metric d, if ∀x ∈ U there exists r > 0

The empty set is open.

Then (X, T ) is a Hausdorff space. We call T the by d induced topology.

Br1 (x) ⊂ U1 and Br2 (x) ⊂ U2 .

It follows by the triangle inequality that

Bmin(r1 ,r2 ) (x) ⊂ U1 ∩ U2 .

4. Let x 6= y. We have that d (x, y) = c > 0.

∀ε > 0 : ∃δ > 0 : dX (x, y) < δ ⇒ dY (f (x) , f (y)) < ε.

Remark. The second condition is called sequential continuity.

dX (x, xδ ) < δ, but dY (f (x) , f (xδ )) > ε.

dY (f (x) , f (xk )) ≥ ε, xk → x, but f (xk ) 6→ f (x) .

Definition 1.22. Let (X, d) be a metric space.

Pn := {f : I → R | f is a ploynomial of degree deg (f ) ≤ n} ,

(P, d) is a metric space. We take

(And also supx∈I |fn (x) − fk (x)| −−−−−→ 0.) However, if

would exist in P , then necessarily, we would need

|fn (x) − g (x)| ≤ d (fn , g) → 0 ∀x ∈ I,

(xj )j∈N = (yj )j∈N :⇔ d (xj , yj ) → 0.

Furthermore, the mapping

J : X → X̃, J (x) := (x)j∈N (constant sequence)

defines an (therefore injective) isometry, that is

d˜(J (x) , J (y)) = d (x, y) ∀x, y ∈ X,

and for any (xj )j∈N ∈ X̃, we have

thus J (X) is dense in X̃.

|d (xj , yj ) − d (xi , yi )| ≤ |d (xj , yj ) − d (xi , yj )| + |d (xi , yj ) − d (xi , yi )|

So this was a Cauchy sequence in R and the limit

d˜(x̃, ỹ) = lim d (xj , yj )

exists. By analogous arguments,

d x2i , yi2 − d x1i , yi1 → 0,

if x̃1 = x̃2 ∈ X̃ and ỹ 1 = ỹ 2 ∈ X̃.

2. Symmetry carries over in the limit.

d xkjk , xljl ≤ d xkjk , xkj + d xkj , xlj + d xlj , xljl

Proof. d˜ xl , x∞ ←−−−− d xlk , x∞

d xlk , x∞ ≤ d xlk , xljl + d xljl , xkjk

• The statements about J are easy to verify.

1.3 Vector spaces and norms

Definition 1.27 (Norm).

• If k·k is a norm on X, then (X, k·k) is called a normed space.

• A mapping with properties 1., 3., 4. is called semi-norm.

Proposition 1.28. Let (X, k·k) be a normed space. By taking

we get a metric space (X, d) .

2. Two norms are equivalent (:⇔ T1 = T2 ), iff ∃C, c > 0 :

c kxk1 ≤ kxk2 ≤ C kxk1 .

1.4 Scalar product

is called a sesquilinear form, if ∀x, x1 , x2 , y, y1 , y2 ∈ X, α ∈ R we have

A sesquilinear form is symmetrical, if

(x, y) = (y, x).

2. Let K ⊂ V ⊂⊂ Ω be compact and > 0. By uniform continuity of f on V , there exists δ > 0