Functional Analysis Notes

Fall 2004
Prof. Sylvia Serfaty

Yevgeny Vilensky
Courant Institute of Mathematical Sciences
New York University

March 14, 2006

ii
Preface

These are notes from a one-semester graduate course in Functional Analysis

given by Prof. Sylvia Serfaty at the Courant Institute of Mathematical Sci-
ences, New York University, in the Fall of 2004. Thanks to Atilla Yilmaz,
Caroline Müller, and Alexey Kuptsov for providing their course notes to help
with the preparation of this typed version. The course was largely based on
Haim Brezis’ Analyse fonctionnelle : thorie et applications, and Michael Reed’s
and Barry Simon’s Methods of modern mathematical physics vol. 1. The ge-
ometric versions of the Hahn-Banach Theorems were taken almost entirely out
of Brezis and the section on Spectral Theory was based entirely on Reed and
Simon.

These notes may be used for educational, non-commercial purposes. You

can reproduce as many copies as you want, but you may not sell them (but you
can give them away for free!).

2004,
c Yevgeny Vilensky, New York, NY

iii
iv
Contents

Preface iii

1 Hahn-Banach Theorems and Introduction to Convex Conjuga-

tion 1
1.1 Hahn-Banach Theorem - Analytic Form . . . . . . . . . . . . . . 1
1.1.1 Theorems on Extension of Linear Functionals . . . . . . . 1
1.1.2 Applications of the Hahn-Banach Theorem . . . . . . . . 3
1.2 Hahn-Banach Theorems - Geometric Versions . . . . . . . . . . . 5
1.2.1 Definitions and Preliminaries . . . . . . . . . . . . . . . . 5
1.2.2 Separation of a Point and a Convex Set . . . . . . . . . . 6
1.2.3 Applications (Krein-Milman Theorem) . . . . . . . . . . . 8
1.3 Introduction to the Theory of Convex Conjugate Functions . . . 9

2 Baire Category Theorem and Its Applications 13

2.1 Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.1.1 Reminders on Banach Spaces . . . . . . . . . . . . . . . . 13
2.1.2 Bounded Linear Transformations . . . . . . . . . . . . . . 13
2.1.3 Duals and Double Duals . . . . . . . . . . . . . . . . . . . 15
2.2 The Baire Category Theorem . . . . . . . . . . . . . . . . . . . . 16
2.3 The Uniform Boundedness Principle . . . . . . . . . . . . . . . . 17
2.4 The Open Mapping Theorem and Closed Graph Theorem . . . . 18

3 Weak Topology 21
3.1 General Topology . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 Frechet Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3 Weak Topology in Banach Spaces . . . . . . . . . . . . . . . . . . 24
3.4 Weak-* Topologies σ(X ∗ , X) . . . . . . . . . . . . . . . . . . . . 28
3.5 Reflexive Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.6 Separable Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.7 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.7.1 Lp Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.7.2 PDE’s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

v
vi

4 Bounded (Linear) Operators and Spectral Theory 37

4.1 Topologies on Bounded Operators . . . . . . . . . . . . . . . . . 37
4.2 Adjoint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.3 Spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.4 Positive Operators and Polar Decomposition (In a Hilbert Space) 46

5 Compact and Fredholm Operators 47

5.1 Definitions and Basic Properties . . . . . . . . . . . . . . . . . . 47
5.2 Riesz-Fredholm Theory . . . . . . . . . . . . . . . . . . . . . . . 49
5.3 Fredholm Operators . . . . . . . . . . . . . . . . . . . . . . . . . 51
5.4 Spectrum of Compact Operators . . . . . . . . . . . . . . . . . . 52
5.5 Spectral Decomposition of Compact, Self-Adjoint Operators in
Hilbert Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

A 57
Chapter 1

Hahn-Banach Theorems
and Introduction to Convex
Conjugation

1.1 Hahn-Banach Theorem - Analytic Form

1.1.1 Theorems on Extension of Linear Functionals
The Hahn-Banach Theorem concerns extensions of linear functionals from a
subspace of a linear space to the entire space.

Theorem 1.1.1 (Real Version of the Hahn-Banach Theorem) Let X be

a real linear space and let p : X → R be a function satisfying:

p(tx) = t · p(x), p(x + y) ≤ p(x) + p(y)

for all t > 0, x, y ∈ X. Let f : Y → R be linear with Y ⊂ X such that

f (x) ≤ p(x) for all x ∈ X. Then, ∃ a linear map Λ : X → R such that for
y ∈ Y, Λ(y) = f (y) and Λ(x) ≤ p(x) for all x ∈ X.

Before beginning the proof of the theorem, we need some definitions and a
reminder of Zorn’s Lemma.

Definition Let P be a set with a partial order relation ”≺”. Q ⊂ P is said to

be totally ordered if ∀a, b ∈ Q we have a ≺ b or b ≺ a. c is an upper bound for
Q if a ∈ Q ⇒ a ≺ c. m is called a maximal element in Q if and only if ∀a ∈ Q
we have that if m ≺ a then a = m.

Lemma 1.1.2 (Zorn’s Lemma) Let P be a non-empty set with a partial or-
dering, such that every totally ordered subset of P admits an upper bound. Then,
P has a maximal element.

1
2

Proof of Real Version of the Hahn-Banach Theorem Let P be the col-

lection of linear functions h, defined on their domain, D(h) ⊃ Y, that extend f
and that satisfy:

h(y) = f (y) ∀y ∈ Y
h(x) ≤ p(x) ∀x ∈ X.

We now define a partial ordering on the set P, so that we can apply Zorn’s
Lemma. In P, we say h1 ≺ h2 if and only if D(h1 ) ⊂ D(h2 ) and h1 = h2 in
D(h1 ).
Certainly, P is nonempty (because it at least contains fS). Now, let (hα )α∈A
be a totally ordered subset of P. Let h be defined on α∈A D(hα ) and let
h(x) = hα (x) if x ∈ D(hα ). This is well-defined because (hα )α∈A is a totally-
ordered set (and so, all hα agree on the intersection). By our definition of ≺, it
follows that h is an upper bound.
So, applying Zorn’s Lemma to (P, ≺), we see that P has a maximal element.
Call this element Λ. We just need to check that D(Λ) = X.
Suppose that D(Λ) 6= X. Then, let x0 ∈ / D(Λ). Then, we claim that there is
an a so that we can extend Λ to h : D(Λ) ⊕ Rx0 → R by:

h(x + tx0 ) = Λ(x) + t · a

and
Λ(x) + t · a ≤ p(x + tx0 )

for all x ∈ D(Λ) and t ∈ R.


Λ(x) + a ≤ p(x + x0 )
⇔
Λ(x) − a ≤ p(x − x0 )
For all x ∈ D(Λ) (just replace x by xt if t > 0 and − xt if t < 0). So, is there
such an a? It is enough to check that:

sup Λ(x) − p(x − x0 ) ≤ inf p(y + x0 ) − Λ(y)

x∈D(Λ) y∈D(Λ)

To show this, note that by the linearity of Λ we have

Λ(x) + Λ(y) = Λ(x + y)

= Λ(x − x0 + x0 + y) ≤ p(x − x0 + x0 + y)
≤ p(x − x0 ) + p(x0 + y)

The last inequality being true, by the subadditivity of p.

⇒ Λ(x) − p(x − x0 ) ≤ p(x0 + y) − Λ(y)

for all x, y. Hence, supx LHS ≤ inf y RHS. Hence, it is possible to choose an a so
we can extend Λ to h such that h(x+tx0 ) = Λ(x)+ta and Λ(x)+ta ≤ p(x+tx0 ).
But this contradicts the fact that Λ was the maximal element.
 3

Theorem 1.1.3 (Complex Version of Hahn-Banach Theorem) Let X

be a complex linear space, p : X → R a map such that:
p(αx + βy) ≤ |α|p(x) + |β|p(y) and p(tx) = t · p(x)
for all x, y ∈ X, t > 0, and α, β ∈ C satisfying |α| + |β| = 1. Let f : Y ⊂ X → C
be linear such that |f (y)| ≤ p(y) for all y ∈ Y.
Then, there exists a linear Λ : X → C such that Λ(y) = f (y) for y ∈ Y and
|Λ(x)| ≤ p(x) for all x ∈ X.

Proof We want to reduce this to the real case. Let l(x) = <f (x). Since
f (ix) = if (x), we have that l(ix) = <f (ix) = <if (x) = −=f (x). So that,
f (x) = l(x) − il(x).
Then, since for any z ∈ C, |<z| ≤ |z|, we get that l(x) ≤ |f (x)| ≤ p(x).
So, we apply the Real Version of the Hahn-Banach Theorem to l, which is
real linear and p satisfies p((1 − α)x + αy) ≤ (1 − α)p(x) + αp(y) ∀α ∈ [0, 1].
Hence, there exists and L defined on all of X such that L(x) ≤ p(x) for all
x ∈ X and l(y) = L(y) for all y ∈ Y. So, we take Λ to be given by Λ(x) =
L(x) − iL(ix). Λ is linear and Λ(y) = L(y) − iL(y) = l(y) − il(y) = f (y) for
y ∈ Y. Furthermore, since |z| is real, for any z ∈ C, we can write |z| = eiθ z
for some θ. So, R 3 |Λ(x)| = eiθ Λ(x) = Λ(eiθ x). Thus, since Λ(eiθ x) is real,
Λ(eiθ x) = L(eiθ x) ≤ p(eiθ x) ≤ |eiθ |p(x) = p(x) (by setting β = 0 and α = eiθ
and applying the assumptions of the theorem).

1.1.2 Applications of the Hahn-Banach Theorem

Definition Let X be a normed linear space. The dual space, denoted X ∗ , is
the space of all bounded linear functions on X :
f : X → K is linear, and kf kX ∗ = sup |f (x)| < ∞
kxkX ≤1

k · kX ∗ defines a norm on X ∗ , called the dual norm. For all

x ∈ X, |f (x)| ≤ kf kX ∗ kxkX .
Lemma 1.1.4 Let f : X → R be linear. The following are equivalent:
1. f is bounded
2. f is continuous
3. f is continuous at a point
Proof It is clear that (2) ⇒ (3).
To show (1) ⇒ (2), suppose that f is bounded. Then, let kf kX ∗ = M. Fix
 > 0. So, letting δ = /M, if kx−yk < δ, then |f (x−y)| < kf kX ∗ δ = M /M = .
Finally, to show (3) ⇒ (1), assume that f is continuous at a point x0 . Then,
∀ > 0, ∃η > 0 such that
kx − x0 k < η ⇒ |f (x) − f (x0 )| <  (Hence, |f (x − x0 )| < ).
4

Hence, for any y such that kykX < η, we have that |f (y)| < . Now, for any
y η η
y ∈ X, y 6= 0 let x = kyk X 2
⇒ kxkX = η/2 ⇒ | 2kyk X
f (y)| = |f (x)| < . So, for
2
all y 6= 0, |f (y)| <  η kykX . Hence, f is bounded.

Remark Sometimes, we denote f (x) =< f, x > .

Corollary 1.1.5 Let X be a normed linear space and f be a linear function

defined on a subspace Y ⊂ X with

kf kY ∗ = sup |f (x)|.
x∈Y, kxkX ≤1

Then, f can be extended to g ∈ X ∗ such that g = f on Y and kgkX ∗ = kf kY ∗ .

Proof Apply the either the Real or Complex Version of the Hahn-Banach The-
orem (depending on the field of scalars K, for X) with p(x) = kf kY ∗ kxkX . It
is easy to check that it satisfies all of the semi-norm properties required for the
assumptions in the Hahn-Banach Theorem and that |f | ≤ p in Y . So, we can
extend f to g with |g(x)| ≤ p(x) = kf kY ∗ kxkX . Hence, kgkX ∗ ≤ kf kY ∗ . On
the other hand, if we take any y ∈ Y ⊂ X, satisfying kykY ≤ 1, we see that
kgkX ∗ ≥ |g(y)| = |f (y)| (from the H-B Theorem). Hence, kgkX ∗ ≥ kf kY ∗ .

Corollary 1.1.6 For all x0 ∈ X, there exists f0 ∈ X ∗ such that f0 (x0 ) = kx0 k2X
and kf0 kX ∗ = kx0 kX .

Proof Take Y = Kx0 , where K is the base field. Define g : Y → K by:

g(tx0 ) = t · kx0 k2X .

So, kgkY ∗ = supktx0 kX ≤1 |g(tx0 )| = supktx0 kX ≤1 |t|kx0 k2X = kx0 kX , the last
equality being true by considering the case of t = kx01kX . So, we can extend g
to f0 ∈ X ∗ such that kf0 kX ∗ = kx0 kX by applying the preceding corollary.

Corollary 1.1.7 For all x ∈ X,

kxkX = sup | < f, x > |

kf kX ∗ ≤1

= max | < f, x > |

kf kX ∗ ≤1

f0
Proof Fix x0 6= 0 and consider g = kx0 k with f0 as in the previous result.
Then,  
 f0 (x0 ) 
sup | < f, x > | ≥ 
  = kx0 kX ,
kf kX ∗ ≤1 kx0 kX 
since f0 (x0 ) = kx0 k2X and kgkX ∗ = 1.
But, | < f, x > | ≤ kf kX ∗ kxkX . Hence, kxkX ≥ supkf kX ∗ ≤1 | < f, x > |. So,
the first equality is proved. For the second one, we note that the sup is achieved
for g = f0 /kx0 kX . Since f0 exists by the previous corollary, the sup becomes a
max .
 5

Remark In light of this result, compare:

kf kX ∗ = sup | < f, x > |

kxkX ≤1

kxkX = sup | < f, x > |

kf kX ∗ ≤1

The first is the definition. The second is the previous result.

Corollary 1.1.8 x = 0 ⇔ ∀f ∈ X ∗ , f (x) = 0

1.2 Hahn-Banach Theorems - Geometric Ver-

sions
In this section, we will investigate a formulation of the Hahn-Banach theorem in
terms of separating convex sets by hyperplanes. For the purposes of this section
we assume that X is a normed linear space where the base field, K, is R.

1.2.1 Definitions and Preliminaries

Definition A hyperplane H is a set of solutions to the equation f (x) = α for
some α ∈ R and f is a non-zero linear function.

Proposition 1.2.1 H is closed if and only if f is bounded.

Definition Suppose A, B ⊂ X.
• The hyperplane {f = α} separates A and B if ∀x ∈ A, f (x) ≤ α, and
∀x ∈ B, f (x) ≥ α.
• The hyperplane {f = α} separates A and B strictly if ∃ > 0 such
that ∀x ∈ A, f (x) ≤ α − , and ∀x ∈ B, f (x) ≥ α + .

Definition A set A is convex if for all x, y ∈ A and for all t ∈ [0, 1],

t · x + (1 − t) · y ∈ A.

Theorem 1.2.2 (Hahn-Banach Theorem - First Geometric Form) Let

A, B ⊆ X be two non-empty disjoint convex sets, A open. Then, there exists a
closed hyperplane separating A and B

The primary tool to be used for proving such a theorem is the idea of a ”gauge”
of a convex set.

Definition Let C be an open convex subset of X, containing the origin. We

define the gauge of C to be a map p : X → R+ by:
x
p(x) = inf{t > 0 : ∈ C}.
t
6

Remark Some books refer to the gauge p as the Minkowski Functional.

Proposition 1.2.3 Let p be the gauge of C. Then p has the following properties:
1. p(tx) = t · p(x), ∀t > 0 ∀x ∈ X
2. p(x + y) ≤ p(x) + p(y), ∀x, y ∈ X
3. 0 ≤ p(x) ≤ M kxkX , ∀x ∈ X
4. p(x) < 1 ⇔ x ∈ C

Proof 1. The proof is clear.

2. From the first property, if λ > p(x) then, λx ∈ C. So, if  > 0, then,
x y
p(x)+ , p(y)+ ∈ C. So, ∀t ∈ [0, 1], since C is convex,

x y
t + (1 − t) ∈ C.
p(x) +  p(y) + 
p(x)+ x+y
So, take t = p(x)+p(y)+2 ∈ [0, 1]. Hence,p(x)+p(y)+2 ∈ C. Since p(x + y)
is defined to be the smallest t such that p(x+y)
t ∈
C, it must be that
p(x+y) ≤ p(x)+p(y)+2. Since,  > 0 was arbitrary, p(x+y) ≤ p(x)+p(y).
3. C is open. So, there is an r > 0 such that C ⊃ B(0, r). So, for all x 6= 0
in X,
x r 2
∈ C ⇒ p(x) ≤ kxkX
kxkX 2 r
x
since p(x) is the inf of all t such that t ∈ C.
4. Suppose x ∈ C. Then, (1 + )x ∈ C for some  > 0 since C is open.
So, reasoning as before regarding the minimality of p(x), we have that
1
p(x) ≤ 1+ < 1. Conversely, if p(x) < 1, then ∃α < 1 such that αx ∈ C.
x
So, α · α + (1 − α) · 0 ∈ C since C is convex. Hence, x ∈ C.

1.2.2 Separation of a Point and a Convex Set

The proof of the following lemma will allow us to prove the First Geometric
Form of the Hahn-Banach Theorem
Lemma 1.2.4 Let C ⊂ X be an open, non-empty, convex set and x0 a point
such that x0 ∈ / C. Then, there exists a bounded linear function f such that
f (x) < f (x0 ), ∀x ∈ C.

Proof Up to translation, we can assume WLOG that 0 ∈ C. Define the func-

tional g : Rx0 → R by g(tx0 ) = t. Then, we apply the Real Version of the
Hahn-Banach Theorem to g and p, the gauge of C.
To do so, we just check that g ≤ p on Rx0 . If t ≥ 0, then
p(tx0 ) = tp(x0 ) ≥ t = g(tx0 ) since p(x) ≥ 1 for x ∈
/ C, by Property 4 in
 7

Proposition 1.2.3. On the other hand, if t ≤ 0, g(tx0 ) = t ≤ 0 ≤ p(tx0 ). In

either case, g ≤ p on Rx0 .
So, from Hahn-Banach, we get a linear functional f such that f = g on
Rx0 and f ≤ p on all of X. So, for x0 , f (x0 ) = g(x0 ) = 1. But, for x ∈ C,
f (x) ≤ p(x) < 1 (by property 2 of Proposition 1.2.3). By Property 2 of the
same Proposition, p is bounded and since f is linear and bounded by p, it
belongs to X ∗ Hence, f separates x0 and C.

Proof of Hahn-Banach Theorem S - First Geometric Form Apply the

preceding lemma to C = A − B = y∈B A − y. One can check by hand that C
is convex. Also, since A is open, C is open as well (being the union of open sets).
/ C, for else, A ∩ B 6= ∅. By the preceding lemma, ∃f ∈ X ∗ such that
Finally, 0 ∈
f (x) < 0, for all x ∈ C (since f is linear, f (0) = 0). So, for all x ∈ A y ∈ B,
f (x − y) < 0. So, by the linearity of f, f (x) < f (y) for any x ∈ A, y ∈ B.
Hence,
sup f (x) ≤ inf f (y).
x∈A y∈B

Therefore, ∃α ∈ R such that f (x) ≤ α ≤ f (y) for all x ∈ A, y ∈ B. Hence, the

hyperplane {f = α} separates A and B.

Theorem 1.2.5 (Hahn-Banach Theorem - Second Geometric Form)

Let A, B 6= ∅ be disjoint convex sets with A closed, B compact. Then, there
exists a closed hyperplane which strictly separates A and B.

Proof Consider the sets A = A + B(0, ), B = B + B(0, ). For  suffi-
ciently small, they are disjoint. Indeed, suppose ∃xn ∈ A, yn ∈ B such that
kxn − yn kX → 0. Then, since B is compact, ∃ subsequence {ynk }k∈Z such that
ynk → l. Hence, xnk → l. Hence, since A is closed, l ∈ A ∩ B, contradicting their
disjointness. So, by the First Geometric Form of the Hahn-Banach Theorem,
∃f ∈ X ∗ , f 6= 0 and α ∈ R such that ∀x ∈ A , ∀y ∈ B , f (x) ≤ α ≤ f (y). So,
∀x ∈ A, y ∈ B, and z ∈ B(0, 1), we have f (x + z) ≤ α ≤ f (y + z). Hence, by
choosing z appropriately, we can get that f (x) ≤ α−kf kX ∗ , f (y) ≥ α−kf kX ∗ ,
∀x ∈ A, y ∈ B. Hence, f separates A and B strictly.

Corollary 1.2.6 Let Y ⊆ X be a subspace such that Y 6= X. Then, ∃f ∈ X ∗

such that f 6= 0 and f (y) = 0 ∀y ∈ Y.

Remark Stated alternatively, Y ⊆ X subspace is dense ⇔ ∀f ∈ X ∗ , f = 0 on

Y implies f = 0.

Proof Assume Y 6= X. Then, ∃x0 ∈ X \ Y . Y is closed and convex. {x0 }

is convex and compact. Hence, by the Second Geometric Form of the Hahn-
Banach Theorem ∃f ∈ X ∗ , f 6= 0 such that f (x) < f (x0 ) for all x ∈ Y . For all
t ∈ R, tf (x) = f (tx) < f (x0 ). Hence, for x ∈ Y , f (x) = 0. Thus, f = 0 on Y,
but f 6≡ 0.
8

1.2.3 Applications (Krein-Milman Theorem)

Definition Let K be a subset in a normed linear space.
• S ⊆ K is an extreme set if:

tx + (1 − t)y ∈ S for some t ∈ (0, 1), x, y ∈ K =⇒ x, y ∈ S.

• A point x0 is an extreme point of K if and only if:

x0 = tx + (1 − t)y, 0 < t < 1, x, y ∈ K =⇒ x = y = x0 .

Definition
• The convex hull of a set is the smallest convex set containing it.
• The closed convex hull of a set is the closure of the convex hull.

Theorem 1.2.7 (Krein-Milman) Let K be a compact and convex set in X.

Then, K is the closed convex hull of its extreme points.

Remark If X is a normed linear space, then X ∗ separates points (i.e.: for all
x, y ∈ X such that x 6= y ∃f ∈ X ∗ such that f (x) 6= f (y)).

Proof of Krein-Milman Theorem Let P be the collection of all extreme

sets in K. We will use the following two properties:
• The intersection of elements of P is in P or empty (check!)
• If S ∈ P and f ∈ X ∗ then if we define Sf = {x ∈ S : f (x) = maxS f },
Sf ∈ P.
To show this, let tx + (1 − t)y ∈ Sf ⊆ S. Then, f (tx + (1 − t)y) = maxS f.
Since S is extreme, x, y ∈ S. Thus,

tf (x) + (1 − t)f (y) = max f (1.1)

S

If f (x) < maxS f or f (y) < maxS f, (i.e.: x or y ∈

/ Sf ) we would have:

tf (x) + (1 − t)f (y) < max f,

S

contradicting Eq. (1.1). Hence, f (x) = f (y) = maxS f . This means that,
x, y ∈ Sf . Hence, Sf is extreme.
0
Now, let S ∈ P. Let P be the collection of all extreme sets in S. By the Hausdorff
0
Maximality Theorem ∃ a maximal, totally ordered subcollection called Ω ⊂ P .
0
Let M = ∩T ∈Ω T. M is an extreme set, ie: M ∈ P .
Then, Mf = M by the definition of M.

⇒ ∀f ∈ X ∗ , ∀x ∈ M, f (x) = max f (x) = const

x∈M
 9

Hence, every f is constant on M. But, since X ∗ separates points, M is a single-

ton. Consequently, ∀S ∈ P, S contains an extreme point.
Let H denote the convex hull of the set of extreme points in K. We want
H = K. For all S ∈ P, S ∩ H 6= ∅. Clearly, H ⊆ K. On the other hand, suppose
∃x0 ∈ K \ H. Then, {x0 } compact and H is closed. By the Second Geometric
Form of Hahn-Banach, ∃f ∈ X ∗ such that f (x) < f (x0 ) ∀x ∈ H. Hence, if we
consider the set Kf = {x ∈ K : f (x) = maxK f }, we see that Kf ∩ H = ∅
since ∀x ∈ H, f (x) < f (x0 ) ≤ maxK f. But, Kf is an extreme set and so must
intersect H. This is a contradiction. Hence, K ⊂ H.

1.3 Introduction to the Theory of Convex Con-

jugate Functions
Let X be a topological space. Consider functions ϕ : X → (−∞, ∞]. Let the
domain of ϕ be defined as D(ϕ) = {x ∈ X : ϕ(x) < +∞} and the epigraph of
ϕ be defined by epi(ϕ) = {(x, λ) ∈ X × R : ϕ(x) ≤ λ}.
Definition We say a function ϕ : X → (−∞, +∞] is lower semicontinuous
(or l.s.c.) if ∀λ ∈ R, the set {x : ϕ(x) ≤ λ} is closed. Equivalently, if epi(ϕ)
is closed. Also, if ∀x ∈ D(ϕ), ∀ > 0, ∃ a neighborhood V of x such that
y ∈ V ⇒ f (y) ≥ f (x) − .

Remark This allows for possible downhill discontinuities. One can similarly
define upper semicontinuous.

Proposition 1.3.1
• If ϕ is l.s.c., and xn → x, then ϕ(x) ≤ lim inf n→∞ ϕ(xn ).
• A supremum of l.s.c. functions is l.s.c. (ie: ϕ(x) = supi ϕi (x) is l.s.c. if
ϕi (x) are l.s.c.).

Definition We say f is convex if ∀x, y ∈ X, and t ∈ [0, 1],

f (tx + (1 − t)y) ≤ tf (x) + (1 − t)f (y). Or equivalently, if epi(f ) is convex.

Definition Let X be a normed linear space, and ϕ : X → (−∞, +∞]. We define

the conjugate function (or Legendre-Fenchel transform) of ϕ as
ϕ∗ : X ∗ → (−∞, +∞] by:

ϕ∗ (f ) = sup (< f, x > −ϕ(x)). (Requires that ϕ 6≡ +∞)

x∈D(ϕ)

Remark Observe that:

• ∀x ∈ D(ϕ), f 7−→< f, x > −ϕ(x) is an affine function (hence, continuous

and convex).
10

• A supremum of affine functions is l.s.c. and convex −→ ϕ∗ is convex and

l.s.c.

Proposition 1.3.2 If ϕ is convex, l.s.c. and ϕ 6≡ +∞, then ϕ∗ 6≡ +∞.

Proof Look at epi(ϕ). It is closed and convex. So, let (x0 , λ̃0 ) ∈ epi(ϕ) (such
a point exists since ϕ 6≡ +∞, so consider a point, (x0 , λ0 ) below epi(ϕ). In
other words, choose λ0 < ϕ(x0 ). {(x0 , λ0 )} is compact and convex. So, apply
the Second Geometric Form of the Hahn-Banach Theorem in X × R to this
set and to epi(ϕ). So, ∃ a linear functional Λ on X × R and an α such that
∀(x, λ) ∈ epi(ϕ), Λ(x, λ) > α > Λ(x0 , λ0 ).
Now, we can write Λ(x, λ) = f (x) + kλ for some f ∈ X ∗ and k ∈ R since Λ
is linear. So, ∀x, ∀λ ≥ ϕ(x), f (x) + kλ > α > f (x0 ) + kλ0 . In particular, for
λ = ϕ(x), and ∀x ∈ D(ϕ),

f (x) + kϕ(x) > α > f (x0 ) + kλ0 (1.2)

We consider the sign of k. At x0 , f (x0 ) + kϕ(x0 ) > f (x0 ) + kλ0 ⇒

k > 0, since (x0 , λ0 ) was chosen so that ϕ(x0 ) > λ0 . So, we divide both sides of
Equation (1.2) by k :

f (x) f (x0 )
+ ϕ(x) > + λ0
k k
f (x) f (x0 )
⇒− − ϕ(x) < − − λ0 ∀x ∈ D(ϕ)
k k
The left hand side is linear in x. So, taking supremums in x, we get that:
 
f (x) f (x0 )
sup − − ϕ(x) ≤ − − λ0
x k k
 
f α
=⇒ ϕ∗ − ≤ − < ∞.
k k

We can also define the bi-conjugate of ϕ in the following manner:

ϕ∗∗ : X → (−∞, +∞], ϕ∗∗ (x) = sup [< f, x > −ϕ∗ (f )].
f ∈D(ϕ∗ )

This function is convex and lower semicontinuous. So, the diagram looks like:

ϕ −→ ϕ∗ −→ ϕ∗∗
6 ≡ convex convex
∞ l.s.c. l.s.c.

Theorem 1.3.3 (Fenchel-Moreau) If ϕ is convex and l.s.c. and 6≡ +∞, then

ϕ∗∗ = ϕ.
 11

Proof First, we show that ϕ∗∗ ≤ ϕ : By definition of ϕ∗ , ∀x ∈ X, f ∈ X ∗ :

< f, x > −ϕ(x) ≤ ϕ∗ (f ) (1.3)

(1.3) =⇒ ∀x ∈ X, sup (< f, x > −ϕ∗ (f )) ≤ ϕ(x).

f ∈X ∗

Hence, ϕ∗∗ (x) ≤ ϕ(x).

Now, assume by contradiction that ∃x0 such that ϕ∗∗ (x0 ) < ϕ(x0 ). Then,
epi(ϕ) lies ”above” (x0 , ϕ∗∗ (x0 )). In other words, we can use the Hahn-Banach
Theorem to separate epi(ϕ) and {(x0 , ϕ∗∗ (x0 ))}. So, there exists, f ∈ X ∗ ,
k ∈ R, α ∈ R such that ∀x ∈ D(ϕ) and λ ≥ ϕ(x) :

f (x) + kλ > α > f (x0 ) + kϕ∗∗ (x0 ) (1.4)

Note that we used a similar technique as in Proposition 1.3.2 to break down the
operator given to us by the H-B Theorem into f and k. Again, we can conclude
that k ≥ 0, for else, we could send λ → +∞ and get a contradiction.
So, we first assume that ϕ(x) ≥ 0. Applying the relation in Equation 1.4 to
λ = ϕ(x), we get that f (x) + kϕ(x) > α. Hence, for all  > 0 :
f (x) α
f (x) + (k + )ϕ(x) > α =⇒ − − ϕ(x) < − ∀x
k+ k+
   
f (x) α f α
=⇒ sup − − ϕ(x) ≤ − =⇒ ϕ∗ − ≤ −
x∈D(ϕ) k +  k +  k +  k +

So, we see that:

   
f f
ϕ∗∗ (x0 ) = sup (< f, x0 > −ϕ∗ (f )) ≥ − , x0 − ϕ∗ −
f ∈X ∗ k+ k+
 
f α
≥ − , x0 +
k+ k+
=⇒ (k + )ϕ∗∗ (x0 ) ≥ < −f, x0 > +α

So, we take  → 0, and get that f (x0 ) + kϕ∗∗ (x0 ) ≥ α, which contradicts
α > f (x0 ) + kϕ∗∗ (x0 ) in Equation 1.4. Hence, ϕ ≥ 0 ⇒ ϕ∗∗ = ϕ.
Now, we consider any ϕ and f0 ∈ D(ϕ∗ ). Define a new function:

ϕ(x) = ϕ(x)− < f0 , x > +ϕ∗ (f0 ).

Fix x. ϕ∗ (f0 ) = supy [< f, y > −ϕ(y)] ≥ f0 (x) − ϕ(x). Since x was arbitrary,
this shows that ϕ ≥ 0. So, we can apply the result we obtained above to see
that ϕ = ϕ∗∗ . But,

ϕ∗ (f ) = sup [< f, x > −ϕ(x)] = sup [< f, x > + < f0 , x > −ϕ(x) − ϕ(f0 )]
x x
= sup [< f + f0 , x > −ϕ(x)] − ϕ∗ (f0 ) = ϕ∗ (f + f0 ) − ϕ∗ (f0 )
x
12

Also,

ϕ∗∗ (x) = sup [< f, x > −ϕ∗ (f )] = sup [< f, x > −ϕ∗ (f + f0 ] + ϕ∗ (f0 )
f f
= sup [< f + f0 , x > −ϕ∗ (f + f0 )]− < f0 , x > +ϕ∗ (f0 )
f
= sup [< g, x > −ϕ∗ (g)]− < f0 , x > +ϕ∗ f0
g
= ϕ∗∗ (x) + ϕ∗ (f0 )− < f0 , x >

Here, we have used the fact that f0 is independent of the sup taken over all f.
Hence, putting everything together, since ϕ∗∗ = ϕ, we get that ϕ∗∗ = ϕ.

Example Say that ϕ(x) = kxk. Surely, ϕ is a convex and lower semicontinous
function from X to R. ϕ∗ (f ) = supx∈X [< f, x > −kxk].
• If kf k ≤ 1, then < f, x >≤ kxk =⇒ ϕ∗ (f ) ≤ 0 =⇒ ϕ∗ (f ) = 0 (since we
can just take x = 0 and the sup will be at least 0).
• If kf k > 1, then ∃x such that f (x) > (1 + )kxk. So, f (x) − kxk > kxk.
If we consider the case of nx and then letting n go to +∞ we see that
ϕ∗ (f ) = +∞.
This means that:

ϕ∗∗ (x) = sup [< f, x > −ϕ∗ (f )] = sup (< f, x >) = kxk = ϕ(x).
f ∈D(ϕ∗ ) kf k≤1

Theorem 1.3.4 (Fenchel-Rockafellar) Assume ϕ, ψ are two convex func-

tions and ∃x0 ∈ X such that ϕ(x0 ) < ∞, ψ(x0 ) < ∞ and ϕ is ontinuous
at x0 . Then,

inf (ϕ(x) + ψ(x)) = sup [−ϕ∗ (−f ) − ψ ∗ (f )] = max∗ [−ϕ∗ (−f ) − ψ ∗ (f )]

x∈X f ∈X ∗ f ∈X

Proof Exercise. The proof is similar to the previous result.

Remark This theory has wide range of applications:

• Optimization (sometimes the dual problem is easier to deal with)

• PDE
• Convex Programming (see Ekeland - Teman, Intro to Convex Analysis).
Chapter 2

Baire Category Theorem

and Its Applications

2.1 Review
2.1.1 Reminders on Banach Spaces
Definition A Banach space is a complete normed linear space (i.e. every
Cauchy sequences converges in that space w.r.t. to its norm).
Example
• Hilbert spaces are Banach spaces
• Lp (X, dµ), 1 ≤ p ≤ ∞ are Banach spaces.
1/p
• lp = {(un )n∈N : ( n |un |p )P < ∞}, 1 ≤ p ≤ ∞ are Banach. (Note that
P
p
lp = L (R, dµ) where µ = n δn and the δn are the Dirac masses at the
integers.)

2.1.2 Bounded Linear Transformations

Definition A bounded linear transformation (or bounded operator) T between
two normed linear spaces (X1 , k · k1 ) and (X2 , k · k2 ) is a linear mapping such
that ∃C ≥ 0 such that:
∀x ∈ X1 , kT xk2 ≤ Ckxk1 .
As we had shown before, T bounded ⇐⇒ T continuous ⇐⇒ T continuous at
a point.

We define the operator norm as:

kT k = sup kT xk2
kxk1 ≤1

13
14

Lemma 2.1.1 The operator norm is a norm.

Proof First, we show the triangle inequality. This follows from the fact that
kT x+Sxk2 ≤ kT xk2 +kSxk2 since k·k2 is a norm and that supkxk1 ≤1 kT xk2 + kSxk2 ≤
supkxk1 ≤1 kT xk2 + supkxk1 ≤1 kSxk2 .
Next, we show that kaT k = |a|kT k. Again, this follows from the fact that
kaT xk2 = |a|kT xk2 since k · k2 is a norm and from the fact that for any positive
number a, sup a· = a sup ·.
Finally, we show that kT k = 0 ⇒ T = 0. In other words, we must show that
T x = 0 ∀x ∈ X1 . Again, kT k = 0 → ∀kxk1 ≤ 1, kT xk2 = 0. Then, by linearity,
this means that ∀x ∈ X1 , kT xk2 = 0. But then, since k · k2 is a norm, this means
that T x = 0. Hence, T = 0.

We denote by L(X1 , X2 ), the space of bounded linear operators from X1 to X2 ,

with norm kT k given by above.

Example L(X, K) = X ∗

Theorem 2.1.2 If Y is a Banach space, then L(X, Y ) is as well.

Proof It is clear that the space is normed, and linear. Remains to show com-
pleteness. Let An be a Cauchy sequence of functions in L(X, Y ). i.e.:

kAn − Am k −→ 0
n, m → ∞

=⇒ ∀x ∈ X, An (x) is Cauchy. Hence, it has a limit, which we will denote by

A(x). The fact that the mapping x 7→ A(x) is linear is clear. Now, we seek to
show boundedness:

kA(x)kY = lim kAn (x)k ≤ lim sup kAn kkxkX .

n→∞

But, the {An }n∈N form a Cauchy sequence. Hence, they are uniformly bounded
by some constant C. Hence, kAxk ≤ CkxkX by above. Hence, A ∈ L(X, Y ).
Finally, we show that it is the limit of the An ’s:

k(An − A)(x)k = lim k(An − Am )(x)k

m→∞
≤ lim sup kAn − Am kkxkX
m→∞
≤ o(1)kxkX

since the An are Cauchy. Hence, kAn − AkL(X,Y ) → 0.

Definition We say two norms, k · k1 and k · k2 are equivalent, if ∃ constants,

C1 , C2 such that C1 k · k1 ≤ k · k2 ≤ C2 k · k1 .

Proposition 2.1.3 In a finite dimensional space, all norms are equivalent.

Equivalent norms define the same topology.
 15

This result is not true in infinite dimensions.

Definition An isomorphism between two normed linear spaces is a bounded
linear map which is bijective with bounded linear inverse.
k · k1 and k · k2 are equivalent ⇔ the function id : (X, k · k1 ) → (X, k · k2 ) is an
isomorphism (continuous with continuous inverse).

Example All separable (there exists a countable Hilbert basis) Hilbert spaces
are isomorphic to l2 .

2.1.3 Duals and Double Duals

Theorem 2.1.4 (Riesz Representation Theorem) Let H be a Hilbert
space. Then, for every F ∈ H ∗ , then ∃ a unique f ∈ H such that ∀x ∈ H,
F (x) =< f, x > .

So, H 7→ H ∗ is an isomorphism between H and H ∗ , through which we can

identify H and its dual.

Example
• (L2 )∗ = L2
0
• For 1 < p < ∞,(Lp )∗ = Lp where 1
p + 1
p0 = 1. In other words,

Theorem 2.1.5 Let 1 < p < ∞. Then, ∀F ∈ (Lp )∗ , there exists a unique
0
f ∈ Lp where p1 + p10 = 1 such that for all ϕ ∈ Lp ,
Z
F (ϕ) = f ϕdµ.

• While (L1 )∗ = L∞ , (L∞ )∗ ) L1 . In fact, (L∞ )∗ = { measures }.

Definition The double dual is the dual of the dual. i.e.: X ∗∗ = (X ∗ )∗ .

Consider J : X → X ∗∗ given by J(x)(v) = v(x), for v ∈ X ∗ . This is the

canonical embedding of X into X ∗∗ . It is an isometric embedding. To see this,
note that:

kJ(x)k = sup kJ(x)(f )k = sup | < f, x > | ≤ sup kf k · kxk ≤ kxk.

kf k≤1 kf k≤1 kf k≤1

On the other hand, by a Corollary to the Hahn-Banach Theorem (Corollary

1.1.6), ∃f ∈ X ∗ such that kf k ≤ 1 and < f, x >= kxk. Hence, the sup is
achieved and there is equality.
In general, J(X) ⊂ X ∗∗ . Formally, we say, “X ⊂ X ∗∗ ” with the canonical
identification. In finite dimensions, we have equality. But this is not necessarily
true in infinite dimensions.
16

Definition If J(X) = X ∗∗ , we say that X is reflexive.

Examples of Reflexive Spaces

• Hilbert spaces, as seen above.
0 00
• Lp spaces for 1 < p < ∞ (since (Lp )∗∗ = (Lp )∗ = Lp where 1
p + 1
p0 =1
and p10 + p100 = 1 ⇒ p00 = p).

• Not L1 since (L1 )∗ = L∞ but (L∞ )∗ ) L1 , as we saw above.

2.2 The Baire Category Theorem

This theorem is used to prove that some sets have non-empty interior.

Example of Usefulness If T is a linear map and T −1 (B(0, 1)) has non-empty

interior, then T is bounded.
To see this, note that if T −1 (B(0, 1)) has non-empty interior, then ∃x0 ,  > 0
such that T −1 (B(0, 1)) ⊃ B(x0 , ) ⇒ B(0, 1) ⊃ T (B(x0 , )) ⇒ ∀y such that
kyk < , kT (x0 + y)k ≤ 1 ⇒ kT (y)k ≤ 1 + kT (x0 )k. For all x,
x 
kT kxk 2 k ≤ C ⇒ kT (x)k ≤ C̃kxk.

Theorem 2.2.1 (Baire Category Theorem) Let X be a complete metric

space and let Fn be a sequence of closed subsets of X with empty interior (i.e.:
int(Fn ) = ∅) then int(∪(Fn )) = ∅.

Complementary Form of Theorem If On is a sequence of dense open sub-

sets of X, then ∩On is also dense.

Remark A subset whose closure has empty interior is called “nowhere” dense.

Proof of the Baire Category Theorem Let On be a sequence of dense open

subsets. Then, ∩On is dense if we can prove that it intersects every open set.
Let W be an arbitrary open set. Let x0 ∈ W. Then, ∃r0 > 0 such that
B(x0 , r0 ) ⊂ W. Since O1 is dense, its intersection with B(x0 , r0 ) is non-empty.
0
Hence, ∃x1 ∈ B(x0 , r0 )∩O1 . Since B(x0 , r0 )∩O1 = O1 is an intersection of open
0
sets, it is itself open. Hence, ∃r1 > 0 such that B(x1 , r1 ) ⊂ O1 and r1 < r20 .
So, by induction we build a sequence of xn ’s such that

B(xn , rn ) ⊂ B(xn−1 , rn−1 ) ∩ On and rn < rn−1 /2.

r0
Hence, d(xn , xn−1 ) ≤ rn−1 ≤ 2n−1 . This is a Cauchy sequence in ∩On . Since X
is complete, the xn have a limit, l ∈ X. Now, since the B(xn , rn ) are closed,
and form a decreasing sequence of sets, for each n, l ∈ B(xn , rn ) ⊂ On . Hence,
l ∈ ∩On and l ∈ ∩B(xn , rn ). Hence, l ∈ B(x0 , r0 ) ⊂ W. Thus, l ∈ W ∩ (∩On ).
Thus, ∩On intersects every open set and is therefore dense.
 17

2.3 The Uniform Boundedness Principle

Theorem 2.3.1 (Uniform Boundedness Principle (Banach-Steinhaus))
Let X be a Banach Space and Y a normed linear space. Let (Ti )i∈I be an arbi-
trary family of elements of L(X, Y ) such that:

∀x ∈ X, sup kTi (x)k < ∞

i∈I

Then:
sup kTi k < ∞
i∈I

Proof Consider the sets:

Fn = {x ∈ X : kTi (x)k ≤ n ∀i ∈ I}.

∪n∈N Fn = X because each x is in some Fn by assumption. Moreover, each Fn

is closed by the continuity of the Ti . The Baire Category Theorem says that if
the Fn have empty interior, then their union must have empty interior as well.
But, ∪n∈N Fn = X which certainly doesn’t have empty interior. Hence, at least
one of the Fn cannot have an empty interior. Suppose Fn0 is such an Fn . Then,
it follows that ∃x0 ,  > 0 such that B(x0 , ) ⊂ int(Fn0 ). So,
y  y 
∀y, kTi (x0 + )kY ≤ n0 =⇒ kTi ( )k ≤ n0 +kTi (x0 )k < n0 +C0 ∀i ∈ I
kyk 2 kyk 2
The last inequality is true since by assumption, for each y, the Ti (y) are bounded
uniformly in i. So, for all y, kTi (y)k < 2(C0+n0 ) kyk. Since each quantity on the
right is independent of i, we can take the sup over all i on both sides and get
that supi kTi k < 2(C0+n0 ) .

Corollary 2.3.2 Let (Tn )n∈N be a sequence of bounded linear functions between
two Banach spaces X and Y such that ∀x ∈ X, Tn (x) converges to a limit
denoted by T x. Then,
• T ∈ L(X, Y ).
• supn∈N kTn k < ∞.
• kT kL(X,Y ) ≤ lim inf n→∞ kTn kL(X,Y ) .

Proof That T is linear ought be clear. ∀x ∈ X, such that kxk = 1,

sup kTn (x)kY < ∞

n

since Tn (x) converges by assumption. Hence, by the Uniform Boundedness

Principle, supn∈N kTn k < ∞. This proves the second claim.
To show the first, ∀x ∈ X, kTn (x)kY ≤ CkxkX ⇒ kT xkY ≤ Ckxk by
passing to the limit (since the RHS in the first inequality is independent of n).
Hence, T ∈ L(X, Y ).
18

Finally, for the third claim, notice that ∀x ∈ X, n ∈ N,

kTn (x)kY ≤ kTn kkxkX . Passing to the limit again, we see that since
Tn (x) −→ T (x),
kT (x)kY ≤ lim inf kTn kkxkY
n→∞
=⇒ kT k ≤ lim inf kTn k
n→∞

Example Limits in distributions. It is sufficient to just have pointwise conver-

gence.
Corollary 2.3.3 Let B be a subset in a Banach space X. If ∀f ∈ X ∗ ,
f (B) = ∪x∈B f (x) is bounded, then B is bounded.
Proof The idea is to apply the Uniform Boundedness Principle to the family
{Tb }b∈B given by:
Tb : X ∗ −→ K, Tb (f ) =< f, b >
for each b ∈ B. But, we have the following:
∀f ∈ X ∗ , sup kTb (f )k < ∞
b∈B
⇐⇒ ∀f ∈ X ∗ , sup | < f, b > | < ∞
b∈B
⇐⇒ ∀f ∈ X ∗ , f (B) is bounded.
Hence, {Tb }b∈B satisfies the hypothesis of the Uniform Boundedness Principle.
So, ∃C such that ∀f ∈ X ∗ , kTb (f )k ≤ Ckf k ∀b ∈ B. ⇐⇒ ∀f ∈ X ∗ , ∀b ∈ B,
| < f, b > | ≤ Ckf k ⇐⇒ kbkX ≤ C.
Corollary 2.3.4 Let X be a Banach space and B 0 a subset of X ∗ . If
∀x ∈ X, B 0 (x) = ∪f ∈B 0 f (x) is bounded, then B 0 is bounded.
Proof Tf (x) = f (x) where Tf : X → K. So,
sup kTf (x)k < ∞ ∀x
f ∈B 0

So, by the Uniform Boundedness Principle,

sup kTf k < ∞
f ∈B 0

and we can finish as above .

2.4 The Open Mapping Theorem and Closed

Graph Theorem
Theorem 2.4.1 (The (Banach) Open Mapping Theorem) Let T be a lin-
ear map from the Banach space X, onto another Banach space Y. Then, T is
open: The image of any open set is open.
 19

Proof By translation and linearity, for any r > 0, it is enough to prove that

T (BX (0, r)) ⊇ BY (0, r0 ) for some r0 .

Define Fn = T (B(0, n)). Since T is onto, we have ∪n∈N Fn = Y. So, by Baire Cat-
egory Theorem, some Fn0 has nonempty interior. By rescaling,
int(T (B(0, 1))) 6= ∅. Hence, we can assume that for some  > 0,
B(0, ) ⊆ T (B(0, 1)) (since it has non-empty interior).
We are going to show that T (B(0, 1)) ⊆ T (B(0, 3)) and therefore, that
B(0, ) ⊆ T (B(0, 3)). Since we’re in a linear space and T is linear, we can
rescale so that B(0, r/3) ⊆ T (B(0, r)) and so we can choose r0 = r/3.
So, let y be in T (B(0, 1)) and we will find x ∈ B(0, 3) such that y = T x. By
the definition of closure there exists x1 ∈ B(0, 1) such that

ky − T x1 k ≤
2
    
1
⇒ y − T x1 ∈ B 0, ⊂ T B 0, .
2 2

By the definition of closure, ∃x2 ∈ B(0, 1/2) such that ky − T x1 − T x2 k ≤ /4.

So, we iterate in this manner to get that:

 1
∀n, ∃xn such that ky − T x1 − . . . − T xn k < and kxn k < n .
2n 2
P∞
So, we take x = i=1 xi , which converges since the sequence is Cauchy. So,
∞
X
⇒ kxk ≤ kxi k ≤ 2 < 3
i=1
⇒ x ∈ B(0, 3) and T x = y
⇒ T (B(0, 1)) ⊂ T (B(0, 3))

Corollary 2.4.2 If T is a bounded linear map between two Banach spaces which
is also bijective, then its inverse is also continuous. Hence, T is an isomorphism.

Theorem 2.4.3 (Closed Graph Theorem) Let X, Y be two Banach spaces

and T : X → Y linear. Then, T is bounded if and only if its graph,
Γ(T ) = {(x, T x) : x ∈ X} ⊂ X × Y is closed.

Remark If X is a Banach space for both k · k1 and k · k2 , and ∃C > 0 such

that kxk1 ≤ Ckxk2 , then ∃C1 such that kxk2 ≤ C1 kxk1 .
Indeed, consider Id : (X, k · k2 ) → (X, k · k1 ). By assumption, it is a bounded
map that is also a bijection. So, by the corollary to the Open Mapping Theorem,
it has bounded inverse.
20

Proof of Closed Graph Theorem Apply the above remark to the norms:
k · k1 , k · k2 given by:

kxk1 = kxkX kxk2 = kxkX + kT xkY

Certainly, kxk1 ≤ kxk2 .

(=⇒) So, assume that Γ(T ) is closed. Is (X, k · k2 ) a Banach space? Well, take
a Cauchy sequence {xn }n∈N in (X, k · k2 ). Then, xn → x for k · k1 , since k · k1
is always bounded from above by k · k2 . Similarly, {T (xn )}n is Cauchy in Y. So,
since Y is Banach, it converges to some y ∈ Y. So, (xn , T (xn )) → (x, y). Hence,
T x = y since the graph of T is closed (and thus, the graph contains all limit
points). Therefore,

kxn − xk2 = kxn − xk1 + kT (xn − x)kY → 0 + 0 = 0.

This proves that (X, k · k2 ) is a Banach space. So, by the remark above,
kT (x)kY = kxk2 − kxk1 ≤ kxk2 ≤ C1 kxk1 = C1 kxkX , as desired.

(⇐=) Now, assume that T is bounded. Hence, it’s continuous. So, let
{(xn , T (xn ))}n be convergent in X × Y such that (xn , T (xn )) → (x, y) ∈ X × Y.
Then, xn → x, so T (xn ) → T (x) by continuity of T. Hence, y = T x and
(x, y) = (x, T x) ∈ Γ(T ). This proves the graph is closed.
Chapter 3

Weak Topology

3.1 General Topology

Definition A topological space is a set S with a distinguished family of subset
τ called the topology (a.k.a. all open sets) satisfying:
• ∅ and S are in τ.
• A finite intersection of elements of τ is in τ.
• An arbitrary union of elements of τ is in τ.

Definition A set S is closed if its complement is open.

Definition A family B ⊆ τ is called a base if every element of τ can be written

as a union of elements of B.

Definition A set N is a neighborhood of x ∈ S if there exists U ∈ τ such that

x ∈ U ⊂ N (the neighborhood does not have to be open).
Definition A family N is a neighborhood base of x if it is a family of neighbor-
hoods of x s.t. for every neighborhood M of x, ∃N ∈ N s.t. N ⊂ M.

Definition A function between two topological spaces is continuous if the in-

verse image of any open set is open.

Definition A topological space is Hausdorff if ∀x, y ∈ S, there exists Ox , Oy ∈

τ such that x ∈ Ox , y ∈ Oy and Ox ∩ Oy = ∅.

Example Metric spaces are Hausdorff.

Definition A set K is compact if every open cover of K has a finite subcover.

21
22

Remark

• The image of a compact set by a continuous function is compact.

• Two extreme cases: τ = ∪x∈S {x}, the discrete topology, where the only
sequences that converge are constant sequences and τ = {∅, S}, the inde-
screte topology, where all sequences converge.

• More generally, the more open sets there are, the harder it is to converge.

Now, let ϕi : X → Yi , i ∈ I be mappings from X to topological spaces Yi . What

is the weakest topology on X that makes all the ϕi continuous?
Obviously, it must contain ϕ−1i (Oi ) where Oi is any open set in Yi , along
with arbitrary unions and finite intersections. So, the answer is:
[ \
τ ={ ϕ−1
i (Oi ) }
arbitrary finite

where the Oi are open in Yi .

3.2 Frechet Spaces

Definition A seminorm ρ on a linear space X is a map from X to [0, +∞)
that satisfies the following:

1. ρ(x + y) ≤ ρ(x) + ρ(y)

2. ρ(λx) = |λ|ρ(x) ∀λ ∈ K

A family of seminorms, {ρα }α∈A is said to separate points if and only if

ρα (x) = 0 ∀α =⇒ x = 0.

Definition A locally convex space is a linear space endowed with a family of

seminorms, {ρα }α∈A , which separate points. The natural topology is the one
that makes all of the ρα continuous, and makes addition in the space continuous.

In a locally convex space, a basis of neighborhoods of 0 is given by sets of the

form:
Nα1 ,...,αN ; = {x ∈ X : ραi (x) < , ∀i = 1, . . . , N }.
A basis of neighborhoods of any point x0 ∈ X is given by sets of the form:

Nα1 ,...,αN ; = {x ∈ X : ραi (x − x0 ) < , ∀i = 1, . . . , N }.

Characterization A linear mapping T is continuous if and only if

∃C > 0 such that kT (x)k ≤ C(ρα1 (x) + . . . + ραN (x)).

Proposition 3.2.1 A locally convex space is Hausdorff

 23

Proof Take x 6= y. Then, ∃α such that ρα (x − y) 6= 0 (otherwise, we’d have

x−y = 0 since the family of seminorms separates points). Now, let η = ρα (x−y)
and let:
Ox = {z ∈ X : ρα (z − x) < η/4}
Oy = {z ∈ X : ρα (z − y) < η/4}
By the definition of the locally convex topology, these sets are open. Further-
more, if z ∈ Ox ∩ Oy , then:
η = ρα (x−y) ≤ ρα (x−z)+ρα (z −y) = ρα (z −x)+ρα (z −y) < η/4+η/4 = η/2,
which yields an obvious contradiction. Hence, Ox ∩ Oy = ∅.
Convergence of Sequences In this topology, xn → x if and only if ∀α ∈ A,
ρα (xn − x) → 0.
Definition
• A convex set in a linear space is called balanced or circled if x ∈ C ⇒
λx ∈ C ∀λ, |λ| = 1.
• It is called absorbing if [
tC = X.
t>0

Remark If ρα is a family of seminorms on X then the sets

Nα1 ,...,αN ; = {x ∈ X : ραi (x) < , ∀i = 1, . . . , N }
are convex, balanced, absorbing sets.
Theorem 3.2.2 Let X be a linear space with a Hausdorff topology in which
addition and scalar multiplication are continuous. Then, X is a locally convex
space if and only if 0 has a basis of neighborhoods which are convex, balanced
(circled) absorbing sets.
Proof (=⇒) This follows from the preceding remark.

(⇐=) What we want to do here is to build the family of seminorms. Take C to

be a convex neighborhood of 0 and let ρC be its gauge:
x
ρC (x) = inf{t > 0 : ∈ C}.
t
It is easy to check that ρC is a seminorm. Also,
{ρC (x) < 1} ⊆ C ⊆ {ρC (x) ≤ 1}.
But that means that the neighborhood basis given by the seminorms is the same
as the original neighborhood basis given by the C’s. Hence, the two topologies
are the same, i.e. the original topology is induced by seminorms. So, the space
is locally convex.
24

Proposition 3.2.3 Let X be a locally convex vector space. The following are
equivalent:

1. X is metrizable (the topology is induced by a distance).

2. 0 has a countable basis of neighborhoods that are convex, balanced, absorb-
ing.
3. The topology is generated by a countable family of seminorms.

Proof
(1) =⇒ (2). Take balls of countable radius (say the rationals).
(2) =⇒ (3). Do the previous construction using gauges.
(3) =⇒ (1). The distance can be given by:
∞
X 1 ρn (x − y)
d(x, y) = n 1 + ρ (x − y)
.
n=0
2 n

Definition A Frechet space is a complete, metrizable locally convex space.

In particular, the Baire Category Theorem applies to Frechet spaces.

Example The Schwartz Class, S of functions of rapid decrease:

S = {f : Rn → C : sup |x|α |∂ β f (x)| < C ∀α ∈ Z, ∀β multi-index of integers }.

x∈Rn

For f ∈ S, define: kf kα,β = supx |x|α |∂ β f (x)|

The set S ∗ (the dual of S = the space of all continuous linear functions on
S) is called the space of all tempered distributions.
S is a Frechet space.

Example Let D(Ω) = C ∞ (Ω) with seminorms given by kf kβ = supx∈Ω |∂ β f (x)|.

Let D0 (Ω) be D(Ω)∗ = dual of D(Ω) = space of distributions.

T ∈ D0 (Ω) ⇐⇒ T is continuous, linear

X
⇐⇒ ∃C, n such that T (f ) ≤ C kf kβ
|β|≤n

n is called the order of the distribution.

3.3 Weak Topology in Banach Spaces

Definition Let X be a Banach space. The weak topology on X is defined as
the weakest topology which makes all of the f ∈ X ∗ continuous. In other words,
it is: [ \
= f −1 (O),
arbitrary finite

where O is open. It is denoted by σ(X, X ∗ ).

 25

Note:

• A weakly open set is always strongly open.

• In infinite dimensions, the weak topology is not metrizable.

• A basis of neighborhoods for x0 is given by sets of the form:

Nf1 ,...,fN ; = {x ∈ X : |fi (x − x0 )| < , ∀i = 1, . . . , N }

Proposition 3.3.1 The weak topology is Hausdorff

Proof Let x 6= y. Apply the Geometric Version of the Hahn-Banach Theorem

to x, y. Then, ∃f ∈ X ∗ such that f (x) < α < f (y). So, define:

O1 = f −1 ((−∞, α)), O2 = f −1 ((α, +∞))

O1 , O2 are weakly open, they separate x and y and are certainly disjoint.

Remark Given a sequence {xn }n , we distinguish between:

1. xn → x strongly means convergence in the X norm.

i.e. kxn − xkX → 0.

2. xn * means that xn → x in the weak topology.

i.e. ∀f ∈ X ∗ , f (xn ) → f (x).

Proposition 3.3.2 Let {xn }n be a sequence in X. Then, the following are

true:

1. xn * x if and only if f (xn ) → f (x) ∀f ∈ X ∗ .

2. If xn → x, then xn * x (The converse is not true, however).

3. If xn * x then, {kxn kX }n is bounded and

kxkX ≤ lim inf kxn kX .

n→∞

4. If xn * x and fn → f in X ∗ , then fn (xn ) → f (x).

Proof
(1) This is the definition of weak convergence.

(2) If xn → x, then:

|f (xn ) − f (x)| ≤ kf kX ∗ kxn − xkX → 0

since kxn − xkX → 0, independent of f. Hence, f (xn ) → f (x) ∀f ∈ X ∗ . So,

xn * x.
26

(3) ∀f ∈ X ∗ , {f (xn )}n is bounded. By a corollary of the Uniform Boundedness

Principle (Corollary 2.3.3), we deduce that {xn }n is bounded. So,
|f (xn )| ≤ kf kX ∗ kxn kX
|f (x)| ≤ lim inf kf kX ∗ kxn kX
n→∞
= kf kX ∗ (lim inf kxn kX ).
n→∞

But, kxkX = supkf kX ∗ ≤1 |f (x)|. So,

kxkX = sup |f (x)| = sup kf k∗X (lim inf kxn kX ) ≤ lim inf kxn kX
kf kX ∗ ≤1 kf kX ∗ ≤1 n→∞ n→∞

(4)
|fn (xn ) − f (x)| ≤ |fn (xn ) − f (xn )| + |f (xn ) − f (x)|
≤ kfn − f kX ∗ kxn kX + |f (xn ) − f (x)| −→ 0
since fn → f and f (xn ) − f (x) → 0 for all f, by the weak convergence of {xn }n
to x, and since {xn }n is bounded because of its weak convergence to x.
Proposition 3.3.3 If dim X < ∞, then weak and strong topologies coincide.
Proof Surely, a weakly open set is strongly open. But is a strongly open set,
weakly open? Let U be strongly open with x0 ∈ U. So, there is r > 0 such that
B(x0 , r) ⊆ U. Let {e1 , . . . , en } be a basis for X with kei k = 1. Let {f1 , . . . , fn }
be the dual basis. In other words, fj (ei ) = P δi,j . The dual basis has the property
that if we can expand any y ∈ X via: y = fi (y)ei . Then the set
r
N = {x ∈ X : |fi (x − x0 )| < ∀i = 1, . . . , n}
n
is weakly open. So,
n
X n
X
x ∈ N ⇒ kx − x0 kX = k fi (x − x0 )ei k ≤ |fi (x − x0 )| < r.
i=1 i=1

Hence, N ⊆ B(x0 , r) ⊆ U. Thus, U is weakly open.

Example If dim X = ∞, then S = {x ∈ X : kxkX = 1} is not weakly closed.
In fact, its weak closure is B̂X = {x ∈ X : kxkX ≤ 1}.
Proof of this fact Let x0 ∈ BX . We will show every weak neighborhood of x0
intersects S. Take any U of the form:
U = {x ∈ X : |fi (x − x0 )| < , ∀i = 1, . . . , n}.
Then, ∃y0 ∈ X such that f1 (y0 ) = . . . = fn (y0 ) = 0. If not, then the function
x 7→ (f1 (x), . . . , fn (x)) would be a one-to-one map, meaning that
dim X ≤ n < ∞. Therefore,
∀t ∈ R, fi ((x0 + ty0 ) − x0 ) = tfi (y0 ) = 0.
 27

Hence, x0 + ty0 ∈ U, ∀t ∈ R. So, take g(t) = kx0 + ty0 k. Then,

g(0) = kx0 k < 1, g is continuous, and g → ∞ as t → ∞. Hence, g must take on
all the values between kx0 kX < 1 and ∞. Hence, ∃t0 such that g(t0 ) = 1. So,
x0 +t0 y0 ∈ S∩U.T hisprovesthattheweakclosureof ScontainsB.W ewilllaterseethatitisB,sinceBisweaklyclosedbyconv

Example BX = {x ∈ X : kxk < 1} is not weakly open. It has empty interior

since every weak neighborhood of x0 ∈ BX contains an element of S.

Theorem 3.3.4 Let C ⊆ X be a convex set. Then, C is weakly closed if and

only if C is strongly closed.

Proof
(=⇒) Since weakly open =⇒ strongly open, taking complements, weakly closed
=⇒ strongly closed.

(⇐=) Assume C is strongly closed. Then, we show that C is weakly closed. i.e.,
we show that C c is weakly open. Let x0 ∈ C c . By the Hahn Banach Theorem
(Second Geometric Form), ∃f ∈ X ∗ , α ∈ R such that f (x0 ) < α < f (x),
∀x ∈ C. So, N = f −1 ((−∞, α)) is a weakly open set (since it is the inverse
image of an open set under a continuous function), containing x0 and included
in C c . Hence, C c is weakly open.

Corollary 3.3.5 Let ϕ be a convex , lower semi-continuous function (for the

strong topology). Then, ϕ is lower semi-continuous for the weak topology. In
particular, if xn * x, then ϕ(x) ≤ lim inf ϕ(xn ).

Proof ϕ strongly lower semi-continuous =⇒ {ϕ(x) ≤ λ} is convex and strongly

closed. =⇒ The set is weakly closed. Hence, ϕ is weakly l.s.c.

Remark Therefore, convex, strongly continuous =⇒ weakly l.s.c.

Example x 7→ kxkX is a convex, continuous function. Hence, it is weakly l.s.c.

So, if xn * x then, kxk ≤ lim inf kxn k is reproved.

Theorem 3.3.6 Let X and Y be two Banach spaces and T : X → Y linear.

Then, T is strongly continuous if and only if it is continuous from σ(X, X ∗ ) to
σ(Y, Y ∗ ).

Proof
(=⇒) Assume that T is strongly continuous. Let f ∈ Y ∗ . So, take any set in
σ(Y, Y ∗ ) of the form, f −1 ((a, b)) ⊂ Y. Then, T −1 (f −1 ((a, b))) = (f ◦T )−1 ((a, b)).
But, f ◦ T : X → Y is continuous and linear. Hence, (f ◦ T )−1 (a, b) is open in
σ(X, X ∗ ), being the inverse image of an open set under a continuous function.
Thus, T is weakly continuous.

(⇐=) Conversely, assume that T is weakly continuous. Γ(T ) is weakly closed

(i.e.: closed in σ(X × Y, (X × Y )∗ )). So, Γ(T ) is strongly closed. Hence, T is
strongly continuous by the Closed Graph Theorem.
28

3.4 Weak-* Topologies σ(X ∗ , X)

On X ∗ we can define the weak topology, σ(X ∗ , X ∗∗ ). But, X ⊆ X ∗∗ . So, tech-
nically, there is something even weaker than the weak topology.

Definition The weak-* topology on X ∗ is defined as the weakest topology which

makes all the maps f 7→ f (x) continuous. σ(X ∗ , X) is weaker than σ(X ∗ , X ∗∗ ).

Proposition 3.4.1 σ(X ∗ , X) is Hausdorff.

Proof If f1 , f2 ∈ X ∗ and f1 6= f2 , then ∃x ∈ X such that f1 (x) 6= f2 (x). So,

∃α such that f1 (x) < α < f2 (x). So, define the following sets:

O1 = {f ∈ X ∗ : f (x) < α}, O2 = {f ∈ X ∗ : f (x) > α}.

O1 and O2 are open in σ(X ∗ , X) and separate f1 and f2 .

A basis of neighborhoods of f0 for σ(X ∗ , X) is given by sets of the form:

Nx1 ,...,xn ; = {f ∈ X ∗ : |(f − f0 )(xi )| < , ∀i = 1, . . . , n}

∗
We say, fn * f (fn converges weakly-* to f ) if fn → f in σ(X ∗ , X). In other
words, ∀x ∈ X, fn (x) → f (x).

Properties
∗
1. fn * f ⇐⇒ ∀x ∈ X, fn (x) → f (x).

2. fn → f in X ∗
=⇒ fn * f in σ(X ∗ , X ∗∗ )
∗
=⇒ fn * f in σ(X ∗ , X)
∗
3. If fn * f, then kfn kX ∗ bounded and kfn kX ∗ ≤ lim inf kfn kX ∗ .
∗
4. If fn * f, and xn → x in X, then fn (xn ) −→ f (x).

Theorem 3.4.2 (Banach-Alaoglu) Let X ∗ be the dual of a Banach space.

Then,
BX ∗ = {f ∈ X ∗ : kf kX ∗ ≤ 1}
is compact for the weak-* topology.

Remark Observe right-away that compactness 6⇒ sequential compactness. It

is only true if the space is metrizable.
Compare this to Riesz’ Theorem which states that the unit ball of a Banach
space is strongly compact if and only the dimension is finite.
 29

Proof of Banach-Alaoglu’s Theorem Tychonoff’s Theorem states that any

product of compact spaces is compact for the product topology.
So, apply Tychonoff’s Theorem to:
Y
A= B(0, kxkX )
x∈X

A is therefore compact for the product topology.

Elements of A are assignments x 7→ g(x). So, they are functions of x which
satisfy |g(x)| < kxkX . Let à be the subset of A containing all linear functions.
So, we can write: \ \
à = Ax,y × Bλ,x ,
x,y∈X x∈X,λ∈K

where:

Ax,y = {f ∈ A : f (x + y) − f (x) − f (y) = 0}

Bx,λ = {f ∈ A : f (λx) − λf (x) = 0}.

These are closed in the product topology. So, Ã is a closed subset of a compact
set and so, Ã is compact for the product topology. But, the product topology
on à is the weak-* topology. Hence, à = BX ∗ is compact in σ(X ∗ , X).

3.5 Reflexive Spaces

Definition X is said to be reflexive if X ∗∗ = X.

Theorem 3.5.1 (Kakutani) Let X be a Banach Space. Then, the closed unit
ball, BX = {x ∈ X : kxk ≤ 1} is compact for the weak topology σ(X, X ∗ ) if and
only if X is reflexive.

Before we prove Kakutani’s Theorem, we need several lemmas by Helly and

Goldstein.

Lemma 3.5.2 (Helly) Let X be a Banach space, f1 , . . . , fn ∈ X ∗ and

α1 , . . . , αn ∈ R. Then, the following conditions are equivalent:

1. ∀ > 0 ∃x kx k ≤ 1 such that:

| < fi , x > −αi | <  ∀i = 1, . . . , n.

Pn Pn
2. ∀βi , | i=1 βi αi | ≤ k i=1 βi fi kX ∗ .

Proof (1) =⇒ (2): From (1), we get that ∀βi ,

 n n
 n
X X  X
βi < fi , x > − βi αi  <  |βi |
 

 
i=1 i=1 i=1
30
 n   n  n
X  X  X
βi αi  ≤  < βi fi , x > +  |βi |
   

   
i=1 i=1 i=1
* n + n
 X  X
=  βi fi , x  +  |βi |
 
 
i=1 i=1

X n Xn
≤ βi fi kx kX +  |βi |

∗
i=1 X i=1

But, kx kX ≤ 1. So, let  → 0. Then,

  n
Xn  X
βi αi  ≤ βi fi .
 

 
i=1 i=1 X∗

(2) =⇒ (1): Assume not. Then, let ϕ ~ (x) = (< f1 , x >, . . . , < fn , x >) . Then,
(α1 , . . . , αn ) ∈
/ϕ~ (BX ). Since {(α1 , . . . , αn )} = {α} is a compact set and ϕ ~ (BX )
is closed and convex, we can apply the Hahn-Banach Theorem and say that ∃γ
and β ~ such that β ~·α
~ >γ>β ~·ϕ~ (x) ∀x ∈ BX . So,
n
X n
X
∀x ∈ BX , βi αi > γ > βi < fi , x >.
i=1 i=1

Changing x to −x above, we get that:

n  n 
X X 
βi fi (x) < γ <  βi αi  .
 

 
i=1 i=1

Taking the sup over x ∈ BX :

n  n 
X X 
βi fi ≤γ< βi αi  ,
 

 
i=1 X∗ i=1

contradicting the assumption made in (2).

Lemma 3.5.3 (Goldstine) J(BX ) is dense in BX ∗∗ for σ(X ∗∗ , X ∗ ). Here,

J : X → X ∗∗ , J(x) =< x, · > .

Proof We prove that for every η ∈ BX ∗∗ , every neighborhood of η for

σ(X ∗∗ , X ∗ ) intersects J(BX ).
So, take η ∈ BX ∗∗ . We can assume that the neighborhood is:

{ζ ∈ X ∗∗ : |< ζ − η, fi >| < , fi ∈ X ∗ , i = 1, . . . , n}.

So, is there x ∈ BX such that |< x − η, fi >| <  for i = 1, . . . , n? This is

equivalent to asking is there x ∈ BX such that |< fi , x > − < η, fi >| <  for
 31

eachPi? Let αi =< P η, fi > . By Helly’s Lemma, this can only happen if and only
n n
if | i=1 βi αi | ≤ k i=1 βi fi kX ∗ . Since, η ∈ BX ∗∗ ,
  n
Xn  X
∀βi ,  < βi fi , η > ≤ βi fi .
 
  ∗
i=1 i=1 X

But, αi =< η, fi > . So, we have that:

 n   n   n  n
X  X  X  X
βi αi  =  βi < fi , η > =  < βi fi , η > ≤ βi fi .
     

     
i=1 i=1 i=1 i=1 X∗

We are now ready to prove Kakutani’s Theorem.

Proof of Kakutani’s Theorem

(⇐=) If X is reflexive, then apply the Banach-Alaoglu Theorem to X ∗ . Since
X = (X ∗ )∗ , the result follows

(=⇒) We must show that X ∗∗ = X. But, this is equivalent to showing that

J(BX ) = BX ∗∗ by linearity of J. By Theorem 3.3.6, if T is a linear operator,
then it is strong-strong continuous if and only if it is weak-weak continuous.
Hence, J is Continuous from σ(X, X ∗ ) to σ(X ∗∗ , X ∗∗∗ ). This is more de-
manding that J being continuous from σ(X, X ∗ ) to σ(X ∗∗ , X ∗ ) since
X ∗∗∗ ⊇ X ∗ . Therefore, J is continuous σ(X, X ∗ ) to σ(X ∗∗ , X ∗ ). Since J(BX ) is
compact for σ(X ∗∗ , X ∗ ), it is closed. So, by Goldstein’s lemma, J(BX ) is dense
in BX ∗∗ and closed. Hence, J(BX ) = BX ∗∗ . This proves that J(X) = X ∗∗ .
Hence, X is reflexive.

Corollary 3.5.4 If M is a closed subspace of a reflexive space X, then M is

reflexive.

Proof BM is a weakly closed subset of the compact set BX because it’s convex.
Hence, BM is weakly compact. Hence, M is reflexive.

Corollary 3.5.5 Let X be a reflexive Banach space. If C is a closed (strong

or weak), convex, bounded set, then C is compact for σ(X, X ∗ ).

Proof C is weakly closed and C ⊆ mBX for some m > 0. Since mBX is
compact for σ(X, X ∗ ), C is compact for σ(X, X ∗ ) as well.

Proposition 3.5.6 Let X be a reflexive Banach space, and ϕ 6≡ +∞, a convex,

lower semi-continuous function from a closed, convex set A to (−∞, +∞] such
that either A is bounded or limx∈A,kxk→∞ ϕ(x) = +∞. Then, ϕ achieves its
minimum on A.

Proof

Property A lower semi-continuous function achieves its min on a compact set.

32

Let λ = ϕ(x0 ) < ∞ for some x0 . Then, if we define:

à = {x ∈ A : ϕ(x) ≤ λ},
is a convex, strongly closed set (because ϕ is l.s.c.). Hence, Ã is weakly closed.
Since it’s bounded by assumption, it is weakly compact. Hence, since ϕ is
convex, l.s.c. it is weakly l.s.c.. Hence, ϕ achieves its minimum on Ã. Since
à ⊂ A, ϕ achieves its minimum on A.

3.6 Separable Spaces

Definition X is separable if X has a countable, dense subset.
Theorem 3.6.1 BX ∗ is metrizable for the σ(X ∗ , X) topology if and only if X
is separable, with the metric given by:
∞
X 1
d(f, g) = n
|< f − g, xn >|,
n=1
2

where the {xn }n is the countable dense set in X.

Remark BX ∗ is metrizable by not X ∗
Corollary 3.6.2 Let X be separable. Let {fn }n be a bounded sequence in X ∗ .
Then, there exists a subsequence {fnk }k converging weakly-*.
Proof We assume WLOG that {fn }n ⊂ BX ∗ . By Banach-Alaoglu, BX ∗ is
weakly-* compact. Since BX ∗ is metrizable, by Theorem 3.6.1, we have that
BX ∗ is sequentially compact.
Proposition 3.6.3 Let X be a reflexive space and {xn }n a bounded sequence
in X. Then, there exists a subsequence {xnk }k which converges in σ(X, X ∗ ).
Proof X reflexive =⇒ BX is compact. Let M = Span{x1 , x2 , . . .}. Then,
M is a separable Banach space, which is also reflexive. So, BM is compact for
σ(X, X ∗ ). Hence, we may extract a convergent subsequence.
Remark These two results show that for a reflexive space X, BX is both com-
pact and sequentially compact.

3.7 Applications
3.7.1 Lp Spaces
0
For 1 < p < ∞, the dual of Lp is Lp where 1/p + 1/p0 = 1. So, what is weak
convergence in Lp ? Answer:
Z Z
0
fn * f in Lp ⇐⇒ ∀g ∈ Lp , fn g → f g.

Recall that the definition of strong Lp convergence is |fn − f |p → 0.

R
 33

Example
• Consider {fn (x) = sin nx}n∈Z on [0, 1]. Then, ∀g ∈ C ∞ ([0, 1]),
Z 1
sin nxg(x) dx → 0.
0
0
Since C ∞ is dense in Lp , we see that sin nx * 0 weakly in Lp .
• On the other hand ∃C > 0 such that ∀n,
Z 1
| sin nx|2 dx = C.
0

Hence, {fn }n 6→ 0 strongly in Lp .

Example For 1 < p < ∞, Lp is reflexive and separable. Hence, the unit ball
B1 is weakly and weakly sequentially compact.

Example
• (L1 )∗ = L∞ , but (L∞ )∗ ) L1
(In fact, (L∞ )∗ = { Bounded Radon Measures }).
• Neither L∞ nor L1 is reflexive.
• L1 is separable, but L∞ is not.
• L1 is not the dual of any space.
• BL1 is not even weakly closed. Hence, it’s not weakly compact (Take
approximate identities and you’ll see that BL1 = B{ measures } .
• BL∞ is weak-* compact, but not weakly compact by Kakutani’s Theorem
(since L∞ is not reflexive.

3.7.2 PDE’s
Suppose we wish to solve the following PDE:
(
4u + |u| · u + u = f on Ω,
(3.1)
u=0 on ∂Ω.

where Ω is a bounded open set in R2 , and f is smooth.

Method of Calculus of Variations: We want to minimize the energy:

Z Z Z Z
1 1 1
F (u) = |∇u|2 + |u|3 + |u|2 − f u.
2 Ω 3 Ω 2 Ω Ω
34

Assume that u minimizes F. Then, ∀g ∈ C0∞ (Ω)(g = 0 on ∂Ω), if we set ϕ(t) =

F (u + tg),
d 
 ϕ(t) = 0, since ϕ(t) ≥ ϕ(0), ∀t.
dt t=0
Z Z Z
1 1 1
F (u + tg) = |∇(u + tg)|2 + |u + tg|3 + |u + tg|2 +
2 Ω 3 Ω 2 Ω
Z
+ f (u + tg)
Ω
Z Z
1 1
= (|∇u|2 + 2t∇u∇g) + (|u|3 + 3tg|u| · u)
2 Ω 3 Ω
Z Z
1
+ (|u|2 + 2tgu) − (f u + tg + O(t2 ))
2 Ω Ω
Z
d 
0 =  F (u + tg) = (∇u · ∇g + g|u| · u + ug − f g)
dt t=0 Ω

Integrating the first term by parts and noticing that g vanishes on ∂Ω, we see
that:
Z
0 = [(−4u)g + g|u| · u + ug − f g]
ZΩ
= [−4u + |u| · u + u − f ] g
Ω

Since g was an arbitrary member of C0∞ , we have that u solves 3.1 in the sense
of distributions.

So, to summarize, the minimizer of the equation:

Z Z Z Z
1 1 1
F (u) = |∇u|2 + |u|3 + |u|2 − f u.
2 Ω 3 Ω 2 Ω Ω

gives a weak solution to:

(
4u + |u| · u + u = f on Ω,
u=0 on ∂Ω.

So, now define F to be a function on the Sobolev space H01 (Ω). First, the norm
on H01 (Ω) is given by:
Z
kukH01 (Ω) = (|∇u|2 + |u|2 ).
Ω

So, with that norm, H01 (Ω) becomes the closure of C0∞ (Ω). Alternatively, we
can define H01 (Ω) to be the set of functions u ∈ L2 (Ω) whose weak derivative
∇u is also in L2 (Ω) and u = 0 on ∂Ω.
Fact: If dim = 1, H 1 ⊆ C 0 .
 35

Ry
Proof Let u ∈ H01 . Then, u(x) − u(y) = x
u0 (t)dt. Hence,
sZ
Z y  √
 y
0

|u(x) − u(y)| ≤  u (t)dt ≤ y − x |u0 (t)|2 dt ≤ CkukH01 (Ω) .
x x

1
by Cauchy-Schwartz. Hence, u ∈ C 0, 2 , the set of Hölder continuous functions
with Hölder exponent 12 .

Fact: If dim = 2, then H 1 is not a subset of C 0 . However, in any any dimension,

there is a continuous embedding of H 1 into Lp for p < ∞ and p ≤ p∗ where
1/p∗ = 1/2 − 1/d. We call p∗ the critical exponent.. By continuous embedding,
we mean that kukLp ≤ CkukH 1 .

Example So, for example, in dim = 2, p∗ = ∞. Hence, H 1 ⊆ Lp , ∀p < ∞. In

dim = 3, we have: H 1 ⊆ Lp , ∀p ≤ 6.

Remark In fact, the embedding H 1 ⊆ Lp is compact (the embedding is a

compact operator). This means that B(0, 1) in H 1 is mapped into a compact
set in Lp . This transforms weak convergence into strong convergence. In other
words, if un * u weakly in H 1 , then un → u strongly in Lp , ∀p < p.

Returning back to the problem, we still have not answered whether min F is
achieved. First, we check that F is coercive, i.e. F → ∞ as kukH01 (Ω) → ∞, so
that: {u : F (u) ≤ 1} is bounded and nonempty (since F (0) = 0).

Proof that F is Coercive

Z Z Z Z
1 1 1
F (u) = |∇u|2 + |u|3 + |u|2 − f u.
2 Ω 3 Ω 2 Ω Ω

for Ω ⊆ R2 . If u ∈ H01 (Ω), then u ∈ Lp (Ω) for all p < ∞. In particular,

u ∈ L3 (Ω). So,
1
F (u) ≥ kuk2H 1 (Ω) − kf kL2 (Ω) kukL2 (Ω) ,
2 0

where kukL2 (Ω) ≤ kukH01 (Ω) . Thus,

1 1
F (u) ≥ kuk2H 1 (Ω) + kuk2H 1 (Ω) − kf kL2 (Ω) kukH01 (Ω) ,
4 0
|4
0
{z }

where the grouped terms are bounded from below by −kf kL2 (Ω) . To see this,
consider the function x 7→ x2 /4 − kf kL2 (Ω) x, which has a minimum at x =
2kf kL2 (Ω) and takes the value −kf kL2 (Ω) there . Thus, F (u) ≥ 14 kuk2H 1 (Ω) − C
0
with C independent of u. Therefore, F (u) → ∞ as kukH01 (Ω) → ∞.

Lemma 3.7.1 F is weakly l.s.c.

36

Proof Note that the following functions are all strongly continuous and convex:
Z
1
u 7→ (|∇u|2 + |u|2 )
4 Ω
Z
1
u 7→ |u|3
3 Ω
Z
u 7→ − fu
Ω

Hence, F is weakly lower semi-continuous and convex. Therefore, by Proposition

3.5.6, F achieves its min on H01 (Ω).
Chapter 4

Bounded (Linear)
Operators and Spectral
Theory

4.1 Topologies on Bounded Operators

Let X, Y be Banach spaces and denote by L(X, Y ) to be the space of bounded
operators from X to Y, with the norm given by:
kT kL(X,Y ) = sup kT xkY
kxkX ≤1

Definition The topology on L(X, Y ) defined by this norm is called the uniform
topology. In that topology, (A, B) 7→ AB is jointly continuous.
Definition We define the strong topology as the weakest topology which makes
all the:
Ex : L(X, Y ) −→ Y, T 7→ T x
continuous (∀x ∈ X). It’s the topology of pointwise convergence. However, in
this topology, multiplication, (A, B) 7→ AB is separately continuous, but not
jointly continuous.
Definition We define the weak operator topology as the weakest topology which
makes all of the:
Ex,l : (X, Y ) −→ C, T 7→< l, T x >
for x ∈ X, l ∈ Y ∗ , continuous.
Remark It is akin to the convergence of all n matrix entries < l, T x > of T.
So, we write:
w
Tn −→ T, if ∀l ∈ Y ∗ , ∀x ∈ X, < l, Tn x > −→ < l, T x > .
uniform > strong > weak.

37
38

Example

|un |2 < ∞ given by:
P
• Bounded operators on l2 = {un }n :
u u 
1 2
Tn : (u1 , u2 , . . .) 7→ , ,... .
n n
It is not difficult to see that Tn −→ 0 uniformly.
• Consider the deletion operators on l2 :
Sn : (u1 , . . . , un , . . .) 7→ (0, 0, . . . , 0, un+1 , un+2 , . . .)
| {z }
n times

Clearly, Sn −→ 0 strongly. However, Sn 6→ 0 uniformly. To see this, fix

n > 0 consider a sequence u whose l2 norm is 1, such that ui = 0 for all
i ≤ n. Then, Sn (u) = u. Hence, kSn kL(X) ≥ 1. On the other hand, for any
u ∈ l2 with l2 norm 1, kSn (u)kl2 ≤ kukl2 . Hence, kSn kL(X) = 1. Since n
was arbitrary, kSn kL(X) = 1 for all n,
• Now, consider the shift operators Wn given by:
Wn : (u1 , u2 , . . .) 7→ (0, . . . , 0, u1 , u2 , . . .)
| {z }
n times

To see that Wn −→ 0 weakly, consider any functional f : l2 → R. Then,

for any u ∈ l2 ,
< f, Wn (u) >= f (0, . . . , 0, u1 , u2 , . . .)) → 0.
| {z }
n times

On the other hand, it is clear that for any u ∈ l2 , kWn (u)kl2 = kukl2 .
Hence, kWn kL(X) = 1 for each n. Hence, Wn 6→ 0 strongly.
Theorem 4.1.1 Let H be a Hilbert space and Tn ∈ L(H) such that ∀x, y ∈ H,
hTn x, yiH converges as n → ∞, then ∃T ∈ L(H) such that Tn → T in the weak
topology.
Proof Given x, ∀y ∈ Y, supn |hTn x, yi| < ∞. Hence, by the Uniform Bound-
edness Principle,
sup sup |hTn x, yi| < ∞ ⇐⇒ sup kTn xkH < ∞.
kykH ≤1 n n

This is true for any x ∈ H. So, again applying the Uniform Boundedness Princi-
ple, we see that supn kTn kL(H) < ∞. Now, we define B(x, y) = limn→∞ hTn x, yi.
One can see that B is sesquilinear. Furthermore,
|B(x, y)| ≤ lim sup kTn kL(H) kxkH kykH ≤ CkxkH kykH
n

Therefore, by a corollary of the Riesz Representation Theorem (not proven in

class), ∃T ∈ L(H) such that B(x, y) = hTn x, yi . Then, it is easy to see that
Tn → T weakly.
 39

4.2 Adjoint
Definition If T ∈ L(X, Y ), where X, Y are Banach spaces, the adjoint,
T 0 ∈ L(Y ∗ , X ∗ ) defined by:
T 0 (l) = l(T x) or < l, T x >=< T 0 l, x >
Theorem 4.2.1 Let X, Y be Banach spaces and T ∈ L(X, Y ). Then, the map
given by T 7→ T 0 is a linear, isometric isomorphism.
Proof
kT kL(X,Y ) = sup kT xkY = sup sup |hl, T xi|
kxkX ≤1 kxkX ≤1 klkY ∗ ≤1

= sup sup |hT 0 l, xi|

klkY ∗ ≤1 kxkX ≤1

= sup kT 0 lkX ∗
klkY ∗ ≤1

= kT 0 kL(X,Y )
This shows the isometry part. Linearity and isomorphism are both trivial.
If H is a Hilbert space, and C is the canonical isomorphism taking H to H ∗ ,
we define the Hilbert space adjoint of T ∈ L(H) as T ∗ = C −1 T 0 C where T 0 is
the Banach space adjoint. With this association, T ∗ ∈ L(H). Equivalently, we
can write this relation in the more familiar manner:
∀x, y ∈ H, hx, T yi = hT ∗ x, yi .
It follows that kT k = kT ∗ k. In fact, we have the following properties:
∗
• T 7→ T ∗ is an isomorphism with (αT ) = αT ∗ .
∗
• (T S) = S ∗ T ∗ .

∗ −1
• T −1 = (T ∗ ) .
The map T 7→ T ∗ is continuous in the uniform and weak topologies, but not in
the strong.
Counterexample Shift in l2 :
Wn : (u1 , u2 , . . .) 7→ (0, . . . , 0, u1 , u2 , . . .)
| {z }
n times

So what is the adjoint of Wn ?

∞
X
hv, Wn ui = vn+i ui = hVn v, ui
i=1

with Vn (v1 , v2 , . . .) = (vn+1 , vn+2 , . . .). Thus, Wn∗ = Vn . Vn −→ 0 in the strong

topology, but Wn = Vn∗ 6→ 0 strongly.
40

Note: kT ∗ T kL(H) = kT k2L(H) .

Definition

• An operator T ∈ L(H) is self-adjoint if T ∗ = T.

• An operator P is a projection if P 2 = P.

• A projection P is orthogonal if P ∗ = P.

4.3 Spectrum
Definition Let X be a Banach space, T ∈ L(X).

• The resolvent set of T, denoted ρ(T ) is the set of scalars λ ∈ R (or C) s.t.
λI − T is bijective with a bounded inverse.

• If λ ∈ ρ(T ), then Rλ (T ) = (λI − T )−1 is called the resolvent of T (at λ).

• If λ ∈
/ ρ(T ), then λ is in the ”spectrum of T ” = σ(T ).

Note: From the Open Mapping Theorem, if λI − T is bijective, then its inverse
is continuous

Definition

1. λ ∈ σ(T ) is said to be an eigenvalue of T if ker (λI − T ) 6= {0}

OR λI − T is not injective
OR ∃x 6= 0 such that T x = λx. If this is the case, we say that x is an
eigenvector .

The set of eigenvalues is called the point spectrum of T.

2. λ ∈ σ(T ) which is not an eigenvalue and for which R(λI − T ) is not dense
is said to be in the residual spectrum of T.

In fact, we can draw the following diagram to describe the relationship among
the various parts of the spectrum.
Point Spectrum = injectivity violated
%
Bijectivity violated = Spectrum −→ Residual Spectrum = injectivity OK, surjectivity too violated
&
other = injectivity OK, surjectivity slightly violated

Note: In infinite dimensions, injective 6⇒ bijective since there’s no pigeonhole

principle.
 41

Theorem 4.3.1 Let X be a Banach space and T ∈ L(X). Then, ρ(T ) is open,
and Rλ (T ) = (λI − T )−1 is an L(X)-valued analytic function of λ on ρ(T ).
Moreover, ∀λ, µ ∈ ρ(T ), Rλ (T ) and Rµ (T ) commute and

Rλ (T ) − Rµ (T ) = (µ − λ)Rλ (T )Rµ (T ).

Proof Let λ0 ∈ ρ(T ). Formally, if T were to be taken as a real number, we

could write:
1 1 1
= =  
λ−T λ0 − T + λ − λ 0 (λ0 − T ) 1 + (λ−λ0 )
(λ0 −T )
∞
1 X (λ0 − λ)n
=
λ0 − T n=0 (λ0 − T )n

Inspired by this calculation, we set

∞
X
R̃λ (T ) = Rλ0 (T ) [Rλ0 (T )]n (λ − λ0 )n .
n=0

This series converges absolutely, since:

∞
X ∞
X
kRλ0 (T )n k|(λ − λ0 )n | ≤ kRλ0 (T )kn |(λ − λ0 )|n
n=0 n=0
 
if |λ − λ0 |kRλ0 (T )k < 1. That is, in B λ0 , kR1λ k , we can define R̃λ (T ) and
0

R̃λ (T )(λI − T ) = (λI − T )R̃λ (T ) = I.

 
Hece, R̃λ (T ) = Rλ (T ) and B λ0 , kR1λ k ⊆ ρ(T ). This proves that ρ(T ) is open
0
and that Rλ (T ) is analytic in λ with coefficients in L(X), since we just wrote
a representation for R̃λ (T ) in this way. Moreover, to show commutativity and
the last identity, note the following:

Rλ (T ) − Rµ (T ) = Rλ (T ) (µI − T )Rµ (T ) − Rλ (T )(λI − T ) Rµ (T )

| {z } | {z }
=I =I
= (µ − λ)Rλ (T )Rµ (T ).

Similarly, Rλ (T ) − Rµ (T ) = −(Rµ (T ) − Rλ (T )) = (µ − λ)Rµ (T )Rλ (T ). This

shows that Rλ (T )Rµ (T ) = Rµ (T )Rλ (T ).

Theorem 4.3.2 Let X be a Banach space and T ∈ L(X). Then, σ(T ) is closed,
non-empty and included in B(0, kT kL(X) )

Remark This shows that the spectrum is a non-empty compact subset of a

disk.
42

Proof

• Formally, for any λ, we can write:

∞
!
1 1 1 1 X Tn
= T
= .
λI − T λ 1− λ
λ n=0 λn

If |λ| > kT kL(X) , then λ1 n T n /λn converges absolutely and provides

P
and inverse to λI − T (one can just check by multiplying on right and
left to get the identity). Hence, if λ > kT kL(X) , then λ ∈ ρ(T ) and
Rλ (T ) = λ1 n T n /λn . Hence, σ(T ) ⊂ B(0, kT kL(X) ).
P

• The fact that the spectrum is closed is clear from the previous theorem.

• If σ(T ) were empty, then Rλ (T ) would be an analytic function on C and

lim|λ|→∞ Rλ (T ) = 0. Hence, Rλ must be constant in λ (by Liouville’s
Theorem). Hence, ∀λ, Rλ (T ) = 0. This is a contradiction. Hence,
σ(T ) 6= ∅.

Definition The spectral radius of T , r(T ), is defined as:

r(T ) = sup |λ|.

λ∈σ(T )

We know that r(T ) ≤ kT kL(X) .

Proposition 4.3.3
1/n
r(T ) = lim kT n kL(X)
n→∞

If A is self-adjoint (on a Hilbert Space), then r(A) = kAkL(H) .

1/n
Proof We admit that lim kT n kL(X) exists.
∞
1 X Tn 1
Rλ (T ) = Think of this as a series in z = .
λ n=0 λn λ
∞
X
= z T nzn
n=0

The radius of convergence is given by:

1 1
= .
lim sup kT n k1/n lim kT n k1/n

This is called Hadamard’s Formula . So, for

 
1 1
 <
 λ  lim kT n k1/n ,
 43

Rλ (T ) converges. Hence, ∀λ > limn→∞ kT n k1/n , λ ∈ ρ(T ). Hence,

n 1/n
r(T ) ≤ limn→∞ kT k .
Conversely, if |λ| > r(T ), that means λ ∈ ρ(T ) and Rλ (T ) is analytic there.
=⇒ λ1 has to be in the disc of convergence of:
∞
1 X Tn
λ n=0 λn
 
1 1
⇒   ≤ .
λ lim kT n k1/n
Hence, |λ| ≥ limn→∞ kT n k1/n . Therefore, r(T ) ≥ lim kT n k1/n . We conclude
that:
r(T ) = lim kT n k1/n .
n→∞
In the case of a self-adjoint operator A on a Hilbert space H,
kA2 kL(H) = kA∗ AkL(H) = kAk2L(H) (Check that this is indeed the case!). Then,
kA2n kL(H) = kAk2n
L(H) . Thus, r(A) = kAk.

Example of the Shift Operator Consider T : l1 → l1 given by:

T (u1 , u2 , . . .) = (u2 , u3 , . . .).

Its adjoint from l∞ → l∞ is given by T 0 (u1 , u2 , . . .) = (0, u1 , u2 , . . .).

• Point Spectrum of T : T u = λu. For |λ| < 1, define uλ = (1, λ, λ2 , . . .).

Then, uλ ∈ l1 . So,
T uλ = λuλ .
Hence, {|λ| < 1} ⊆ σ(T ) and kT k = kT 0 k = 1. Therefore, σ(T ) ⊂ B(0, 1).

But, what happens for |λ| = 1? Then, if we solve T u = λu = u, we get

that |u1 | = |u2 | = |u3 |, = . . . . But, this means that either ui = 0 ∀i, or
u∈/ l1 . Thus, |λ| = 1 is not in the point spectrum.
• T 0 has no point spectrum: T 0 u = λu gives:

λu1 = 0
λu2 = u1
·
·

This means that u1 = u2 = . . . = 0.

• If λ is in the point spectrum of T, then Ran (λI − T 0 ) is not dense:
Take f ∈ (l1 )∗ :

< (λI − T 0 )(f ), x >=< f, (λI − T )x > .

| {z }
∈(l1 )∗ =l∞
44

Let |λ| < 1 and apply this to x = uλ = (1, λ, λ2 , · · · ). We see that

< (λI − T 0 )(f ), uλ >= 0 ∀f ∈ (l1 )∗ . From this, we deduce that
Ran (λI − T 0 ) is not dense, for if it were, every L ∈ (l1 )∗ could be approx-
imated by functions of the form (λI − T 0 )fn where fn ∈ (l1 )∗ , leading to
the conclusion that < L, uλ >= 0 ∀L ∈ (l1 )∗ =⇒ by Hahn Banach that
uλ = 0, a clear contradiction.

• λ ∈ residual spectrum of T =⇒ λ ∈ point spectrum of T 0 : Suppose

λ ∈ residual spectrum of T.
=⇒ Ran (λI − T ) is not dense.
=⇒ ∃f ∈ (l1 )∗ s.t. < f, (λI − T )x >= 0 ∀x.
=⇒ < (λI − T 0 )(f ), x >= 0 ∀x.
=⇒ λ is an eigenvalue of T 0 .

• If |λ| = 1 then λ ∈ residual spectrum of T 0 : Take |λ| = 1. Then

2
the element, c = (1, λ, λ , . . .) ∈ l∞ . We will show that B(c, 21 ) does not
intersect Ran (λI − T 0 ). So, the range is not dense and thus, λ ∈ residual
spectrum of T 0 .
Assume d ∈ B(C, 12 ) and ∃e ∈ l∞ such that d = (λI − T 0 )e. Then,

d1 = λe1
d2 = λe2 − e1
d3 = λe3 − e2
·
·
n+1 Pn k
More generally, we can write: en = λ k=1 λ dk . We just need to
k
check that this is not in l∞ . First, note that λ ck = 1 for all k. Also, note
that |dk − ck | < 1/2 since d ∈ B(c, 21 ). Therefore, since |λ| = 1, we get
that

1/2 > |λk dk − λk ck | = |λk dk − 1| =⇒ <(λk dk ) ≥ 1/2

n
!
X
k
=⇒ < λ dk ≥ n/2 =⇒ |en | ≥ n/2 =⇒ e ∈ / l∞
k=1

But, this is a contradiction. Hence, B(c, 12 ) does not intersect

Ran (λI − T ), and λ ∈ residual spectrum of T 0 .
0

Summary of Results for Shift Operator on l1 :

Spectrum Point Spectrum Residual Spectrum
T |λ| ≤ 1 |λ| < 1 ∅
T0 |λ| ≤ 1 ∅ |λ| ≤ 1
 45

In general, for any T ∈ L(X, Y ) for any Banach spaces X, Y, we have the
following:

Proposition 4.3.4

1. If λ ∈ Residual spectrum of T, then λ is in the point spectrum of T 0 .

2. If λ ∈ point spectrum of T, then λ ∈ point spectrum of T 0 or λ ∈ Residual

spectrum of T 0 .

Theorem 4.3.5 Let H be a Hilbert space and A ∈ L(H) be self-adjoint. Then,

1. A has no residual spectrum.

2. σ(A) ⊆ R.

3. Eigenvectors corresponding to different eigenvalues are orthogonal.

Proof

1. If λ were in the residual spectrum of A, then λ would be in the point spec-

trum of A∗ = A. But, this is a contradiction since the residual spectrum
and the point spectrum are disjoint.

2. kAx − (λ + iµ)xk2 = kAx − λxk2 + µ2 kxk2 + 2<< Ax − λx, iµx >. Now,
< Ax − λx, iµx >= iµ < Ax, x > −iλµkxk2 = imaginary
since < Ax, x > =< x, Ax >=< Ax, x >, thus showing that < Ax, x > is
real. Hence,
kAx − (λ + iµ)xk2 ≥ µ2 kxk2 . (4.1)

So, assume µ 6= 0. We will show that λ + iµ ∈ ρ(T ). If µ 6= 0, then, we

deduce that A − (λ + iµ)I is one-to-one. Therefore, λ + iµ is not in the
point spectrum. So, now we will check that Ran (A − (λ + iµ)I) is closed.
Suppose that yn = Axn − (λ + iµ)xn −→ y. Since {yn }n is a Cauchy
sequence, we apply the inequality, 4.1 to get that:

kyn − ym k2 ≥ µ2 kxn − xm k2 ,

showing that {xn }n is also Cauchy, hence ∃x such that xn −→ x. More-

over, by continuity, (A − (λ + iµ))xn −→ (A − (λ + iµ))x. Hence,
y = (A − (λ + iµ))x and is thus in Ran (A − (λ + iµ)I).
If Ran (A − (λ + iµ)I) were not dense, then λ + iµ would be in the residual
spectrum of A. But, A has no residual spectrum. Hence,
Ran (A − (λ + iµ)I) is dense and closed. Therefore, it must be that
Ran (A − (λ + iµ)I) = H =⇒ A − (λ + iµ)I is onto. Therefore,
it is invertible since we showed earlier that it is one-to-one. Therefore,
(λ + iµ) ∈ ρ(T ). Therefore, λ + iµ is in the spectrum only if µ = 0, as
desired.
46

4.4 Positive Operators and Polar Decomposition

(In a Hilbert Space)
Definition A ∈ L(H) is said to be positive if for every x, < Ax, x > ≥ 0. We
write A ≥ 0. Also, A ≥ B means that A − B ≥ 0.
Proposition 4.4.1 Every positive operator on a complex Hilbert space is self-
adjoint.
Proof < Ax, x > is real if A is positive. Hence,
< Ax, x >= < Ax, x > =< x, Ax > .
Now, ∀x, y ∈ H, this means that:
< x+y, A(x+y) >=< A(x+y), x+y > < x−y, A(x−y) >=< A(x−y), x−y >
Subtracting accordingly, we get that < x, Ay >=< Ax, y > .
Note: ∀A ∈ L(H), A∗ A ≥ 0 since < x, A∗ Ax >=< Ax, Ax >= kAxk2 ≥ 0.
Proposition 4.4.2 (Existence of Square Roots) Let A be a positive oper-
ator in L(H). Then, ∃ a unique positive operator B such that A = B 2
√ p
Proof By scaling, reduce to kI √ − Ak < 1. Compute A = I − (I − A)
through the series expansion of 1 − z :
∞
√ X f (n) (0) n
f (z) = 1−z = z ,
n=0
n!
√
with f (n) (0) ≥ 0 ∀n. Hence A is positive.
√
Definition |A| = A∗ A (in L(H)).
Definition U ∈ L(H) is an isometry if kU xk = kxk ∀x ∈ H. It is a partial
isometry if it is an isometry restricted to (ker U )⊥ .

Proposition 4.4.3 Let U be a partial isometry. Then, U ∗ U = P  is an

 (ker U )⊥

orthogonal projection on (ker U )⊥ and U U ∗ = P  .


Ran U
Conversely, if U satisfies these properties, then U is a partial isometry.
Theorem 4.4.4 (Polar Decomposition) Let A ∈ L(H). Then, there exists
a partial isometry U such that A = U |A|. This U is uniquely determined by the
requirement ker U = ker A. Moreover, Ran U = Ran A.
Example
A = right shift in l2
A∗ = left shift in l2 .
A∗ A = I =⇒ |A| = I.
In the polar decomposition, A = U (|A|) = U. So, we see that U = A is not an
isometry since A is not invertible.
Chapter 5

Compact and Fredholm

Operators

5.1 Definitions and Basic Properties

Definition Let X and Y be Banach spaces. T ∈ L(X, Y ) is said to be compact
if T (BX ) is compact.

⇐⇒ T maps bounded sets into precompact sets (i.e. sets with compact clo-
sure). ⇐⇒ T maps bounded sequences into sequences which have convergent
subsequences.

Proposition 5.1.1 If xn * x, then T (xn ) → T (x) strongly in Y.

Proof If xn * x then {xn }n is bounded. Hence, {T (xn )}n has a conver-

gent subsequence that converges to some y ∈ Y. Since T is continuous in the
strong-strong topology, it is also continuous in the weak-weak topology. Hence,
T (x) = y and T (xnk ) → T (x). A sequence whose every convergent subse-
quence converges to T (x) and which is bounded, converges to T (x). Hence,
T (xn ) → T (x).

Definition T ∈ L(X, Y ) is said to be an operator of finite rank if dim Ran (T ) <

∞.

Remark Finite rank operators are obviously, compact (since a closed and
bounded subset of a finite-dimensional space is compact).

Proposition 5.1.2

1. If Tn are compact operators in L(X, Y ) and Tn → T in the L(X, Y )−norm,

then T is compact.
2. T is compact =⇒ T 0 is compact.

47
48

3. If T is compact and S is bounded, then T ◦ S and S ◦ T are compact.

Proof

1. For  > 0, n ≥ N, kTn − T kL(X,Y ) < . Since TN is compact, TN (BX ) is

precompact. Hence, it can be covered by a finite number of balls of radius
. So, [
TN (BX ) ⊂ B(y; ).
finite

But, ∀x ∈ BX , kTN (x) − T (x)k < . Therefore,

[
T (BX ) ⊆ B(y, 2).
finite

Since this is true ∀ > 0, T (BX ) is precompact.

2. In homework (Due 11/19).
3. Since S bounded, S(BX ) is bounded. Since T compact, therefore, T (S(BX ))
is precompact. Hence, T ◦ S is compact. On the other hand, if T compact,
then T (BX ) is precompact. Since S is bounded, S(T (BX )) is precompact
as well by continuity. Hence, S ◦ T is compact.

Note: This theorem shows that limits of finite rank operators are compact!

Conversely: Can any compact operator be approximated by finite rank oper-

ators? Not always. Yes if we’re in a Hilbert space:

Theorem 5.1.3 Let H be a Hilbert space and T ∈ L(H) compact. Then, T is

the uniform limit of finite rank operators.

Proof Let K = T (BH ), compact. Given  > 0, there exists a covering of K:

[
K⊂ B(y, ).
finite

Let Y be the space spanned by the yi0 s. dim Y < ∞. Let PY be the orthogonal
projection onto Y. Take T = PY ◦ T. Let x ∈ BH . Therefore, ∃io such that
kT x − yi0 k < . By projection,

kPY T x − PY (yi0 )k <  =⇒ kT x − yi0 k < 

=⇒ kT x − T xk < 2
=⇒ kT − T kL(H) < 2

Important Example (Kernel
of Integral Operator)

Let X = C 0 ([0, 1]), k · k∞ . K(x, y) ∈ C 0 ([0, 1] × [0, 1]). For all f ∈ X, define:
Z 1
TK f (x) = K(x, y)f (y) dy.
0
 49

Proposition 5.1.4 For each K defined as above, TK is a compact operator.

R1
Proof Say f ∈ BX , kf k∞ ≤ 1. Thus, |TK f (x)| < 0 |K(x, y)||f (y) dy ≤ kKk∞ ,
independent of f. Hence, TK ∈ L(X), with kTK k ≤ kKk∞ .
To show that TK (BX ) is precompact, we will use Ascoli’s Theorem , which
says that if a uniformly bounded family is equicontinuous, every subsequence
has a limit point. So, all remains to show is that TK f are an equicontinuous
family. First, ∀f ∈ BX , kTK (f )k ≤ kKk∞ . Hence, TK f are uniformly bounded.
Remains to show equicontinuity.
K ∈ C 0 ([0, 1] × [0, 1]). Since K is a continuous function on a compact set,
it is uniformly continuous on that set. Therefore, ∀ > 0, ∃δ > 0 such that
∀x, y ∈ [0, 1], with |x − x0 | < δ, |K(x, y) − K(x0 , y)| < . Then,
Z 1
|TK f (x) − Tk f (x0 )| < |K(x, y) − K(x0 , y)||f (y)| dy < 
0

for any f ∈ BX . Hence, we have shown equicontinuity of the family. Then, the
conclusion of Ascoli’s Theorem gives us that TK (BX ) is precompact.

5.2 Riesz-Fredholm Theory

Lemma 5.2.1 (Riesz) Let X be a Banach space and M ⊆ X, a closed lin-
ear subspace of X, M 6= X. Then, ∀ > 0, there exists kxk = 1 such that
dist(x, M ) ≥ 1 − .

Proof Take x ∈ X \ M and let d = dist(x, M ) 6= 0 (since M is closed).

d x−y
Therefore, ∃y ∈ M such that kx − yk < 1− . Take v = kx−yk . Now, we want to
calculate dist(v, M ).
∀m ∈ M,
∈M
z }| {
kx − (y + kx − ykm) k dist(x, M )
kv − mk = ≥ ≥ 1 − .
kx − yk kx − yk

So, dist(v, M ) ≥ 1 − . Hence, v is the one we want.

Definition Let X be a Banach space and Y be a subspace of X. Then, Y ⊥ is

the subspace of X ∗ defined by:

Y ⊥ = {f ∈ X ∗ : ∀y ∈ Y, f (y) = 0}

Remark Y ⊥ is always closed. If X = X ∗ , then (Y ⊥ )⊥ = Y . If things are

closed, (ker T 0 )⊥ = Ran T, (Ran T 0 )⊥ = ker T.

Definition codim Y = dim Y ⊥

50

Theorem 5.2.2 (Fredholm Alternative) Let T ∈ L(X) be a compact oper-

ator on a Banach space. Then,
1. ker(I − T ) is finite dimensional.
⊥
2. Ran (I − T ) is closed and = (ker(I − T 0 ))
3. ker(I − T ) = {0} ⇐⇒ Ran (I − T ) = X.
4. dim ker(I − T ) = dim ker(I − T 0 ).

The Alternative Let A = I − T. Either, “ker A = {0} and Ran A = X” OR

“ker A 6= {0} and Ran A 6= X”. In other words, either, Ax = b has a unique
solution or Ax = 0 has non-trivial solutions.

Motivation T is compact (think of an integral operator). Want to solve,

T ϕ − ϕ = f. This either has solutions ∀f or T ϕ = ϕ has non-trivial solutions.

Example

• Let ϕ0 − ϕ00 = f, ϕ(0) = ϕ(1) = 0.

• Or in PDE: 4ϕ − ϕ = f.

Proof of Fredholm Alternative

1. Let N = ker(I − T ). Then, ∀x ∈ N , T (x) = x. BN = T (BN ) ⊂ T (BX ).

Hence, BN is precompact. Recall the theorem of Riesz that BN compact
⇐⇒ dim N < ∞.
2. Ran (I − T ) is closed: Let fn = xn − T (xn ), fn → f. Is f ∈ Ran (I − T )?
Since by the above, ker (I − T ) is finite-dimensional, hence closed. Hence,
dn = dist(xn , ker (I − T )) is achieved. So, ∃vn ∈ ker (I − T ) such that
dn = kxn − vn k, vn = T vn . So, we can write:

fn = xn − vn − T (xn − vn ) (5.1)

If kxn − vn k is bounded, then by compactness of T, we can assume (up

to extraction), that T (xn − vn ) → l. So, pass to the limit in Eqn. 5.1, to
obtain xn − vn → l + f. Again, passing to the limit in Eqn. 5.1, we get
that f = l + f − T (l + f ) =⇒ f = (I − T )(l + f ). So, f ∈ Ran (I − T ).
All that remains to do is to check that {kxn − vn k}n is bounded.
Suppose not. Then, divide the quantity in Eqn. 5.1 by kxn − vn k. Then,
 
xn − vn xn − vn
−T →0 (5.2)
kxn − vn k kxn − vn k
| {z }
≡un

{un }n is certainly bounded. By compactness of T we can assume that

T (un ) → z. From Eqn. 5.2, un → z. Therefore, by uniqueness of limits, we
 51

can conclude that T (z) = z. Hence, z ∈ ker (I − T ). But,

dist(xn , ker (I − T )) = kxn − vn k. Hence, dist(un , ker (I − T )) = 1. But,
this contradicts the fact that un → z ∈ Ker(I − T ). This proves that
kxn − vn k is bounded. Hence, we’re done.
3. ker (I − T ) = {0} ⇐⇒ Ran (I − T ) = X :
(=⇒) Assume not. In other words, ∃x ∈ X \ Ran (I − T ). Let
X1 = Ran (I − T ). It is closed. Therefore, it is a Banach space. Also,
T (X1 ) ⊆ X1 since if y = x − T (x), then

T (y) = T (x) − T 2 (x) = (I − T )(T (x)) ∈ Ran (I − T ) = X1 .

 
Now, consider T X Then, let X2 = Ran (I − T X ). Inductively, let
1 1
Xn = (I − T )n (X). Then, Xn ( Xn+1 (Why?) If Xn = Xn−1 , then
Ran (I − T )n = Ran (I − T )n−1 . So, applying this to x, we see that
(I − T )n−1 x = (I − T )n y for some y. But, I − T is injective. So,
(I − T )y = x ∈ Ran(I − T ). This is a contradiction of the original as-
sumption that x ∈ X \ Ran (I − T ).
Now, apply Riesz’ Lemma (Lemma 5.2.1) and we find a xn ∈ Xn such
that kxn k = 1 and dist(xn , xn + 1) ≥ 1/2. Now, consider xn , xm m < n.
Then,
T xn − T xm = (T − I)xn − (T − I)xm + xn − xm .
| {z } | {z } |{z} |{z}
∈Xn+1 ∈Xm+1 ∈Xn ∈Xm

So, kT xn − T xm k > dist(xm , Xm+1 ) ≥ 1/2. Hence, {T xm }m is not a

Cauchy sequence, which contradicts the fact that {xn }n is bounded and
T is compact. Therefore, Ran (I − T ) = X.

(⇐=) If Ran (I − T ) = X, then, ker (I − T 0 ) = {0}. So, apply the (⇒)

direction to T 0 , which is also compact. This gives that Ran (I − T 0 ) = X ∗ .
Hence, ker (I − T ) = {0}.
4. dim ker (I − T ) = dim ker (I − T 0 ) (Check as an exercise!)

5.3 Fredholm Operators

Definition A Fredholm Operator is an operator A ∈ L(X, Y ) such that:
• ker (A) is finite-dimensional
• Ran (A) is closed and has finite codimension (codim Ran (A) = dim (Ran (A))⊥ ).
The index of A is given by:

Ind (A) = dim (ker (A)) − codim (Ran (A))

Example From Riesz-Fredholm Theorem (by parts (1), (2), and (4)) of Fred-
holm Alternative), if T is compact, then I − T is Fredholm of index 0.
52

Theorem 5.3.1

1. The set of Fred (X, Y ) is open in L(X, Y ) and A 7→ Ind A is continuous,

and therefore constant on each connected component of Fred (X, Y ).

2. Every Fredholm operator is invertible modulo finite rank operators.

∃B ∈ L(X, Y ) such that BA − IX and AB − IY have finite rank. Con-
versely, if A ∈ L(X, Y ) is an operator such that ∃B ∈ L(X, Y ) with AB−I
and BA − I compact, then A is Fredholm.

3. If A is Fredholm and T compact, then A + T is Fredholm and

Ind (A + T ) = Ind A.

4. If A and B are Fredholm, then AB is also and Ind (AB) = Ind A + Ind B.

5. If A is Fredholm, then A0 is Fredholm and Ind A0 = −Ind A.

Example

• Right-shift in lp : Consider the operator, A : (u1 , u2 , . . .) 7→ (0, u1 , u2 , . . .).

Then, ker A = {0}. Also, Ran A = {(ui )i : u1 = 0}. It is closed and
codim Ran A = 1. Hence, Ind A = −1.

• Lef-shift in lp : Consider the operator A : (u1 , u2 , . . .) 7→ (u2 , u3 , . . .).

Then, ker A = {(u, 0, 0, . . .) : u ∈ R}. Hence, dim ker A = 1 and
Ran A = lp . Hence, codim Ran A = 0. Therefore, Ind A = 1.

• Erasure in lp : Consider the operator A : (u1 , u2 , . . .) 7→ (0, u2 , u3 , . . .).

Then, ker A = {(u, 0, 0, . . .) : u ∈ R}. =⇒ dim ker A = 1. Also,
Ran A = {(un )n : u1 = 0}. It follows, then, that codim Ran A = 1
and Ind A = 0.

5.4 Spectrum of Compact Operators

Theorem 5.4.1 (Riesz-Schauder) Let T ∈ L(X) be a compact operator and
dim X = ∞. Then, the following hold:

• 0 ∈ σ(T )

• σ(T ) \ {0} consists of eigenvalues of finite multiplicity (i.e. the dimension

of the λ-eigenspace (ker (T − λI)) has finite dimension ∀λ ∈ σ(T ) \ {0}.

• σ(T ) \ {0} is either empty, finite or a sequence converging to 0 (i.e. it is

a discrete set with no limits other than 0).
 53

Proof
• If 0 ∈ / σ(T ) then T is invertible (i.e.: ker T = {0}). Therefore,
T · T −1 = I. =⇒ since T and T −1 are compact, that I is compact.
Hence, dim X < ∞ (by Riesz’ Theorem), a contradiction! This shows
that in infinite dimension, a compact operator is never invert-
ible.
• From Riesz-Fredholm Theorem (i.e. Fredholm Alternative), if
λ ∈ σ(T ) \ {0} then since λ 6= 0, if ker (I − Tλ ) = {0}, then
Ran I − Tλ = X (since T compact =⇒ Tλ compact). This would show

that I − Tλ is invertible, a clear contradiction. So, ker I − Tλ 6= {0}.

Hence, λ is an eigenvalue. Moreover, dim(ker I − Tλ ) < ∞ by Riesz-

Fredholm.
• Suppose to the contrary, that ∃ a sequence of non-zero eigenvalues,
λn → λ 6= 0. Each λn is an eigenvalue, so take en to be an eigenvector.
The en are
Pnlinearly independent P (to see this, by induction P assume that
n n
en+1 = i=1 α i e i . Then, λ n+1 ( i=1 α i ei ) = T (en+1 ) = i=1 λi αi ei .
Hence, since e1 , . . . , en are linearly independent, λi αi = λn+1 αi for each i.
But, we assumed that λn 6= λn+1 we get that αi = 0, a contradiction). So,
let Xn = Span(e1 , . . . , en ). Xn 6⊆ Xn+1 . Moreover, (T − λn I)Xn ⊂ Xn−1 .
By Riesz’ Lemma, take a sequence un ∈ Xn , such that for each n, kun k = 1
and dist(un , Xn−1 ) ≥ 1/2. Then,


T un T u m λ n u n T u m − λ m u m

− = T un − − + u − u m = ?.

n

λn λm λn λm |{z}
∈Xn−1

So, take n > m =⇒ m ≤ n − 1 =⇒ Xm ⊆ Xn−1

T u n − λn un
=⇒ ∈ (T − λn I)Xn ⊂ Xn−1
λn
T u m − λm u m
∈ Xm ⊂ Xn−1
λm
T u n − λn un T u m − λm u m
=⇒ − ∈ Xn−1
λn λm
=⇒ ? = kun − |{z} x̃ k > 1/2
∈Xn−1

If λn , λm → λ 6= 0, this contradicts the fact that T un is a Cauchy sequence.

But, kun k = 1 and T is compact. This is a contradiction. Therefore,
λ = 0.
Remark Conversely, if αn → 0, one can build a compact operator whose spec-
trum is exactly that sequence. For example, consider l2 and take
{un }n 7→ {αn un }n . This can be approximated by finite rank operators. Hence,
it is compact.
54

5.5 Spectral Decomposition of Compact, Self-

Adjoint Operators in Hilbert Space
Proposition 5.5.1 Let T ∈ L(H) be a self-adjoint operator on a Hilbert space
(recall that self-adjoint operators in Hilbert space have real spectrum... i.e.:
σ(T ) ⊂ R). Then if we define:
M = sup < x, T x >, m = inf < x, T x >
kxk=1 kxk=1

Then, σ(T ) ⊂ [m, M ] with m, M ∈ σ(T ).

Proof Private Exercise!
Corollary 5.5.2 If T is self-adjoint and σ(T ) = {0} then T = 0.
Proof If σ(T ) = 0, then m = M = 0, in the notation of the preceding propo-
sition. Then, ∀x, < x, T x >= 0. So, polarize to get < x, T y >= 0, ∀x, y ∈ H.
Hence, T ≡ 0.
Theorem 5.5.3 (Hilbert-Schmidt Theorem) Let T be a compact self-adjoint
operator on a Hilbert space. Then, ∃ a complete orthonormal basis of H formed
of eigenvectors such that
T ϕn = λn ϕn ∀n.
(If H is separable, then you can find a countable basis. If H is not, then there
is possibly an uncountable basis of ker T ). Also,
lim λn = 0 λ0 = 0, (σ(T ) \ {0} = {λn }n )
n→∞

Proof Take En = ker (T − λn I). Then, by Riesz-Schauder Theorem,

dim En < ∞. If x ∈ En , y ∈ Em , for n 6= m ⇒ < x, y >= 0. This can
be shown by noting that:
λm < x, y >=< x, λm y >=< x, T y >=< T x, y >=< λn x, y >= λn < x, y >
But, λn 6= λm . Hence, < x, y >= 0. So, let M be the sum of the En and ker T.
M is stable under T : T (M ) ⊂ M (since En is space of eigenvectors). Hence,
M ⊥ is also stable under T (take x ∈ M, y ∈ M ⊥ ⇒ < x, y >= 0 ⇒ M is
stable under T ⇒ T x ∈ M ⇒ < x, T y >=< T x, y >= 0 since T is self-adjoint
and T x ∈ M, y ∈ M ⊥ . Hence, T y ∈ M ⊥ ).
So, consider T M ⊥ . It’s also a compact (self-adjoint) operator.

(=⇒) σ(T M ⊥ ) \ {0} is formed by eigenvalues.
(=⇒) in M ⊥ there are eigenvectors  for T, but they are all in M.
(=⇒) The only possibility is σ(T M ⊥ ) = {0}.
So, by previous corollary, T M ⊥ ≡ 0. Hence, M ⊥ ⊂ ker T ⊂ M.
=⇒ M ⊥ = {0} (since M ∩ M ⊥ = {0}).
So, choose, an orthonormal basis in each En (each is finite dimensional) and
an orthonormal basis of ker T. This provides a complete orthonormal family,
which is countable if H is separable.
 55

T can
P∞be approximated by finite rank operators in the following manner: If
x = n=0 xn , xn ∈ En , E0 = ker T.
∞
X
Tx = λn xn
n=0
N
X
TN x = λn xn
n=0
kTN − T kL(H) −→ 0 as N → ∞

Now, we prove a result that essentially says that: “compact operators on a

Hilbert space, can be ‘diagonalized’ over an orthonormal basis”
Theorem 5.5.4 (Canonical form for Compact Operators) Let T be a com-
pact operator in L(H). Then, there exist orthonormal sets (not necessarily com-
plete) {ϕn }n and {ψn }n and a sequence, {λn }n , with λn → 0 such that:
∞
X
T = λn < ψ n , · > ϕ n (SVD)
n=0
√
The λn are eigenvalues of |T | = T ∗ T and are called singular values of T.

Proof T ∗ T is compact, self-adjoint. Call its eigenvalues, µn → 0 and let {ψn }n

be the corresponding orthonormal basis of eigenvectors. Then,
∞ ∞
!
X X
Tψ = T < ψn , ψ > ψn = < ψ n , ψ > T ψn .
n=0 n=0

T ψn √
Let ϕn = λn where λn = µn .
∞
X
=⇒ T ψ = < ψ n , ψ > λ n ϕn .
n=0

Check that that the ϕn defined this way are indeed orthonormal. But, this is
clear since the µn are the eigenvalues of T ∗ T.
56
Appendix A

Definition

• Let X be a metric space. A family of functions, {fα }α defined on a subset

U ⊂ X is said to be uniformly bounded if ∃C > 0 such that:

sup |fα (x)| ≤ C.

α
x∈U

• Let X be a metric space. A family of functions, {fα }α defined on a subset

U ⊂ X is said to be equicontinuous if ∀ > 0 there exists δ > 0 such that

dist(x, y) < δ =⇒ sup |fα (x) − fα (y)| ≤ C,

α

for all x, y ∈ U.

Theorem A.0.5 (Ascoli’s Lemma) Let K be a uniformly bounded, equicon-

tinuous, family of functions on a compact metric space X. Then, any sequence
contains a subsequence that is uniformly convergent in X to a continuous func-
tion.

Corollary A.0.6 Let X be a compact metric space. A family K of functions in

X ∗ is precompact if and only if K is both uniformly bounded and equicontinuous.

57
58
Index

H01 , 34 distribution
X ⊥ , 49 tempered, 24
dual
absorbing, 23 double dual, 15
adjoint, 39 norm, 3
Hilbert space, 39 space, 3
Ascoli’s Theorem, 49
eigenvalue, 40
Baire Category Theorem, 13–16 eigenvector, 40
balanced energy minimizer, 33
circled, 23 epigraph, 9
Banach-Alaoglu’s Theorem, 28–29 equicontinuous, 57
basis extreme
topological, 21 point, 8
biconjugate, 10 set, 8
bounded
Fenchel-Moreau Theorem, 10
operator, 13
Fenchel-Rockafellar, 12
closed, 21 finite rank operator, 47
Closed Graph Theorem, 19 Frechet space, 24
Fredholm Alternative, 50
codimension, 49
Fredholm Operator, 51
coercive, 35
compact, 21 gauge
compact embedding, 35 Minkowski Functional, 5
compact operator, 47–49 Goldstine’s Lemma, 30
canonical form, 55 graph, 19
conjugate function
Legendre-Fenchel transform, 9 Hadamard’s Formula, 42
continuous embedding, 35 Hahn-Banach Theorem, 1–9
continuous function, 21 Complex Version, 3
at a point, 3 Real Version, 1
convex Hausdorff, 21
function, 9 Hausdorff Maximality Theorem, 8
set, 5 Helly’s Lemma, 29–30
convex hull, 8 Hilbert-Schmidt Theorem, 54
closed, 8 hyperplane, 5
critical exponent, 35 separates, 5

59
60

separates strictly, 5 point spectrum, 40

residual, 40
index, 51 square root
integral operator, 48 of an operator, 46
isometry SVD, 55
on a Hilbert space, 46
partial, 46 topological space, 21
isomorphism, 14 topology
discrete, 22
Kakutani’s Theorem, 29–31 indescrete, 22
Krein-Milman Theorem, 8 strong operator, 37
uniform, 37
locally convex space, 22 weak, 22
lower semi-continuous, 9 weak operator, 37
weak-*, 28
metrizable, 24, 32 total ordering, 1
neighborhood, 21 Uniform Boundedness Principle
Banach-Steinhaus Theorem, 16
Open Mapping Theorem, 18 upper bound
orthogonal, 40 for totally ordered sets, 1
partial ordering, 1 Zorn’s Lemma, 2
polar decomposition, 46
positive operator, 46
precompact, 47
projection, 40

reflexive, 29
resolvent, 40
resolvent set, 40
Riesz Representation Theorem, 15
Riesz’ Lemma, 49
Riesz-Fredholm Theorem
Fredholm Alternative, 50
Riesz-Schauder Theorem, 52

Schwartz Class, 24
self-adjoint, 40
seminorm, 22
separable space, 32
separate points, 22
shift operator, 43–45
singular values, 55
Sobolev Space, 34
spectral radius, 42
spectrum, 40