Notes for Functional Analysis

(CIMAT, Fall 2020)

November 23, 2021

Contents

Notations and conventions 3

1 What is this course about? 4

2 Spaces of functions I: Normed vector spaces 6

2.1 Open mapping and closed graph theorems, Uniform boundedness principle 11
2.2 Hilbert spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

3 Spaces of functions II: Topological vector spaces 30

3.1 Elementary theory of topological vector spaces . . . . . . . . . . . . . . . 30
3.2 Some famous examples of Frechet spaces . . . . . . . . . . . . . . . . . . 36
3.3 An F -space which supports no continuous linear functional . . . . . . . . 39

4 Duality theory I: Hahn-Banach theorem 41

5 Duality theory II: Weak and weak star topologies 48

5.1 Some compactness theorems . . . . . . . . . . . . . . . . . . . . . . . . . 51

6 Duality theory III: Krein-Milman theorem 58

7 Duality theory IV: Summary of results 60

8 Compact and Fredholm operators 65

9 Spectral theory I: Commutative Banach and C* algebras 68

10 Spectral theory II: Bounded normal operators 72

11 Spectral theory III: Unbounded normal operators 74

11.1 Unbounded operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
11.2 Spectral theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
11.3 Closed range theorem for unbounded operators . . . . . . . . . . . . . . . 75

12 Implicit function theorem for Banach spaces 77

13 GNS construction 79

14 Semigroups 80

1
References 81

2
Notations and conventions

• N, Z, Q, R and C are respectively the set of nonnegative integers, integers, rational

numbers, reals and complex numbers. F denotes either the field of real numbers or the
field of complex numbers. Vector spaces are always on F.

• For a multi-index α = (α1 , . . . , αn ) ∈ Nn and a = (a1 , . . . , an ) ∈ Rn , we set |α| = αj ,

P
Q αj ∂ α1 +···+αn
αj !, aα = aj and ∂ α = ∂x
Q
α! = α1 αn , where (x1 , . . . , xn ) is the standard
1 ···∂xn
coordinates on Rn . Also, 0α := 1 and ∂ 0 := id if α = (0, . . . , 0).

• Neighborhoods are always open.

• Topological vector spaces (TVS) are assumed to be Hausdorff. LCTVS stands for
“locally convex topological vector space”.

• “Subspace” always means in the sense of topology, namely a subset with subspace
topology. If the subspace is also closed under the vector space structure we use the
phrase “linear subspace”.

• The domain, kernel (null space) and range of a linear map A are denoted by DomA ,
KerA and RanA .

• C(X), X compact topological space, is the algebra of continuous functions f : X → C,

equipped with the uniform norm kf k := supx∈X |f (x)|. C(X; R) is the real version.

• C k (U ), k ∈ {0, 1, . . . , ∞}, U ⊆ Rn open, is the set of functions f : U → C which has

continuous derivative ∂ α f up to total order k, namely |α| ≤ k.

• C k (U ), k ∈ {0, 1, . . . , ∞}, U ⊆ Rn open, is the set of functions f : U → C which admit

an extension to a C k function on a neighborhood of U . Alternatively, by a classical
theorem of Borel [Lee, page 27], it is exactly the set P of C k (U ) functions f such that
each partial derivative f (α) , α = (α1 , . . . , αn ) ∈ Nn , αj ≤ k, admits a continuous
extension to U .

• Sequence spaces lp , c, c0 , c00 are introduced in Example 1.

• If α is a linear functional on a vector space X and x ∈ X then α(x) is sometimes

denoted by hx, αi.

• The term “operator”, unless otherwise stated, refers to a continuous linear map.

3
Chapter 1

What is this course about?

Mathematical analysis studies the properties (say differentiability, integrability, Fourier

harmonics, etc.) of functions living on topological spaces (say open subsets of Rn ). This
study can be done at two levels:

1. Personal life. To study each function separately. This is called hard analysis.

2. Social life. To study spaces of functions. This is called soft analysis, functional
analysis or infinite dimensional analysis. Almost always these spaces of functions
turn out to be infinite-dimensional vector spaces, and since mathematical analysis
is about limiting phenomena, these vector spaces are equipped with a structure to
measure nearness, namely a norm, a family of norms, a metric or more generally
a topology. If the topology is complete (namely Cauchy sequences converge) then
geometric intuitions can be developed. Sometimes these spaces are even closed un-
der pointwise multiplication, so form an algebra. (For example, the vector space of
continuous complex-valued functions f living on a compact topological space X and
equipped with the uniform norm kf k = supx∈X |f (x)| is an algebra.) This way, func-
tional analysis is a common ground for analysis, topology, geometry and algebra.

In a real problem these two aspects are interwoven. This course emphasizes the soft
aspects but will try to enlighten the connections to hard ones.
Functional analysis is a vast subject. Here are some aspects of it:

• Linear algebra is the study of finite dimensional vector spaces. It is a well-developed

subject with many highlights such as the theory of determinants, Jordan structure
theorem for linear maps, duality theory (for example: Ax = y is solvable if and only
if y is orthogonal to the kernel of A∗ ), Fredholm alternative (A square matrix is ei-
ther invertible or it is neither injective nor surjective.), etc. On a finite dimensional
vector space there is only one Hausdorff topology compatible with the linear structure
(Theorem 2.(1) and Chapter 3.(26)). This is why topological issues does not appear in
linear algebra. However, topology is an important aspect of infinite dimensional anal-
ysis. One mission of functional analysis is to generalize the fundamental theorems of
linear algebra to infinite-dimensional setting. For example, spectral theory generalizes
principal axis theorem of normal matrices (Chapters 10-11); there is a well-developed

4
 duality theory (Chapters 4-7); there are theories for trace and determinant [Lax, chap-
ter 30]; there is a Fredholm alternative (Theorem 61), etc. In the same way, functional
analysis generalizes some fundamental theorems of differential and integral calculus
on finite dimensional Euclidean spaces Rn to infinitely many dimensions. For exam-
ple, the implicit function theorem and the exponential matrix solution x(t) = eAt x(0)
for the first order linear constant-coefficient system of ordinary differential equations
dx/dt = Ax are generalized in Theorems 79 and 82, respectively. Not all theorems of
functional analysis are of this sort. For example, the theory of compact and Fredholm
operators (Chapter 8) has no finite dimensional counterpart.

• Continuous linear maps between topological vector spaces of functions are called op-
erators; their study, the so-called operator theory, is a branch of functional analysis
[Alp, Pea]. To get deep results one is usually directed to study specific classes of opera-
tors, for example compacts and Fredholms (Chapter 8), integral and pseudodifferential
operators ([Grf, chapters 4,6,8][Shu, Tay-PDO]), self-adjoints and normals (Chapters
10-11), isometries and dilations (or contractions) [SFBK], shifts [Nik], Toeplitz oper-
ators [Dou, chapter 7][BS, BG, Upm], Hankel operators [Zhu-OT, chapters 7, 9], etc.
Generally speaking, the structure theory for a class of operators is called the spectral
theory for that class of operators; however, this phrase usually refers to the structure
theory of normal operators on Hilbert spaces (Chapters 9-11).

• The most important slogan of functional analysis is that continuous linear functionals
can be used to detect many notions in functions spaces. For example, they can be used
to check boundedness of subsets, density of linear subspaces, surjectivity and range-
closedness of operators, etc. Theorem 52 gathers many instances of such phenomena.

• One way to grasp the taste of a branch of mathematics is through the deep theorems
proved there. Here is one in operator theory. Beaurling characterized all closed linear
subspaces of l2 (N) (the space of square-summable sequences of complex numbers) which
are invariant under the shift operator S : l2 (N) → l2 (N), (a0 , a1 , . . .) 7→ (0, a0 , a1 , . . .).
The final answer is extremely hard to express without using a particular function theory
specifically developed for this purpose [Rud-RCA, 17.23]. There are very few operators
whose lattice of invariant closed linear subspaces is completely known. It is an open
problem, the so-called invariant subspace problem, whether every operator on a
separable infinite-dimensional Hilbert space has a nontrivial invariant closed linear
subspace.

Chapters 2-7 constitute the core of this course. Later chapters discuss various topics.
We are going to start by introducing several general classes of spaces of functions appear-
ing in mathematical analysis: topological vector spaces, normed vector spaces, Banach
spaces, reflexive Banach spaces, uniformly convex Banach spaces and Hilbert spaces, in
decreasing order of generality. Only for Hilbert spaces one can develop a structure theory.

5
Chapter 2

Spaces of functions I: Normed vector

spaces

References: [Fol, chapter 5][DS, chapter 2][Dou, chapters 1,3][Bre, chapters 2,5].

The simplest spaces of functions which appear in mathematical analysis are those whose
topology is induced by a norm. This chapter is devoted to them.
Let F be either the field of real numbers or the field of complex numbers, and let X
be a vector space over F. X is called a normed vector space if there exists a unary
operation k − k : X → [0, ∞) on X, called a norm, which satisfy:

1. Homogeneity: kaxk = |a|kxk for every a ∈ F and x ∈ X.

2. Triangle inequality: kx + yk ≤ kxk + kyk for every x, y ∈ X.

3. Positivity: kxk = 0 implies x = 0.

Note that the norm is always finite. The triangle inequality shows that the norm is a
continuous function. If we drop the third positivity axiom from the definition of the norm
we have a seminorm. There is a standard procedure to make a norm out of a seminorm:
pass to the quotient X/{x ∈ X : kxk = 0}, and set the norm of the equivalence class of
each x ∈ X equal to the seminorm of x. A normed vector space can be made a metric
space by the distance function d(x, y) := kx − yk. The distance function in turn induces a
topology on X: A basis for the topology is given by the open balls {y ∈ X : d(x, y) < },
x ∈ X,  > 0. Recall that all metric space topologies are Hausdorff and first countable.
When we talk about the topology of a normed vector space we always mean this topology.
Exercise: Two norms k − k1 and k − k2 on a vector space X are equivalent if there
exist C1 , C2 > 0 such that C1 kxk1 ≤ kxk2 ≤ C2 kxk1 for every x ∈ X. Show that:
(1) Equivalent norms induce
P the same ptopology.
P (2) l1 , l2 and l∞ norms on Cm , given
respectively by kzk1 = |zj |, kzk2 = 2
|zj | and kzk∞ = max |zj | for z = (z1 , . . . , zm ),
are all equivalent.
A linear map T : X → Y between normed vector spaces is called an operator if any
of the following equivalent conditions holds:

6
 • T is continuous.
• T is continuous at a point.
• T is bounded in the sense that it maps bounded subsets to bounded ones; equiv-
alently, kT xk ≤ Ckxk for some C > 0 and every x ∈ X. (If this happens then
the infimum of such C is called the operator norm of T , denoted by kT k; by
homogeneity we have kT k = supx6=0 kT xk/kxk = supkxk=1 kT xk = supkxk<1 kT xk.
Here we are putting the trivial case X = {0} aside.)
The set of all (bounded) operators from normed vector space X to Y is denoted by
B(X; Y ). (B, L or L is also used instead of B in the literature.) One can easily check that
B(X; Y ) is itself a normed vector space via the operator norm. B(X; X) is abbreviated
into B(X). The good news that B(X) is an algebra over F under composition of operators.
B(X; F) is abbreviated into X ∗ , the so-called dual space of X. Elements of X ∗ are called
continuous linear functionals. A linear functional on X is just a linear map X → F.
The transpose of an operator T : X → Y is given by T ∗ : Y ∗ → X ∗ , β 7→ β ◦ T . T
is an isometry if kT xk = kxk for every x ∈ X. An isometric isomorphism between
normed vector spaces is an isometry which is also a homeomorphism. A Banach space
is a normed vector space which is complete in the sense that Cauchy sequences converge.
Completeness can be stated
P equivalentlyP by requiring that absolutely convergent series
are convergent, namely xj converges if kxj k < ∞. (Exercise: Prove this equivalence.
Hint. Suppose that absolutely convergent series are convergent and let xj be a Cauchy
sequence. One can find subsequence jk such that kxjk − xjk+1 k < 2−k for each k. Note
that a Cauchy sequence is convergent if it has a convergent subsequence.)
PnExercise: Consider a linear map A : Rn → Rm given by matrix coefficients (Ax)j =
n m
k=1 ajk xk . If R and R are equipped with lp and lq norms respectively
P then the op-
eratorPnorm of A is denoted by kAkp,q . Show that: kAk1,1 = maxk j |ajk |, kAk∞,∞ =
maxj k |ajk |, kAk1,∞ = maxj,k |ajk |, and that kAk2,2 equals the square root of the great-
est eigenvalue of At A, where At is the transpose matrix. (Hint. For kAk2,2 use Lagrange
multiplier theorem along with the identity kAxk22 = xt At Ax. The interesting point is
that computing the norm kAk∞,1 is an NP-hard problem [Roh].)
Exercise: Assume measure space (X, µ) and 1 < p < ∞. (a) Given ϕ ∈ L∞ , find
the transpose of the multiplication operator Mϕ : Lp → Lp , f 7→ f´ϕ. (2) Given K :
X × X → C such that the integral operator T : Lp → Lp , (T f )(x) = K(x, y)f (y)dµ(y)
is bounded, find the transpose of T . (Hint. Refer Theorem 53 for the dual of Lp spaces.)
Here are some useful constructions:
• If X is a normed vector space then the closure of each linear subspace of X is again a
linear subspace. If X is Banach so is the closure.
• If X and Y are normed vector spaces then the Cartesian product X ×Y is also a normed
vector space with the norm k(x, y)k = max(kxk, kyk). Using other norms gives rise to
normed spaces with the same topology (Theorem 2.(1)). If X and Y are Banach so is
X ×Y.
• If X is a normed vector spaces and Y ⊆ X a closed linear subspace then the quotient
vector space X/Y = {x + Y : x ∈ X} of cosets of Y can be given the norm kx + Y k =

7
 inf y∈Y kx + yk. (If Y is not closed then this is just a seminorm.) If X is Banach so is
X/Y .

• Recall the process of the completion of metric spaces [Mun, pages 268-72][Kpy, pages
88-92]: For every metric space X there exists a complete metric space X e which contains
X as a dense subset; if Xe 0 is another such space then there is an isometric homeomor-
phism X e →X e 0 which is identity on X. If X is a normed vector space then X e inherits
a vector space structure from X. Here are two ways for completing normed vector
spaces:

1. Xe is the set of all Cauchy sequences of elements of X, modulo the equivalence

relation (xj ) ∼ (yj ) if and only if kxj −yj k → 0. The normed vector space structure
on Xe is given by (xj ) + (yj ) = (xj + yj ), a(xj ) = (axj ), k(xj )k = lim kxj k. X sits
inside Xe diagonally via x 7→ (x, x, . . .). Details can be found in [Wei, 4.11].
2. In the view of Theorem 27, one can identify X e as the closure of {bx : x ∈ X} inside
∗∗
the double dual X .

Exercise: Suppose a normed vector space X, a closed linear subspace Y ⊆ X and the
canonical quotient map π : X → X/Y . Show that: (1) π is continuous and open. (2) The
topology on X/Y induced by the canonical norm coincides with the quotient topology
namely Z ⊆ X/Y is open if and only if π −1 (Z) ⊆ X is open. (3) If X is Banach so is X/Y .
(4) X is Banach if and only if both Y and X/Y are so. (Hint for (3). Let zj be a Cauchy
sequence in X/Y . After passing to a subsequence one can assume kzj − zj+1 k < 2−j .
Then find xj ∈ X such that π(xj ) = zj and kxj −xj+1 k < 2−j+1 . Continue! Alternatively,
show that every absolutely convergent series in X/Y converges. Hint for the if part in
(4). Let xj be a Cauchy sequence in X. xj + Y is Cauchy in X/Y , so xj + Y → x + Y
for some x ∈ X. Find yj ∈ Y such that kxj − x − yj k ≤ kxj − x + Y k + 2−j . Show that
yj is Cauchy. Continue!)

Example 1. Here are the most important examples of Banach spaces:

1. The vector space B(X) of all bounded functions f : X → C on a set X, equipped with
the uniform norm kf k := supx∈X |f (x)|, is Banach. (Exercise: Prove this.) If X has a
topology then the linear subspace of continuous functions BC(X) is closed in B(X),
hence Banach. When X is a compact topological space then BC(X) equals the vector
space C(X) of continuous functions on X.

2. Lp (X, µ), 1 ≤ p ≤ ∞, (X, µ) measure space, the vector space of all (equivalence classes
´
1/p
of) Lp integrable functions X → C, equipped with the norm kf kp := |f |p dµ , is
Banach [Fol, 6.6, 6.8]. If µ is the counting measure then Lp (X, µ) is denoted by lp (X).
lp stands for lp (N).

3. Lpa (X) = Lp (X, µ) ∩ {analytic}, 1 ≤ p ≤ ∞, X ⊆ Cm open, µ Lebesgue measure, is

a closed subspace of Lp (X, µ) [Zhu-FT, page 42][Hal-S, pages 187-8], hence Banach.
These are the Bergman spaces. For more examples of Banach spaces of analytic
functions appearing in mathematical analysis refer [Zhu-FT].

8
4. Cbk (U ), k ∈ {0, 1, 2, . . .}, U ⊆ Rn open, the set of all C k functions f : U → C such
that X
kf k := sup |∂ α f (x)| < ∞,
x∈U
α∈Nn ,|α|≤k

is Banach. (Multi-index notation is introduced in page 3.)

5. W p,s (U ), 1 ≤ p ≤ ∞, s ∈ N, U ⊆ Rn open, the vector space of all (equivalence classes

of) Lp functions f : U → C which all their distributional derivatives of total order ≤ s
are represented by Lp functions, equipped with the norm
 1/p
X
kf k :=  k∂ α f kpLp (U )  ,
α∈Nn ,|α|≤s

is Banach. These are the Sobolev spaces.

6. Λα ([0, 1]), α ∈ (0, 1], the vector space of all functions f : [0, 1] → C such that
 
|f (x) − f (y)|
kf k := |f (0)| + sup : x, y ∈ [0, 1], x 6= y < ∞,
|x − y|α

is Banach. These are Hölder spaces. Λ1 ([0, 1]) is the Lipschitz space.

7. B(X; Y ), X normed vector space, Y Banach space, is Banach. Most importantly, the
dual space of every normed vector space is Banach.

8. Let l∞ be (Banach) space of all bounded sequences of complex numbers, equipped

with the supremum norm. The subset c of all convergent sequences is a closed linear
subspace of l∞ , hence Banach. The same is true for the subset c0 of all sequences
converging to 0. However, the linear subspace c00 of all sequences with finitely many
nonzero terms, is not Banach: xj = (1, 1/2, . . . , 1/j, 0, 0, . . .) is a sequence of points of
c00 which is Cauchy but not convergent (to any point of c00 ).

9. Let U be an open subset of Rn . The vector space of all continuous functions f :

´ → C compactly supported in U , with the norm kf k given by the Riemann integral
U
|f |dx, is not Banach. (Exercise: Why?) Its completion can be identified with
1
L (U ), the normed vector space of Lebesgue integrable functions [Fol, 2.41]. This
gives a functional analysis approach to develop Lebesgue measure and integral: First
comes the function space L1 , next the Lebesgue integral, and finally the Lebesgue
measure. Details can be found in [Lax, appendix A] or [Ped, chapter 6].
More examples of Banach spaces can be found in [DS, chapter 4][LT]. 
Here is the fundamental theorem on finite dimensional normed vector spaces:
Theorem 2 (Riesz). Let X be a normed vector space. Then:
(1) All norms on a finite-dimensional vector space are equivalent.
(2) Every finite-dimensional linear subspace of X is closed. More generally, the sum
of a closed linear subspace and a finite-dimensional linear subspace is closed.

9
 (3; Riesz pseudoorthogonality lemma) If Y is a proper closed linear subspace of X
then for every 0 <  < 1 there exists x ∈ X such that kxk = 1 and dist(x, Y ) > 1 − .
(4) A linear functional on X is continuous if and only if its kernel is closed.
(5) X is locally compact if and only if it is finite-dimensional.
Proof. (1) Let Y be a vector space with basisPe1 , . . . , en . In this proof always consider
Y with the topology induced by the norm k aj ej k∞ = max |aj |, aj ∈ F. Assuming
another norm k − k on Y , since K := {y ∈ Y : kyk∞ = 1} is compact by the Heine-
Borel theorem and k − k : Y → [0, ∞) is continuous by the triangle inequality, it follows
that k − k, restricted to K, takes values in a finite interval [C1 , C2 ]. By homogeneity,
C1 kyk∞ ≤ kyk ≤ C2 kyk∞ for every y ∈ Y .
(2) Let Y be a finite-dimensional linear subspace
P of X spanned by linearly independent
elements e1 , . . . , en . Assume that the sequence ni=1 aji ei , aji ∈ F, converges x ∈PX. By
the proof of (1), aji is Cauchy for each i, hence converges some bi . Therefore x = bi ei ∈
Y . The second statement can be easily reduced to the first by moding out the closed
linear subspace. (Refer to the first part of Section 3.1.(27), where by the way, another
proof for the first statement is given for a larger class of spaces X.) Here is another proof
for the second statement using the techniques of Chapter 4. By induction it suffices to
show that each Y + Fx, x ∈ X \ Y , is closed. By Theorem 25.(3) there exists a continuous
linear functional α on X such that α|Y ≡ 0 and α(x) 6= 0. If yj + aj x, where yj ∈ T
and aj ∈ F, is a sequence of points in Y + Fx converging z ∈ X then applying α to this
sequence implies that aj is convergent, so yj is convergent, hence x ∈ Y + Fx.
(3) Choose z ∈ X \ Y . Since Y is closed it follows that δ := dist(z, Y ) > 0. Choose
a sequence yj in Y such that dist(z, yj ) → δ. Setting xj := (yj − z)/kyj − zk we have
kxj k = 1 and

yj − z
dist(xj , Y ) = inf kxj − yk = inf −y =
y∈Y y∈Y kyj − zk
inf y∈Y kyj − z − kyj − zkyk δ
= ,
kyj − zk kyj − zk
approaches 1 as j → ∞. Therefore, for every  > 0, choosing j large enough, we have
dist(xj , Y ) ∈ (1 − , 1].
(4) Only if part is trivial. To prove the if part, contrapositively, assume a linear
functional α on X with closed kernel and a sequence (xj )j≥1 of norm-1 elements in X
such that |α(xj )| > j. The sequence yj := x1 − α(x1 )/α(xj )xj is in the kernel of α and
yj → x1 , so x1 ∈ Kerα . This contradicts |α(x1 )| > 1. Other proofs: The proof of Theorem
11 (for Hilbert spaces X), Section 3.1.(11) (for topological vector spaces X), the proof of
Theorem 25.(5) neglecting the first line (for locally convex topological vector spaces X).
(5) If part is by the famous Heine-Borel theorem that compact subsets of Rn are ex-
actly those which are both closed and bounded. Assuming that X is infinite-dimensional
we need to show that the closed unit ball B of X is not compact. By (2,3), inductively
construct sequence xj in (the boundary of) B such that kxj − xk k > 1/2 whenever j 6= k.
Such sequence has no convergent subsequence. Another argument. S If B1 is compact then
there exists finitely many point x1 , . . . , xn in B such that B ⊆ xj + 2 B. Let Y be the
linear space of xj . Then Y is closed by (2), so we have the canonical map π : X → X/Y

10
between normed vector spaces. Then π(B) ⊆ 12 π(B), so π(B) = 0, namely B ⊆ Y . This
can not happen if X is infinite-dimensional. 
Exercise: Justify the appellation “pseudoorthogonality lemma”.

2.1 Open mapping and closed graph theorems, Uni-

form boundedness principle
The proof of the fundamental results in the title of this section is based on the following
point-set topology result:
Theorem 3 (Baire category theorem). Let X be complete metric space or a locally com-
pact Hausdorff topological space. Then:
(1) Countable intersections of dense open subsets is dense.
(2) Countable unions of nowhere dense subsets1 has empty interior, so can not be the
whole space.
Proof. (1) First assume X to be a complete metric space. Let (Uj )j≥1 be a sequence of
dense opens of X. Assuming an arbitrary nonempty open ballTBr0 (x0 ) ⊆ X of radius
r0 ∈ (0, 1) around x0 ∈ X, we need to find a point in Br0 (x0 ) ∩ Uj . Since Br0 (x0 ) ∩ U1
is nonempty open it contains an open ball Br1 (x1 ) of radius r1 ∈ (0, 1/2) around x1 ∈ X.
Similarly, Br1 (x1 )∩U2 contains an open ball Br2 (x2 ) of radius r2 ∈ (0, 1/4) around x2 ∈ X.
Inductively, one can find sequences xj ∈ X and rj ∈ (0, 2−j ) such that Brj−1 (xj−1 ) ∩ Uj
contains Brj (xj ). Clearly, xj is a Cauchy sequence, so converges to T some x∞ ∈ X. By
our construction and the triangle inequality we have x∞ ∈ Br0 (x0 ) ∩ Uj .

The proof for locally compact Hausdorff space X is similar. This time instead of balls
Brj (xj ) we will have nonempty opens Bj resulting from local compactness: Bj is compact
T T
and contained in Bj−1 ∩ Uj . Note that Bj , clearly a subset of B0 ∩ Uj , is nonempty
because of compactness.
1
A subset Y of a topological space X is called nowhere dense if the interior of the closure of Y is
empty; in other words, every nonempty open subset of X has a point not in Y .

11
 subsets of X. Then X \ Yj S
(1)⇒(2) Let Yj be a sequence of nowhereS dense T is a
sequence of dense open subsets of X, so X \ Yj = X \ Yj is dense. Therefore Yj
has no interior. 

Application 4. There exists a continuous function [0, 1] → R which is not differentiable

at any point; in fact, nowhere differentiable functions are dense in C([0, 1]; R) with respect
to the uniform topology.

Proof. Let X := C([0, 1]; R) be the vector space of continuous functions [0, 1] → R,
equipped the with uniform topology. If f ∈ X is differentiable at x = a then |f (x) −
f (a)| ≤ (|f 0 (a)| + 1)|x − a| on some neighborhood of a; since the graph of f is compact
outside this neighborhood it follows that |f (x) − f (a)| ≤ K|x − a| on whole [0, 1] for
some large enough K. Contrapositively, this shows that the set of nowhere differentiable
functions in X contains ∞
T
j=1 X \ Y j where

Yj = {f ∈ X : |f (x) − f (a)| ≤ j|x − a|, ∃a ∈ [0, 1]∀x ∈ [0, 1]} .

The intuitive picture about elements f ∈ Yj is that their graph are confined in the
biconical region with vertex centered at some (a, f (a)) and border slopes ±j. By Baire
category theorem we need to show that each X \ Yj is open and dense. To show that Yj is
closed, assume a sequence fn in Yj which converges f ∈ X. Since fn ∈ Yj there exists an
such that |fn (x) − fn (an )| ≤ j|x − an | for every x. By Bolzano-Weiertrass theorem, after
passing to a subsequence one can assume that an converges some a. Since the convergence
of fn is uniform it follows that fn (an ) → f (a) and we have

|f (x) − f (a)| = lim |fn (x) − fn (an )| ≤ lim j|x − an | = j|x − a|, ∀x ∈ [0, 1],
n n

hence f ∈ Yj . To prove the denseness of X \ Yj , fixing f ∈ X and  > 0 we should

find some h ∈ X with kh − f k <  and h 6∈ Yj . By the uniform continuity of f one
can find a piecewise linear function (with finitely many slopes) in any neighborhood of f ,

12
so without loss of generality we can assume f is piecewise linear. The idea is to mount
a sharp sawtooth function on f . Let M ≥ 0 be the maximum slope of f , measured in
absolute value. (That is if f has slopes s1 , . . . , sn then M := max{|s1 |, . . . , |sn |}.) Choose
an integer N strictly larger that (M + j)/, and construct the sawtooth function g ∈ X
with vertices g(j/N ) = (−1)j for j = 0, . . . , N . We assert that h := f + g/2 ∈ X works
as our desired function. Clearly, kh − f k = /2 < . On the other hand, h is piecewise
linear, and the absolute values of the slopes of h exceed

 2
M− = |M − N | > j.
2 1/N

Therefore h 6∈ Yj . 
Here are some concrete examples of continuous nowhere differentiable functions:
−j
√
exp( −13j x) [Grf, 3.7.3][Har].
P
• j≥0 2

−j
{10j x}, where {x} denotes the distance of the real number x to the
P
• j≥0 10
nearest integer [DiB, page 228].

Exercise: The vector space dimension of a Banach space can not be ℵ0 , the cardi-
nality of integers. (Hint. Use the Baire category theorem along with the fact that finite
dimensional linear subspaces of a normed vector space are closed.)

Theorem 5 (Banach). (1- Uniform boundedness principle) A family Tα : X → Y , α ∈ A,

of bounded operators from Banach space X to normed vector space Y is uniformly equi-
bounded (namely supα∈A kTα k < ∞) if and only if it is pointwisely equibounded (namely
supα∈A kTα xk < ∞ for every x ∈ X).
(2- Open mapping theorem) A bounded operator between Banach spaces is surjective
if and only if it is open (namely maps opens to opens).
(20 - Inverse mapping theorem) A bounded operator between Banach spaces is bijective
if and only if it a homeomorphism.
(3- Closed graph theorem) A linear map T : X → Y between Banach spaces is contin-
uous (namely for every sequence xj in X such that xj → x it is the case that T xj → T x)
if and only if its graph GT := {(x, T x) : x ∈ X} is closed in X × Y (namely for every
sequence xj in X such that xj → x and T xj → y it is the case that T x = y).

A family of bounded operators from a Banach space to a normed vector

space is uniformly equibounded if and only if it is pointwisely equibounded.

A bounded operator between Banach spaces is surjective (respectively,

bijective) if and only if it open (respectively, a homeomorphism).

To show that a linear map T between Banach space is continuous

it is enough to check that xj → 0 and T xj → y implies y = 0.

13
Proof. (1) Only if part is trivial. For the other direction, since X is the union of closed
subsets Xj := {x ∈ X : sup kTα xk ≤ j}, j = 1, 2, . . ., by Baire category theorem some
Xj contains a nonempty open ball B (x0 ) = {x ∈ X : kx − x0 k < }. This means
that kx − x0 k <  implies sup kTα xk ≤ j, hence sup kTα (x − x0 )k ≤ 2j. Therefore,
sup kTα k ≤ 2j/.
(2) Let T : X → Y be a bounded operator between Banach spaces. For any positive
real r let Xr be the open ball of radius r in X around the origin; similarly define Yr . Since
T is linear and all nonempty open balls in a normed vector space are homeomorphic it
follows that T being open is equivalent to T X1 containing someS Yr . If part is now trivial.
Conversely, assume that T is surjective. Therefore Y = j∈N T Xj . Since all T Xj are
homeomorphic it follows by Baire category theorem that T X1 ⊇ Br (y0 ), open ball in Y
of radius r > 0 around some y0 . From this, just using the linearity of T , we will show
that T X1 ⊇ Yr/4 , as follows. Since y0 ∈ T X1 it follows that there exists x1 ∈ X1 such
that ky0 − y1 k < r/2 where y1 = T x1 . Since X1 ⊆ B2 (x1 ) and Br (y0 ) ⊇ Br/2 (y1 ) it
follows that T B2 (x1 ) ⊇ Br/2 (y1 ). By linearity T X2 ⊇ Yr/2 , or equivalently, T X1 ⊇ Yr/4 .
Replacing r with 4r we deduce that T X1 ⊇ Yr , or more generally T X2−j contains Y2−j r
for every j ∈ N. We assert that T X1 ⊇ Yr/2 . Assume y ∈ Yr/2 . Find x1 ∈ X1/2
such that ky − T x1 k < r/4. Since y − T x1 ∈ Yr/4 one can find x2 ∈ X1/4 such that
ky − T P x1 − T x2 k < r/8. Continuing this
P process we find a sequence xj ∈ X2−j such that
j −j−1
ky −T k=1 xk k < P r2 . The
Pseries xj is absolutely convergent, so it converges some
−j
x ∈ X with kxk ≤ kxj k ≤ 2 < 1. We found x ∈ X1 with y = T x.
(2)⇒(20 ) If part is trivial. Only if part is immediate from (2).
(20 )⇒(2) Let T : X → Y be a surjective bounded operator between Banach spaces.
Since T is continuous it follows that KerT ⊆ X is closed. T equals the canonical quotient
map X → X/KerT composed with the bijective linear map S : X/KerT → Y , x+KerT 7→
T x. For every x ∈ X, the computation

kS(x + KerT )k = inf kT x0 k ≤ kT k inf kx0 k ≤ kT kkx + KerT k

x0 ∈x+KerT x0 ∈x+KerT

shows that S is bounded. Therefore, S is a homeomorphism by (2). T is open because

one can easily show that every canonical quotient map between normed vector spaces is
open.
(20 )⇒(3) Consider the bijective linear map S : X → GT , x 7→ (x, T x). Note that GT
is a closed linear subspace of a Banach space, so is itself Banach. Applying (20 ) to the
(continuous) set-theoretic inverse π1 : GT → X of S implies that S is continuous. The
continuity of T follows. 
To see this fundamental theorem of Banach in more generality refer Remark 24. If
the Banachness assumption is dropped from any of the statements in Theorem 5 then
the conclusion does not hold. (For examples refer [Fol, exercises 29–31] or one of the
exericses in page 15.
Here is an application of the uniform boundedness principle to Fourier analysis.

Application 6 (duBois Reymond). There exist a continuous periodic function R → C

whose Fourier series diverges at a point.

14
Proof. We will prove the existence of some f ∈ C([0, 1]) whose Fourier series diverges at
the origin. Recall that the partial sums of the Fourier series of any f ∈ C([0, 1]) is given
by ˆ 1
X √ √
2π −1kx
Sj (x) = ck e where ck = f (x)e−2π −1kx dx, k ∈ Z.
|k|≤j 0

Therefore at x = 0 we have
ˆ 1
Sj := Sj (0) = f (x)Dj (x)dx,
0

where the Dirichlet kernel Dj (x) is given by

(
√ sin(π(2j+1)x)
X
−2π −1kx sin(πx)
, 0 < x < 1,
Dj (x) = e = .
|k|≤j
2j + 1, x = 0 or x = 1

Now consider Sj as a sequence of linear maps from Banach space C([0, 1]) to normed
vector space C. Each Sj is a bounded operator whose norm does not exceed Cj :=
´1
0
|Dj (x)|dx < ∞. On the other hand, since each Dj (x) has a finite number of simple
zeros on [0, 1] it follows that its sign function sgn Dj (x) has a finite number of jump
discontinuities; therefore for every  > 0, by modifying this sign function on a small
neighborhood of each of its discontinuities, one can construct f ∈ C([0, 1]) such that
kf k = 1 and |Sj f | ≥ Cj − . We have shown that kSj k = Cj . If we prove that
Cj → ∞ then by the uniform boundedness principle there exists f ∈ C([0, 1]) such that
the sequence |Sj f | is unbounded, and we are done. There exist positive constants K1 and
K2 such that:
ˆ 1 ˆ 1
2 |sin(π(2j + 1)x)| 2|sin(π(2j + 1)x)| πx
Cj = 2 dx = 2 dx =
0 sin(πx) 0 πx sin(πx)
ˆ 1 ˆ j+ 1 j−1 ˆ k+1
2 |sin(π(2j + 1)x)| 2 |sin(πx)| X |sin(πx)|
≥ K1 dx = K1 dx ≥ K1 dx =
0 x 0 x k=0 k x
j−1 ˆ 1 j−1 ˆ 3 j−1
X |sin(πx)| X 4 dx X k + 43
K1 dx ≥ K2 = K2 log 1.
k=0 0 x + k k=1 4
1 x + k k=1
k + 4

The last series is asymptotically equivalent to the harmonic series, hence diverges. 

Remark 7. For an explicit functions satisfying Application 6 refer [Grf, volume I, exercise
3.3.6]. Kolmogorov constructed an L1 function whose Fourier series diverges everywhere
[Kol][Zyg, volume I, pages 310-4]. Carleson and Hunt proved that the Fourier series of an
Lp function, 1 < p < ∞, converges almost everywhere [Grf, chapter 11]. It is relatively
easy to show that the Fourier series of a C 1 function (or even a function of bounded
variation) converges everywhere [Apo-A, 11.12][Grf, 3.3.9]. 
1
P Exercise: Let N+ be the set of positive integers, Y := l (N+ ) and X := {x ∈ Y :
j|xj | < ∞}. Show that: (1) X is not complete. (2) T : X → Y , (T x)j = jxj , is

15
bijective, closed, but not bounded. Why the closed graph theorem does not apply? (3)
The set-theoretic inverse S : Y → X of T is bounded, neither open nor a homeomorphism.
Why the open mapping theorem or the inverse mapping theorem does not apply? (Hint
for (1). By truncating series show that X $ Y is dense. Hint for (2). To show that T is
not bounded consider x = (q j ), q ∈ R, or x given by xn = 1 and zero everywhere.)
Exercise: Prove the following if you are familiar with Hilbert spaces. A linear map
T : X → X on Hilbert space X which satisfies hT x, yi = hx, T yi for every x, y ∈ X
(so-called self-adjoint operator) is continuous. (Hint. Use the closed graph theorem.)
Exercise: Let k−k1 and k−k2 be two norms on a vector space which makes it Banach.
If there exists C > 0 such that kxk1 ≤ Ckxk2 for every x ∈ X then there exists D > 0
such that kxk2 ≤ Dkxk1 for every x ∈ X, so the norms are equivalent. (Hint. Use the
inverse mapping theorem.)
Exercise: Let Tj : X → Y be a sequence of bounded operators between Banach spaces
such that Tj x converges for every x ∈ X. Then the limit map T : X → Y , x 7→ lim Tj x,
is a bounded operator. This is called Banach-Steinhaus Theorem. (Hint. Use the
uniform boundedness principle.)
Exercise: A bilinear map B : X × Y → Z, with X Banach and Y, Z normed vector
space, is continuous if and only if it is separately continuous (namely B(x, −) and B(−, y)
are continuous for every x ∈ X and y ∈ Y ) if and only if there exists C > 0 such
that kB(x, y)k ≤ Ckxkkyk for every x ∈ X and y ∈ Y . (Hint. To deduce the last
statement from the second apply the uniform boundedness principle to the family of
bounded operators Ty := B(−, y)/kyk, y ∈ Y \ {0}.)
Exercise: Let Lpa (D) be the Bergman space on the unit disk of the complex plane
(Example 1). Justify the use of closed graph theorem in the following excerpt from [DS,
page 61]: “The´ problem here is to characterize those finite Borel measures µ on D with the
property that |f |p dµ < ∞ for every
´ f ∈ L2a (D). If µ is any such measure it follows from
the closed graph theorem that |f | dµ ≤ Kkf kpp for some constant K > 0 depending
p

only on p.”
The following strong version of injectivity of operators is very important in operator
theory:
Theorem 8. An operator T : X → Y between Banach spaces is injective and range-
closed if and only if it is bounded from below in the sense that kT xk ≥ Ckxk for
some C > 0 and every x ∈ X.
Proof. Let T be bounded from below. T is clearly injective. To prove that RanT is closed
assume a sequence xj in X such that T xj converges some y ∈ Y . Then T xj is Cauchy, so
by the bounded-from-below inequality xj is also Cauchy. Therefore, xj converges some
x ∈ X, and by continuity T x = y. This proves that RanT is closed. Conversely, let T
be injective and range-closed. Then the inverse mapping theorem applied to X → RanT ,
x 7→ T x, gives the bounded-from-below inequality. 

2.2 Hilbert spaces

As usual let F be either the field of reals or the field of complex numbers, and let X be
a vector space over F. X is called a pre-Hilbert space if it is equipped with an inner

16
product namely a binary operation h−, −i : X × X → F satisfying:

1. F-linearity with respect to the second argument.2

2. Conjugate (or Hermitian) symmetry: hx, yi = hy, xi for every x, y ∈ X. (When

F = R this reduces to the usual symmetry hx, yi = hy, xi.)

3. Positivity: hx, xi ≥ 0 for every x ∈ X, with the equality happening only for x = 0.

We develop the theory for complex vector spaces but the theory for real vector spaces
is similar. Here are some facts about pre-Hilbert spaces:

• Cauchy-Schwarz inequality:
| hx, yi | ≤ kxkkyk,
with the equality
p happening exactly when x and y are linearly dependent. Here kxk is
defined by hx, xi, and similarly for kyk. We will show in the next item that this in
fact defines a norm on X.
(Proof. Put the trivial case y = 0 aside. Expand kx − hy, xi y/ hy, yik2 ≥ 0. Another
argument. For real variable t the quadratic expression kx − tyk2 = kyk2 t2 −2tRe hx, yi+
2
kxk2 is everywhere nonnegative, so its discriminant ∆ = 4 (Re hx, 2
√ yi) − 4kxk kyk is
2

nonnegative, namely |Re hx, yi | ≤ kxkkyk. Replacing y by y exp( −1θ), θ ∈ R, implies

|Re hx, yi cos θ − Im hx, yi sin θ| ≤ kxkkyk. We are done since this is true for every θ.)
p
• kxk := hx, xi is a norm. When we talk about the topology of a pre-Hilbert space we
always mean the topology induced by this norm, namely balls {y ∈ X : kx − yk < },
x ∈ X,  > 0, are basic opens.
(Proof. Triangle inequality is immediate from the Cauchy-Schwarz inequality.)

• The inner product is continuous with respect to each of its arguments separately.
(Proof. Immediate from the Cauchy-Schwarz inequality.)

• Parallelogram law:

kx + yk2 + kx − yk2 = 2 kxk2 + kyk2 .

(Proof. Add up two identities kx ± yk2 = kxk2 + kyk2 ± 2Re hx, yi.)

• Polarization identity: The inner product can be reconstructed from the norm it induces,
more precisely:

 1 P3 √ j √ j

2
−1 −1 x + y , complex pre-Hilbert spaces
hx, yi = 4 j=0 .
 1 (kx + yk2 − kx − yk2 ) , real pre-Hilbert spaces
4
2
This is physicists’ convention. Mathematicians usually assume linearity with respect to the first
argument.

17
• Jordan-von Neumann theorem:
p A norm on a vector space is induced by an inner
product (namely via kxk = hx, xi) if and only if it satisfies the parallelogram law.
(Proof. Let the norm k − k satisfy the parallelogram identity, and define the inner
product by the polarization identity. The only nontrivial thing to check it the linearity
of the inner product with respect to its second argument. It is straightforward to show
that hx, yi + hx, zi = 2 hx, (y + z)/2i. Setting z = 0 and by induction one can deduce
hx, yi + hx, zi = hx, y + zi and hx, ayi = a hx, yi for a = 2−j k where j, k ∈ N. Since
scalars of the form 2−j k are dense among positive reals it follows that hx, ayi = √
a hx, yi
for every positive real a. That the same identity is true for a = −1 and a = −1 is
straightforward. Details can be found in [Wei, 1.6].)

Exercise: Prove that there is no inner product on C([0, 1]) which induces the uniform
norm. p
A pre-Hilbert space which is complete with respect to the norm kxk = hx, xi induced
by its inner product is called a Hilbert space. (Completeness means that Cauchy
sequences converge.) An Hilbert spaces isomorphism is a bijective linear map between
Hilbert spaces which preserves the inner product. (Such a map and its inverse are clearly
isometries.) Two elements of a Hilbert space are said to be orthogonal to each other
if their inner product equals zero. For a subset S of a Hilbert space, S ⊥ , called the
orthogonal complement of S, is the set of all elements in the Hilbert space which are
orthogonal to every element of S. Note that S ⊥ is always a closed linear subspace of the
ambient Hilbert space.
Here are some easy facts and constructions:
• The closure of a linear subspace of a Hilbert space is again a Hilbert space.

• The Cartesian product of two Hilbert spaces X, Y is a Hilbert space via the inner
product h(x1 , y1 ), (x2 , y2 )i = hx1 , x2 i + hy1 , y2 i. It is usually denoted by X ⊕ Y .

• Recall the process of completion of normed vector spaces by Cauchy sequences (page
b is the completion of pre-Hilbert space X then h(xj ), (yj )i = lim hxj , yj i makes
8). If X
Xb a Hilbert space. More conceptually, since the inner product on X is a continuous
function it can be uniquely extended to an inner product on X. b

• Suppose Hilbert spaces X, Y over field F, and let Z be their algebraic tensor product,
that is the free vector space generated by the set of symbols {x ⊗ y : x ∈ X, y ∈ Y }
mod out by the relations

(x + x0 ) ⊗ y = x ⊗ y + x0 ⊗ y, x ⊗ (y + y 0 ) = x ⊗ y + x ⊗ y 0 , (ax) ⊗ y = x ⊗ (ay),

where x, x0 ∈ X, y, y 0 ∈ Y , a ∈ F. Z can be made a pre-Hilbert space by the inner

product * +
X X X
xj ⊗ y j , x k ⊗ yk = hxj , xk i hyj , yk i .
j k j,k

The completion of Z is called the tensor product of X, Y , and denoted by X ⊗ Y .

This construction is of fundamental importance in quantum mechanics, basically in

18
 the description of the state space of a composite system in terms of the state spaces of
its parts [Hall, pages 430, 432, 434]. Here is the construction. Refer [Wei, section 3.4]
for more details. Exercise: After learning the notion of orthonormal bases of Hilbert
spaces, prove that if {xα : α ∈ A} and {yβ : β ∈ B} are orthonormal beses for Hilbert
spaces X and Y , respectively, then {xα ⊗ β : α ∈ A, β ∈ B} is an orthonormal basis
for X ⊗ Y .

Example 9. Here are the most important examples of Hilbert spaces:

Pn
• Cn is a Hilbert space with inner product h(x1 , . . . , xn ), (y1 , . . . , yn )i =
xj yj . j=1
´
• L2 (X, µ), (X, µ) measure space, is a Hilbert space with inner product hf, gi = f gdµ.
When µ is the counting measure L2 (X, µ) = l2 (X). Theorem 21 shows that these latter
constitute all Hilbert spaces up to isomorphisms of Hilbert spaces.

• L2a (X) = L2 (X) ∩ {holomorphic}, X ⊆ Cm open, is a closed subspace of L2 (X), hence

a Hilbert space.

• W 2,s (U ), s ∈ N, U ⊆ Rn open, is a Hilbert space with the inner product

X ˆ
hf, gi = ∂ α f (x)∂ α g(x)dx,
α∈Nn ,|α|≤s

where dx denotes the Lebesgue measure. (Refer Example 1 for the definition of Sobolev
spaces.) 

Here is the most fundamental variational or geometric property of Hilbert spaces, all
the other future theorems are based on:

Theorem 10 (Orthogonal projection). Let X be a Hilbert space and Y ⊆ X a nonempty

closed convex subset. Then:
(1) For every x ∈ X there exists a unique y ∈ Y of shortest distance to x. y is the
only element in Y such that Re hx − y, y 0 − yi ≤ 0 for every y 0 ∈ Y .
(2) If Y is furthermore assumed to be a linear subspace then y, obtaind in (1), is the
only element in Y such that x − y ∈ Y ⊥ . The mapping P : X → X, x 7→ y, is called
the orthogonal projection in X onto Y . P is a bounded operator of norm 1, unless
Y = {0}.

19
Proof. (1) Let δ ≥ 0 be the distance of x to Y , namely inf y∈Y kx − yk, and let yj be a
sequence in Y realizing δ, namely kx − yj k → δ. yj is a Cauchy sequence because the
expression

kyj − yk k2 = k(x − yj ) − (x − yk )k2 = 2kx − yj k2 + 2kx − yk k2 − 4kx − (yj + yk )/2k2 ≤

2kx − yj k2 + 2kx − yk k2 − 4δ 2 ,
can be arbitrary small if j, k are sufficiently large. (We have used the parallelogram
identity. (yj + yk )/2 ∈ Y because Y is convex.) Let yj → y. Then y ∈ Y because Y is
closed. The continuity of the norm implies kx − yk = δ. If ye is another point of distance
δ to x then again by the parallelogram identity and the convexity of Y we have
2
y + ye
ky − yek2 = k(x − y) − (x − ye)k2 = 2kx − yk2 + 2kx − yek2 − 4 x − ≤ 4δ 2 − 4δ 2 = 0,
2
hence y = ye. We next show that
Re hx − y, y 0 − yi ≤ 0, ∀y 0 ∈ Y, (2.1)
is equivalent to y being a point in Y of shortest distance to x. If (2.1) holds then
kx − y 0 k2 = k(x − y) − (y 0 − y)k2 = kx − yk2 − 2Re hx − y, y 0 − yi + ky 0 − yk2 ≥ kx − yk2 .
For the other direction fix y 0 ∈ Y . For every real number t ∈ [0, 1] we have

kx − yk2 ≤ kx − (ty 0 + (1 − t)y)k2 = k(x − y) − t(y 0 − y)k2 =

kx − yk2 − 2tRe hx − y, y 0 − yi + t2 ky 0 − yk2 ,
or equivalently
0 ≤ −2Re hx − y, y 0 − yi + tky 0 − yk2 , ∀t ∈ (0, 1]. (2.2)
Tending t → 0+ gives (2.1).
(2) We first show that x − y ∈ Y ⊥ . Fixing y 0 ∈ Y , for every real number t ∈ R we
have

kx − yk2 ≤ kx − (ty 0 + (1 − t)y)k2 = k(x − y) − t(y 0 − y)k2 =

kx − yk2 − 2tRe hx − y, y 0 − yi + t2 ky 0 − yk2 ,
which happens exactly when Re hx √ − y, y 0 − yi = 0. Since y 0 ∈ Y was arbitrary we have
0 0
Re hx − y, y i = 0. Replacing y by −1y 0 we have Im hx − y, y 0 i = 0, hence hx − y, y 0 i =
0. For uniqueness, set z := x − y and let ye be another element of Y such that ze := x − ye ∈
Y ⊥ . Then y − ye = ze − z belongs to Y ∩ Y ⊥ = {0}. 
Compare this proof with the proof of the finite dimensional case given in [Apo-C,
volume I, Theorems 15.13–16].
Exercise: Let D be a bounded open subset of the complex plane. Note that then
the Bergman space L2a (D) contains all polynomial functions. Given arbitrary distinct
points aj ∈ D, j = 1, . . . , n and arbitrary complex numbers bj , prove that there exists
f ∈ L2a (D) of smallest norm such that f (aj ) = bj for every j. (See also [DS, Section 1.4].)

20
Theorem 11 (Riesz representation theorem). Let X be a Hilbert space. The conjugate-
linear mapping X → X ∗ , x 7→ hx, −i, is an isometric isomorphism of normed vector
spaces. In other words, for every continuous linear functional α on X there exists a
unique vector x ∈ X such that α(y) = hx, yi for every y ∈ X, and kαk = kxk. There
is a unique inner product on X ∗ which induces the operator norm, and it is given by
hhx, −i , hy, −iiX ∗ = hy, xiX . Furthermore, X is reflexive (defined in Theorem 27.)

Proof. For every x ∈ X let αx denote hx, −i. Assume α ∈ X ∗ , and let Y be the kernel
of α. Note that Y is a closed subspace. If α is the zero functional then x = 0 works,
otherwise since X = Y + Y ⊥ , one can find a nonzero element z ∈ Y ⊥ . Since the linear
functional β := αz vanishes wherever α does it follows that β = Cα for some scalar
C. (Proof. Choose z0 ∈ X such that α(z0 ) = 1. For each z ∈ X, since α vanishes at
z − α(z)z0 it follows that 0 = β(z − α(z)z0 ) = β(z) − β(z0 )α(z), so C := β(z0 ) works.
Q.E.D.) Evaluating β = Cα at z gives C = kzk2 /α(z). This means that x := α(z)kzk−2 z
works. Uniqueness: If αx = αx0 then x − x0 ∈ X ⊥ = {0}. Next we compute the operator
norm kαx k = supkyk≤1 | hx, yi |. By the Cauchy-Schwarz inequality we have kαx k ≤ kxk.
Setting y := x/kxk (in case x 6= 0) gives kαx k ≤ kxk, hence kαx k = kxk. This latter
equation persists when x = 0. Clearly, hαx , αy i := hy, xi is an inner product. It induces
the operator norm because hαx , αx i = hx, xi = kxk2 = kαx k2 . Uniqueness comes from
the polarization identity. For reflexivity, assume a continuous linear functional F on X ∗ .
Since X ∗ is again a Hilbert space, by what we have proved so far it follows that there
exists αx such that F (αy ) = hαx , αy i = hy, xi = αy (x) = xb(αy ) for every y ∈ X. Since
∗
every β ∈ X is of the form αy it follows that F = x b. 
The linear algebra argument in the parenthesis of the latter proof can be generalized
to prove the following linear T nullstellensatz: If α1 , . . . , αn , β are linear functionals
on a vector space X Pthen Kerαj ⊆ Kerβ (namely β vanishes as soon as all αj do)
if and only if β = Cj αj for some scalars Cj . Proof. To prove the only if part one
can assume that αj are linearly independent. Choose dual basis xj , namely elements
x1 , . P
. . , xn in X such P the Kronecker delta δij . For every x ∈ X then
T that αi (xj ) equals
x − αj (x)xj ∈ Kerαj , hence β = β(xj )αj . Other proofs can be found in [Rud-FA,
3.9][HK, page 110]. Q.E.D.

Theorem 12. For every bounded operator T : X → Y between Hilbert spaces there exists
a unique bounded operator T ∗ : Y → X, called the adjoint of T , such that hT x, yi =
hx, T ∗ yi for every x ∈ X and y ∈ Y . Furthermore, ∗ : B(X; Y ) → B(Y ; X), T 7→ T ∗ , is
a conjugate linear operation with extra properties

T ∗∗ = T, kT ∗ k = kT k, kT T ∗ k = kT k2 .

Proof. There is only one way to construct T ∗ , as follows. Fix y ∈ Y . Consider the
linear functional X → C, x 7→ hy, T xi. Its norm is bounded by kT kkyk according to the
Cauchy-Schwarz inequality, so by the Riesz representation theorem, there exists a unique
element z ∈ X such that hy, T xi = hz, xi, or equivalently, hT x, yi = hx, zi. Set T ∗ y = z.
For every x ∈ X and y ∈ Y we have

hy, T ∗∗ xi = hT ∗ y, xi = hx, T ∗ yi = hT x, yi = hy, T xi ,

21
hence T ∗∗ = T . Taking supremum over {x ∈ X : kxk ≤ 1} of the inequality
kT xk2 = | hx, T ∗ T xi | ≤ kT ∗ T xk ≤ kT ∗ kkT xk ≤ kT ∗ kkT kkxk,
gives kT k2 ≤ kT ∗ T k ≤ kT ∗ kkT k. Together with T ∗∗ = T these inequalities give kT ∗ k =
kT k and kT T ∗ k = kT k2 . 
Example 13. S : l2 (N) → l2 (N), (a0 , a1 , a2 . . .) 7→ (0, a0 , a1 , . . .), is called the (unilat-
eral) forward shift, one of the most important operators in functional analysis. Its
adjoint (a0 , a1 , a2 . . .) 7→ (a1 , a2 , . . .), is called the (unilateral) backward shift. 
Exercise: Assume a Hilbert space X, and finitely many elements x1 , . . . , xm of it.
Show that the analysis and synthesis operators
X
X → Cm , x 7→ (hx, xj i), Cm → X, (aj ) 7→ aj x j
are adjoint to each other.
Exercise: Show that an operator T on a Hilbert space X is the orthogonal projection
onto closed linear subspace Y ⊆ X if and only if T ◦ T = T = T ∗ and Y = RanT . (Hint.
T ◦ T = T implies RanT = Ker1−T .)
Exercise: Let T be an operator on a Hilbert space X such that hT x, xi = 0 for every
x ∈ X. Show that: (1) If F = C then T = 0. (2) If F = R and T = T ∗ then T = 0.
(3) T = √T ∗ can not be dropped in (2). (Hint. For (1), inspect hT z, zi for z = x ± y and
z = x ± −1y. For (3), think about rotations.)
Theorem 14. Let T : X → Y be an operator between Hilbert spaces. Then:
(1) T is an isometry (namely kT xk = kxk for every x ∈ X) if and only if it preserves
the inner product (namely hT x, T yi = hx, yi for every x, y ∈ X) if and only if T ∗ T = 1.
(2) T is an isomorphism of Hilbert spaces (namely a bijective linear map which pre-
serves the inner product) if and only if it is a surjective map which preserves the norm if
and only if it is unitary in the sense that T ∗ T = 1 and T T ∗ = 1.
(3; Mazur-Ulam) Every map between normed real vector spaces which preserves the
distance and the origin is linear.
Proof. (1) kT xk = kxk for every x ∈ X implies hT x, T yi = hx, yi for every x, y ∈ X via
the polarization identity.
(2) Immediate from (1).
(3) The general case is proved in [Lax]. Here we treat the Hilbert space case. The only
property of Hilbert spaces that we use is that: The triangle inequality kx+yk ≤ kxk+kyk
is strict unless one of x or y is a nonnegative multiple of the other. (This is straightforward
to check recalling the equality condition in the Cauchy-Schwarz inequality.) Let T : X →
Y be such a map. Fix x, y ∈ X and set z := (x + y)/2. Then

k(T x − T z) + (T z − T y)k = kT x − T yk = kx − yk = 2k(x − y)/2k = kx − zk + kz − yk =

= kT x − T zk + kT z − T yk,
implies that one of T x − T z or T z − T y is a nonnegative multiple of the other. Since both
has norm kx − yk/2 it follows that T x − T z = T z − T y, namely T x + T y = 2T ((x + y)/2).
Putting y = 0 gives T x = 2T (x/2), hence T (x+y) = T x+T y. By induction T (ax) = aT x
for rational scalars a. By continuity T (ax) = aT x for real scalars a. 

22
 Exercise: Why Mazur-Ulam theorem is not true for complex normed vector spaces?

Lemma 15 (Parseval Inequality). Let S be an orthonormal subset of a Hilbert space X

in the sense that all elements ofP
S are of norm one, and every two distinct elements are
orthogonal to each other. Then s∈S | hx, si |2 ≤ kxk2 for every x ∈ X.3

Proof. For every finite subset F ⊆ S we have:

2 * +
X X X X
x− hx, sis = kxk2 − 2Re x, hs, xis + | hx, si |2 = kxk2 − | hx, si |2 .
s∈F s∈F s∈F s∈F

Theorem 16 (Orthonormal basis). For an orthonormal subset S of a Hilbert space X

the followings are equivalent:
maximal among orthonormal sets, or equivalently S ⊥ = {0}.
(1) S is P
(2) x = s∈S hx, sis for every x ∈ X. (Note that by Lemma 15 the summation on the
right hand side has only countably many nonzero terms. The meaning of the convergnce
of this series is that the series converges in the topology of X no matter how its terms
are enumerated.)
(3) kxk2 = s∈S | hx, si |2 for every x ∈ X. This is called the Parseval identity.
P
P
Equivalently, hx, yi = s∈S hx, si hy, si for every x, y ∈ X.
(4) X → l2 (S), x 7→ (hs, xi)s∈S , is an isomorphism of Hilbert spaces.
(5) The closed linear span of S is X. (Closed linear span is defined in page 25)
If any of these conditions hold then S is called an orthonormal basis for X.

Proof. (1)⇒(2) Let sj , j ∈ N, be an enumeration of all those s ∈ S such that hx, si =

6 0.
2
P P
The series hx, sj isj is Cauchy because | hx, sj i | is so:

m n 2 m n
X X X X
2
hx, sj isj − hx, sj isj = | hx, sj i | − | hx, sj i |2 , ∀m > n.
j=0 j=0 j=0 j=0

Let y ∈ X be the convergence point of the series. By continuity, hy, si = hx, si for all
s ∈ S, so according to (1) we have y = x.
(2)⇒(3) Immediate from the continuity of the norm.
(3)⇒(1) Trivial.
(3)⇒(4) (3) says that the map is an isometry. To prove surjectivity assume (cs )s∈S ∈
l2 (S). Let sj be an enumeration of the countable
P set of all those s ∈ S such that cs 6= 0.
Then the arguments in (2) shows that csj sj converges and the convergence point is
a preimage of (cs )s∈S . Another argument for surjectivity. The range of the map in the
3
Here is the definition and some elementary facts about the arbitrary sums of nonnegative
P numbers.
If (xα )α∈A is a collection
P of nonnegative real numbers indexed over a set A then x α is defined to be
the supremum of all α∈F xα , F varying over finite subsets of A. P Let B be the set of all α ∈ A such
that xα > 0. One can easily show that if B is uncountable then xα = ∞, but if B = {β1 , β2 , . . .} is
P P∞ Pk
countable then xα equals the usual series j=1 xβj namely limk→∞ j=1 xβj . Refer [Fol, page 11]
for details.

23
statement of (4) is dense and closed. (A standard Cauchy sequences argument shows that
the range of any isometry is closed.)
(4)⇒(3) The map is an isometry.
(2)⇒(5) Trivial.
(5)⇒(1) Assume x ∈ S ⊥ . Given  > 0, find a finite linear combination y of elements of
S such that kx − yk < . Since x is orthogonal to y, it follows that kxk2 ≤ kxk2 + kyk2 =
kx − yk2 < 2 . Therefore, x = 0. 
Exercise: Prove that unitary operators are exactly the operators of the change of
bases, namely, an operator T on a Hilbert space X is unitary if and only if there exist
orthonormal bases {xα : α ∈ A} and {yα : α ∈ A} for X such that T xα = yα for all α.
There is an alternative way, based on the notion of nets [Fol, section 4.3], to make
sense of the series like the one in the right hand side of the equation in Theorem 16.(2).
Here are the definitions. A poset (A, ≤) is called a directed set if for every α, β ∈ A
there exists γ ∈ A such that α ≤ γ and β ≤ γ. A net (xα )α∈A in a set X is a collection
of elements xα ∈ X indexed over a directed set A, namely a function A → X. Now let X
be a normed vector space. One says that the net (xα ) converges to x ∈ X if for every
 > 0 there exists α0 ∈ A such that kxα − xk <  for every α ≥ α0 . It is straightforward
to show that X is Banach if and only if every Cauchy net (namely for every  > 0 there
exists α ∈ A such thatP kxβ − xγ k <  for every β, γ ≥ α) converges P [Dou, page 3]. One
says that the series α∈A xα converses to x ∈ X if the net yF = α∈F xα indexed over
the directed set ({F : F ⊆ A, F finite}, ⊆) of all finite subsets of A ordered by inclusion,
converges to x. Here is a basic fact:
P
Theorem 17. If xα is a P collection of orthogonal points in a Hilbert space X then xα
2
converges if and
P only if kx α k converges. If so then there is only countably many
nonzero xα , xα converges to the
P usual series of any enumeration of α’s such that
2 2
P
xα 6= 0, and we have k xα k = kxα k .

Proof. Only ifPpart is2 immediate from the continuity of the inner product. Conversely,
assume thatP kxα k 2 < ∞. P Therefore 2
for every  > 0 there exists finite subset F0 ⊆ A
such that α∈F kxα k − α∈F0 kxα k <  for every finite subset F ⊆ A which contains
F0 . For every finite subsets F1 , F2 ⊆ A which contain F0 we have
2
X X X X
xα − xα ≤ kxα k2 − kxα k2 < .
α∈F2 α∈F1 α∈F1 ∪F2 α∈F0
P
This shows that xα is Cauchy, hence convergent [Dou, page 3]. The rest follows by
applying Theorem 16 to the Hilbert space Y of the closed linear span of {xα }. 
Exercise: Let (xj )j∈N be a sequence in a normed vector space X. Show that: (1) If
P Pk
j∈N xj converges x (in the meaning introduced on page 24) then limk→∞ j=1 xj also
Pk P
converges x. (2) If limk→∞ j=1 kxj k converges
P then j∈N xj converges (in the meaning
introduced on page 24). (3) If X = C, then j∈N xj converges if and only if x1 + x2 + · · ·
and any of its rearrangements converge. (The latter is equivalent to the convergence of
|x1 | + |x2 | + · · · [Apo-A, 8.32-3].)

24
 The closed linear span span(S) of a subset S of a normed vector space X is the
smallest closed linear subspace of X containing S. It is the closure of the set of all linear
combinations of elements of S. Here is a very useful theorem:

Theorem 18 (Closed linear span). Let S be a subset of a Hilbert space X.

(1; Hahn-Banach density theorem) The closed linear span of S is the whole space X
if and only if S ⊥ = {0}. More generally, the closed linear span of S is given ⊥⊥
Pby S .
(2) If S is orthonormal then the closed linear span of S is the set of all cα sα which
|cα |2 < ∞.
P
cα ∈ C, sα ∈ S and

Proof. (1) Set Y := span(S). We start with the first statement. Only if part is trivial.
For the if part assume x ∈ X, and let x = y + z, y ∈ Y , z ∈ Y ⊥ , be its orthogonal
decomposition. Since Y ⊥ = S ⊥ = {0}, it follows that z = 0, hence x = y ∈ Y . Now
we prove the second statement. Clearly, S ⊥⊥ is a closed linear subspace containing S, so
S ⊥⊥ ⊇ Y . For the other containment fix x ∈ S ⊥⊥ , and let x = y + z, y ∈ Y , z ∈ Y ⊥ , be
its orthogonal decomposition. We have kzk2 = hz, x − yi = 0 − 0 = 0, so z = 0, hence
x=y ∈Y.
(2) Let Y be the closed linear span of S. Then Y is a Hilbert space with orthonormal
basis S (Theorem 16.(5)), so the statement follows from Theorem 16.(4). Refer [Dou,
page 68] for a direct proof. 

Theorem 19. For every bounded operator T : X → Y between Hilbert spaces we have:
(1)
KerT = Ran⊥T∗, RanT = Ker⊥
T∗.

(2)
kT k = sup{| hT x, yi | : x, y ∈ X, kxk ≤ 1, kyk ≤ 1}.

Proof. (1) The first identity is immediate from the definitions:

x ∈ KerT ↔ T x = 0 ↔ hT x, yi = 0, ∀y ∈ Y ↔ hx, T ∗ yi = 0, ∀y ∈ Y ↔ x ∈ Ran⊥

T∗.

Replacing T by T ∗ in the first identity, taking orthogonal complement, and using

Theorem 18.(1) gives the second identity.
(2) We have
kT k = sup kT xk = sup sup | hT x, yi |,
kxk≤1 kxk≤1 kyk≤1

by the isometry part in the Riesz representation theorem. 

Exercise: Let X be a Hilbert space and Y ⊆ X be a closed linear subspace with
orthonormal
P basis {yα }. Show that the orthogonal projection P in X onto Y is given by
P x = hyα , xi yα for every x ∈ X.

Example 20. Here are some famous orthonormal bases:

1. Let S be a set. Given s ∈ S, let es : S → C be the function given by es (t) = 1 if t = s

and 0 otherwise. Then the functions es , s ∈ S, constitute an orthonormal basis for
l2 (S).

25
 q
2. The functions j+1 π
z j , j ∈ N, constitute an orthonormal basis for the Bergman space
L2a (D) on the unit disc of the complex plane.
(Proof. Orthonormality is clear by a computation in polar coordinates:
ˆ ˆ 1 ˆ 2π √ π
j k
z z dµ(z) = rj+k+1 e −12π(−j+k) dθdr = δjk ,
D 0 0 j+1
where δ is the Kronecker tensor. Given f ∈ L2a , consider its shrinked version fr (z) =
f (rz), z ∈ D, r ∈ (0, 1). Each fr is a continuous function on the closure of D,
so according to the Stone-Weierstrass theorem (Application 51), can be sufficiently
approximated by linear combinations of the basis elements in the uniform topology,
hence in the L2 topology´as well. It remains to show that fr → f in L2 as r → 1.
Decompose the integral D |f − fr |2 dµ(z) into the one over |z| ≤ 1 −  and over
1 −  < |z| < 1, where  ∈ (0, 1). The first integral approaches 0 as r → 1− and for
each , because of the uniform continuity of f on |z| ≤ 1 − . The second integral
approaches 0 as  → 1 by the Lebesgue dominated convergence theorem. The whole
argument shows that L2 is the closed linear span of our basis. (Exactly the same
argument shows that polynomials are dense in Lpa (Bm ), p ∈ [1, ∞], where Bm is the
unit openPball of Cm [Zhu-FT, page 43].) A Hilbert spaces argument [DS, page 11].
Fix f = aj z j ∈ L2a , with the uniform convergence on compacts subsets of D. We
need to verify that the same representation hold in L2 topology. Let sj be the partial
aj z j . Given r ∈ (0, 1), we have
P
sums of
ˆ j
2
X π
|sj (z)| dµ(z) = |ak |2 r2k+2 .
|z|≤r k=0
k+1

Since sj → f uniformly on |z| ≤ r, it follows that

ˆ ∞
2
X π
|f (z)| dµ(z) = |ak |2 r2k+2 .
|z|≤r k=0
k + 1

By the Lebesgue dominated convergence theorem, kf k2 = ∞ π 2

P
k=0 k+1 |ak | . With the
π
same arguments one can show that f, z k = k+1 ak . We have proved the Parseval
2
P 2
p k
identity kf k = | hf, ek i | , ek := (k + 1)/πz .)
√
3. The functions exp( −12πjx), j ∈ Z, constitute an orthonormal basis for L2 ([0, 1], dx),
called the Fourier basis [Fol, 8.20]. Here dx denotes the Lebesgue measure.
(Proof. Orthonormality is clear. Each element of L2 can be sufficiently approximated
by continuous function. By another approximation, these continuous functions can
be assumed to have equal values at the endpoints 0 and 1. √ These latter continuous
functions can be thought to live on the unit circle S1 = {exp( −12πx) : x ∈ [0, 1]}, a
compact space. By the Stone-Weierstrass theorem the finite linear combinations of our
basis elements are dense in C(S1 ) in the uniform topology, hence in the L2 topology.
The whole argument shows that L2 is the closed linear span of our basis. Remark.
Instead of the Stone-Weierstrass theorem, one can use the Cesaro summability of the
Fejer means [Apo-A, 11.15, 11.17].)

26
 √
4. The functions exp( −12πjx)1[0,1] (x − k), j, k ∈ Z, constitute an orthonormal basis
for L2 (R, dx), sometimes called the Gabor basis [Dau, page 108].

5. Consider the functions ψ : [0, 1] → R given by

(
1, 0 ≤ x < 21
ψ(x) = .
−1, 12 ≤ x < 1

Then the constant function 1 together with ψjk (x) = 2j/2 ψ (2j x − k), j ∈ N, k =
0, . . . , 2j − 1, constitute an orthonormal basis for L2 ([0, 1], dx).
(Proof. Orthonormality is clear. The key observation is that the characteristic function
of every dyadic interval [m2−n , (m + 1)2−n ), n ∈ N, m = 0, 1, . . . , 2n − 1, is a finite
linear combination of basis elements. On the other hand, L2 ([0, 1]) is the closed linear
span of dyadic characteristic functions (by the proof of [Fol, Theorem 2.10]).)

6. Consider the function ψ : R → R given by


1,
 0 ≤ x < 21
ψ(x) = −1, 12 ≤ x < 1 . (2.3)

0, otherwise


Then the functions ψjk (x) = 2j/2 ψ (2j x − k), j, k ∈ Z, constitute an orthonormal
basis for L2 (R, dx), sometimes called the Haar basis [Dau, pages 10–13]. Any
ψ ∈ L2 (R, dx) such that its dilated translations ψjk (x) = 2j/2 ψ (2j x − k), j, k ∈ Z,
constitute an orthonormal basis for L2 (R, dx) is called a wavelet. There exists smooth
wavelet functions in the literature. Wavelet theory is a rich realm of pure mathematics
with a lot of applications. A good start is [Dau].

7. In the Hilbert space L2 ([0, 1], dx), the monomials 1, x, x2 , . . . constitute a countable
dense sequence according to the Stone-Weierstrass theorem. Applying the Gram-
Schmidt process to this sequence gives an orthonormal basis, according to the proof of
Theorem 22. The j-th element of this orthonormal basis is given by dj (x2 − 2)j /dxj ,
up to a constant Cj , and is called the Legendre polynomial of degree j. 

Exercise: Consider the generator ψ of the Haar wavelet (2.3), and let χ be the char-
acteristic function of the interval [0, 1). Prove that χ(x − k), k ∈ Z, together with
ψjk (x) = 2j/2 ψ (2j x − k), j ∈ N, k ∈ Z, constitute an orthonormal basis for L2 (R, dx).
(Hint. Imitate the proof in Example 20, part 5.)

Theorem 21 (Structure theorem of Hilbert spaces). Every Hilbert space has an or-
thonormal basis, and any two orthonormal bases have the same cardinality, called the
dimension of that Hilbert space. Furthermore, two Hilbert spaces are isomorphic as
Hilbert spaces if and only if they have the same dimension.

Every Hilbert space is an l2 (X) space.

27
Proof. Consider the poset of all orthonormal subsets of a Hilbert space X, ordered by
inclusion. Every chain in this poset has an upper bound: the union of the elements of
the chain. By Zorn lemma this poset has a maximal element, which is an orthonormal
basis by Theorem 16.(1). Next, assuming two orthonormal bases S and T for X we
want to show that they have the same cardinality. The result follows from linear algebra
if at least one of S or T is finite [HK, page 44], so assume both are infinite. To any
s ∈ S one can assign the nonempty countable collection of those points t ∈ T such that
hs, ti =
6 0. (Nonemptyness is because T is an orthonormal basis and s 6= 0.) Any element
of T belongs to at least one of the collections in the range this assignment (because
S is an orthonormal basis), hence |T | ≤ |S||N| = |S|. Similarly |S| ≤ |T |. By the
Cantor-Bernstein theorem we have |S| = |T |. The rest is straightforward. 
Exercise: Show that every orthonormal set in a Hilbert space is contained in some
orthonormal basis. (Hint. Apply Zorn lemma.)
Theorem 22. A Hilbert space is separable if and only if it has countable orthonormal
basis (equivalently, its dimension is countable).
Proof. If S is a countable orthonormal basis for a Hilbert √ space X then the finite linear
combinations of elements of S with coefficients in Q + −1Q ⊆ C is dense in X. Con-
versely, assume a Hilbert space X with a countable dense sequence xj in it. Inductively
discarding those xj which belong to the linear span of x1 , . . . , xj−1 , and then refresh-
ing indices one can assume x1 , x2 , . . . to be linearly independent (in the sense of linear
algebra). Apply the Gram-Schmidt process
j−1
X hxj , xk i
y 1 = x1 , yj = xj − xk , j = 2, 3, . . . ,
k=1
hxk , xk i

to make {y1 , y2 , . . .} an orthogonal set. Since the linear span of x1 , . . . , xj equals the
linear span of y1 , . . . , yj for each j it follows that the closed linear span of y1 , y2 , . . . is the
whole X. Replacing yj by yj /kyj k gives an orthonormal basis for X. Another argument. √
If S is an orthonormal basis of a Hilbert space X then the open balls of radius 1/ 2
around elements of S are pairwise disjoint, so if S is uncountable then X can not have a
countable dense subset. 
Putting Theorem 16 and Example 20.(2) together we have:
Application 23 (Riesz-Fischer-Parseval).
´ The Fourier series of an L2 ([0, 1], dx) function
1
f converges f in L2 -norm namely 0 |f (x) − Sj (x)|2 dx → 0 as j → ∞, where Sj (x) =
P √ ´1 √
|k|≤j ck exp(2π −1kx) and ck = 0 f (x) exp(−2π −1kx)dx, k ∈ Z. Furthermore,
´1 2
|ck |2 holds. Conversely, for any sequence (ck )k∈Z
P
the Parseval identity 0 |f (x)| dx =
|ck |2 < ∞ there exists f ∈ L2 ([0, 1], dx) such that ck =
P
of
´ 1 complex numbers which
√
0
f (x) exp(−2π −1kx)dx.
P
Exercise: Let aj be a sequenceP of real numbers such that aj bj < ∞ for every
2 2
P
sequence of real numbers bj which bj < ∞. Prove that aj < ∞. (Hint. Use the
Banach-Steinhaus theorem along with the Riesz representation theorem.)

28
 Exercise: A sequence xj in a Hilbert space X is said to weakly converge x ∈ X if
hxj , yi → hx, yi for every y ∈ X. Prove that a linear map T : X → Y between Hilbert
spaces is continuous if and only if for every sequence xj in X which weakly converges x
it is the case that T xj weakly converges T x. (Hint. Use the closed graph theorem.)

29
Chapter 3

Spaces of functions II: Topological

vector spaces

References: [Rud-FA, chapter 1].

In some areas of mathematrical analysis functions spaces appear whose topology can
not be induced by a single norm, or they are not even metrizable. Functional analysts
developed a beautiful theory for a very general class of function spaces applicable in
such situations: the theory of topological vector spaces. This is the subject of this
chapter. Even a special class of these spaces (called locally convex topological vector
spaces) allows a rich duality theory. Here is a second reason to study such spaces. In the
study of normed vector spaces there are topologies (different from the one induced by the
norm) which are very useful but are rarely metrizable (Theorem 41, Example 43); these
are weak and weak-star topologies to be discussed in Chapter 5. The good news is that
these topologies are locally convex most of the time (Theorem 34). One last emphasize:
Topological vector spaces are important because on one hand they are so general to cover
most spaces appearing in mathematical analysis, and on the other hand they allow a rich
theory (a duality theory (Theorem 52) and the fundamental theorems of Banach space
theory (Remark 24).)
Let F be either the field of reals or the field of complex numbers. A topological
vector space (TVS for short) X is a F-vector space equipped with a T1 topology
(namely, singletons are closed) such that the scalar multiplication and addition operations
of vector spaces are continuous.1 For example, all normed vector spaces are such. The
most important elementary fact about TVSs is that the translation by a fixed vector as
well the multiplication by a nonzero scalar are homeomorphisms.

3.1 Elementary theory of topological vector spaces

Here are some fundamental definitions and facts of the theory. Let X be a topological
vector space, and S ⊆ X a subset.
1
Some references drop T1 separation axiom. We are following Rudin [Rud-FA]. Hausdorff axiom (T2 )
is then deduced.

30
1. A local basis (of the origin) for X is a collection of open neighborhoods of the origin
such that any other neighborhood of the origin contains at least one member of this
family. Therefore every open subset of the topology is a union of translations of some
members of a local basis.

2. S is called balanced if aS ⊆ S for every a ∈ F with |a| ≤ 1.

Exercise: Describe all balanced subsets of R2 and C as TVSs over R and C, respectively.

3. Every neighborhood U of the origin of X contains a balanced neighborhood V of the

origin such that V + V ⊆ U , or even V + V + V + V ⊆ U . Every convex neighborhood
U of the origin of X contains a balanced, convex neighborhood V of the origin.
(Proof. Since the addition of vectors is continuous at the point 0 + 0 = 0 there exist
neighborhoods V1 , V2 of 0 such V1 + V2 ⊆ U . By the continuity of the scalar multipli-
cation in X there exists a neighborhood V10 of 0 such that aV10 ⊆ V1 for every |a| ≤ 
00 0
S
and some  > 0. Then V1 := |a|≤ aV1 is a balanced neighborhood of 0 contained in
V1 . Similarly construct V200 . Then V := V100 ∩ V200 is a balanced neighborhood of 0 such
that V + V ⊆ U . To get T V + V + V + V ⊆ U repeat the process. Now assume that
U is convex. Set A := |a|=1 aU and let V be a balanced neighborhood of the origin
such that V ⊆ U . Since V is balanced if follows that V ⊆ A, so the interior int(A)
of A is a neighborhood of 0. Being an intersection of convex sets, A is convex, hence
so is int(A). It remains to show that T int(A) is balanced.
T For every 0 < r < 1 and
b ∈ F of modulus 1 we have rbA = |a|=1 rbaU = |a|=1 raU . Since aU is a convex set
containing 0 it also contains raU . Therefore rbA ⊆ A.)

4. X has a balanced local basis. If X is locally convex (to be defined in (17)) then X
has a balanced convex local basis.
(Proof. Immediate from (3).)

5. X is Hausdorff.
(Proof. Assume a nonzero element x ∈ X. Since 0 belongs to the open subset X \
{x} ⊆ X, there exists a neighborhood U of 0 which does not contain x. Find a
balanced neighborhood V of 0 such that V + V ⊆ U . Then V and x + V are disjoint
neighborhoods of 0 and x, respectively.)

6. Here is a stronger separation result that will be needed for proving the most important
result of this course (Theorem 25.(4)): For every compact K ⊆ X and closed C ⊆ X
which are disjoint one can find neighborhood U of the origin such that (K + U ) ∩ (C +
U ) = ∅.
(Proof. For every x ∈ K, since x 6∈ C, one can find balanced neighborhood Ux of 0 such
that (x + Ux + Ux + Ux + Ux ) ∩ C = ∅, which implies (x + Ux + Ux +TUx ) ∩ (C + Ux ) = ∅.
Compact K S can be covered by S finitely many xj + Uxj . Then U := Uxj works because
K + U ⊆ xj + Uxj + U ⊆ xj + Uxj + Uxj .)

7. The closure of S equals the intersection of all S + U where U ranges over a local basis.
The closure of a linear subspace is a closed linear subspace.

31
 (Proof. If x ∈ S then since x − U is T
a neighborhood of x it follows that x − U intersects
S, so x ∈ S + U . Conversely, if x ∈ S + U , since for every neighborhood V of 0 there
exists some U such that U ⊆ −V it follows that x ∈ S + U ⊆ S − V , namely x + V
intersects S. For the second statement let Y ⊆ X be a linear subspace of X. Assume
x ∈ Y . For every neighborhood U of the origin we have x + u = y for some u ∈ U
and y ∈ Y , so ax + au = ay for every a ∈ F, hence ax ∈ Y . Assume x1 , x2 ∈ Y . For
every neighborhood V of x1 + x2 , by the continuity of the addition of vectors, there
exists neighborhoods U1 of x1 and U2 of x2 such that U1 + U2 ⊆ V . Y intersects both
U1 and U2 , so it intersects U .)

8. S is called a bounded if for every neighborhood V of the origin there exists r > 0
such that S ⊆ sV for every s > r. Since every neighborhood of the origin contains a
balanced one it follows that an equivalent definition of boundedness is that for every
neighborhood V of the origin there exists r > 0 such that S ⊆ rV . For a general
metrizable space X this notion of boundedness is not equivalent with the famous one:
There exists C > 0 such that d(x, y) < C for every x, y ∈ S. (Refer [Rud-FA, page
23] for a counterexample.)
S
9. Let U be a neighborhood of the origin. Then X = j∈N jU . If U is bounded then
{j −1 U : j = 1, 2, . . .} is a local basis.
(Proof. For every x ∈ X, by the continuity of the scalar multiplication at 0 × x = 0,
there exists  > 0 such that ax ∈ U whenever |a| < . Therefore x ∈ jU for every
integer j > 1/. For the second statement, assuming a neighborhood V of 0, there
exists r > 0 such that U ⊆ sV for every s > r. Therefore j −1 U ⊆ V for every integer
j > r.)
Exercise: Why boundedness is needed in the latter statement?

10. A linear map T : X → Y between TVSs is called bounded if it maps bounded subsets
to bounded ones. It is straightforward to check that continuity implies boundedness,
but the converse is not true [Rud-FA, page 39, exercise 13]. If X is metrizable then
T is bounded if and only if it is continuous if and only if T xj → 0 for every sequence
xj → 0 [Rud-FA, 1.32].

11. A linear functional α : X → F on X is continuous if and only if Kerα is closed.

(Proof. The only if part is due to Kerα = α−1 (0). Conversely, putting the trivial case
α ≡ 0 aside, fix x ∈ X \ Kerα . There exists a balanced neighborhood U of 0 such that
α never vanishes on x + U . By linearity α(U ) ⊆ F is balanced. α(U ) 6= F because
otherwise α(y) = −α(x) for some y ∈ U , hence the contradiction x + y ∈ Kerα ∩ x + U .
Therefore α(U ) contains an open ball around 0. Continuity follows.)

12. Recall that a topological space is metrizable if its topology is induced by a metric.
X is metrizable if and only if it has a countable local basis; and if so then the metric
can be taken to be translation invariant, namely d(x, y) = d(x + z, y + z) for every
x, y, z [Rud-FA, 1.24].

13. S is called absorbing if for every x ∈ X there exists r > 0 such that x ∈ rS.

32
 Exercise: Give examples of an absorbing subset which is not balanced, and a balanced
subset which is not absorbing.

14. Seminorms µ on a vector space X correspond (not necessarily in a one-to-to-one fash-

ion) to convex absorbing balanced subsets S ⊆ X via:

µ 7→ Sµ := {x ∈ X : µ(x) < 1},

S 7→ µS , µS (x) = inf {x ∈ tS}.

t>0

µS is called the Minkowski (or Gauge) functional associated to S. The intuition

about the gauge functional comes from the observation that if S is the unit ball of a
normed vector space X then µS retrieves the norm. If X is a TVS then under the
same maps, continuous seminorms are in one-to-to-one correspondence with convex
balanced neighborhood of the origin, namely S = SµS and µ = µSµ . In this situation
S = SµS = {µS < 1} reveals the usefulness of the gauge functional: It makes convex
balanced neighborhoods of the origin look like the open unit ball.
(Proof. All these statements are straightforward to check [Rud-FA, 1.34-6]. Let us just
check that if S is a convex balanced neighborhood of the origin then S = {µS < 1}. S
is absorbing by (9). If x ∈ S by openness of S there exists t < 1 such that x ∈ tS, so
µS (x) < 1. If x 6∈ S then the set {t > 0 : x ∈ tS} does not contain 1, so its infimum
must by ≥ 1, because by convexity and absorbingness of S, for every x ∈ X the set
{t > 0 : x ∈ tS} has the property that if it contains t then it contains [t, ∞).)

15. Let pα : X → [0, ∞), α ∈ A, be a family of seminorms on a vector space X. Then

the weakest (namely smallest) topology such that all pα are continuous is the one with
subbasis Uα,x, = {y ∈ X : pα (x − y) < }, α ∈ A, x ∈ X,  > 0. Equivalently,
a net (xi )i∈I in X converges to x in this topology if and only if pα (xi − x) → 0 for
every α ∈ A. (Refer to page 48 for a review of the corresponding notions from general
topology.) This family of seminorms is called separating if the topology it generates
is Hausdorff. This happens if and only if there is no nonzero x ∈ X such that all pα (x)
vanish.

16. Let T : X → Y be a linear map between TVSs X and Y whose topologies are induced,
respectively, by the families of seminorms {pα }α∈A and {pβ }β∈B . Then T is continuous
if and only ifPfor every β ∈ B there exists C > 0 and a finite subset F ⊆ A such that
pβ (T x) ≤ C α∈F pα (x) for every x ∈ X.
(Proof. Assume that T is continuous. For every β ∈ B, since F −1 {y ∈ Y : pβ (y) < 1}
is open and contains 0, it follows that there exists finitely many α1 , . . . , αn in A and
 > 0 such that pβ (T x) < 1 whenever pαj (x) <  for all j. We assert that C := 1/
works. Fix x ∈ X. If all pαj (x) vanish then all pαj (rx) vanish for every r > 0, hence
pβ (T x) <
P1/r, which only happens if pβ (T x) =
P0. If at least one pα (x) is nonzero then
pαj (x/ pαj (x)) <  for all j, hence pβ (x/ pαj (x)) < 1. If part is trivial.)

17. X is called a locally convex (LCTVS for short) if it has a local basis whose members
are convex. There is another equivalent definition which is the most common way
these spaces appear in practice [Rud-FA, 27-29]: X is locally convex if and only

33
 if its topology is induced by a separating family of seminorms pα , α ∈ A, on X,
namely a local basis of the origin is given by finite intersections of sets of the form
Uα, := {x ∈ X : pα (x) < }, α ∈ A,  > 0. (Of course  > 0 can be replaced
by 1/j, j = 1, 2, . . ..) If part is easy to verify. Specially, T1 axiom comes from
“separating”, and local convexity is because: If x, y ∈ Uα, then for every λ ∈ (0, 1) we
have pα (λx+(1−λ)y) ≤ λpα (x)+(1−λ)pα (y) < , hence λx+(1−λ)y ∈ Uα, . For only
if part, assuming a convex local basis for X, first construct a balanced convex local
basis {Uα }α∈A by (4); then the gauge functionals µUα , α ∈ A, constitute a separating
family of seminorms which induces the topology of X [Rud-FA, pages 27-29]. Here is a
useful fact: If the topology of X is given by a separating family of seminorms pα , α ∈ A,
then a subset E ⊆ X is bounded if and only of each pα is bounded on E in the sense
that there exist Mα > 0 such that pα (x) < Mα for every x ∈ E and every α ∈ A. If
2−j p (x−y)
pj , j ∈ N, is a countable separating family of seminorms then d(x, y) = max 1+pjj(x−y)
is a translation-invariant metric which induces the same topology [Rud-FA, 1.38.(c)].

LCTVSs ↔ separating families of seminorms ,

metrizable LCTVSs ↔ countable separating families of seminorms .

18. X is called locally bounded if it has a neighborhood of the origin which is bounded.

19. X is called normable if its topology is induced by a norm. This happens if and only
if X is locally bounded and locally convex [Rud-FA, 1.39].

20. X is called complete if every Cauchy net converges. (A net (xα )α∈A in X is called
Cauchy if for every neighborhood U of the origin there exists α0 ∈ A such that
xα − xβ ∈ U for every α, β ≥ α0 .) If X is first-countable then X is complete if every
Cauchy sequence converges.

21. X is called an F -space if the topology is complete and given by a translation-invariant

metric.

22. X is called a Frechet space if it is a locally convex F -space. Equivalently, the

topology of X is complete and given by a countable separating family of seminorms.

23. Every compact subset A ⊆ X is closed and bounded. (Proof. Closedness is because X
is Hausdorff [Mun, 26.3]. For boundedness assume a neighborhood U of 0. Since every
neighborhood
S of 0 contains a smaller balanced one we can assume U is balanced. By
(9), A ⊆ j∈N jU . The compactness of A and balance of U implies that A ⊆ jU for
some j.) X is said to have the Heine-Borel property if every closed and bounded
subset is compact.

24. Let Y a closed linear subspace of X. Then the vector space X/Y := {x + Y : x ∈ X}
of cosets of Y equipped with the quotient topology turn out to be a TVS, called the
quotient space, and the quotient map π : X → X/Y , x 7→ x + Y , is a continuous
linear map. (Interpret x + Y as the set {x + y : y ∈ Y }, therefore two cosets x + Y
and x0 + Y are equal if and only it x − x0 ∈ Y . Vector space operations are defined

34
 naturally: (x + Y ) + (x0 + Y ) = (x + x0 ) + Y and a(x + Y ) = ax + Y , for every x, x0 ∈ X
and a ∈ F. X/Y is topologized by declaring A ⊆ X/Y to be open if and only if π −1 (A)
is open is X.) Note that X/Y is satisfies T1 axiom because the preimage under π of a
singleton subset {x + Y } of X/Y , being equal to x + Y , is closed in X. Here are some
useful facts with straightforward proofs [Rud-FA, 1.41]:

• π is open, namely maps open subsets to open ones.

• If {Uα } is a local basis for X then {π(Uα )} is a local basis for X/Y .
• If X satisfies any of the following properties then X/Y satisfies the same: local
convexity, local boundedness, metrizability, normability, F -space, Frechet space,
Banach space.

Some remarks about the proof:

• If X is normable by k − kX then X/Y is normable by kx + Y k = inf{kx − ykX :

y ∈ Y }, the distance of x to Y .
• If X is metrizable by invariant metric dX then X/Y is metrizable by invariant
metric d(x + Y, x0 + Y ) = inf{dX (x − x0 , y) : y ∈ Y }.

25. If X is locally bounded then it has a countable local basis (equivalently, metrizable).
If X is locally bounded and has the Heine-Borel property then it is locally compact.
(Proof. Let U be a bounded neighborhood of 0. Then j −1 U , j = 1, 2, . . ., is a count-
able local basis. For the second statement it suffices to show that U is bounded,
because then by the Heine-Borel property U is also compact. Let V be an arbitrary
neighborhood of 0. Since U is bounded it follows that there exists r > 0 such that
U ⊆ sV if s > r. Assuming x ∈ U , since x − V is a neighborhood of x it follows
that x − v = u = rv 0 for some v, v 0 ∈ V , u ∈ U , hence x = v + sv 0 . This shows that
U ⊆ (s + 1)V .)

26. If X has finite vector space dimension then there is exactly one topology on X which
makes it a TVS.
(Proof. Assume a vector space isomorphism F : Fm → X.PWith respect to a basis
(e1 , . . . , em ) for Fm the mapping F acts by (a1 , . . . , am ) 7→ aj F (ej ), hence continu-
ous. Let B be the open unit ball of Fm and S be the boundary of B. K := F (S) is
compact and does not contain 0. By compactness there exists a (balanced) neighbor-
hood V of 0 such that V and K do not intersect. Since F is bijective it follows that
U := F −1 (V ) and S do not intersect. Since F is linear if follows that U is balanced,
hence connected. Therefore U ⊆ B (otherwise U intersects S), which clearly implies
that F −1 is continuous at the origin, hence everywhere by linearity.)

27. If Y and Z are, respectively, a closed and finite dimensional linear subspace of X then
Y + Z is closed.
(Proof. First we reduce to Y = 0 case. Consider the quotient π : X → X/Y , and
observe that Y + Z = π −1 (π(Z)). π(Z) is finite dimensional because π being linear
maps linear spanning sets of Z to linear spanning sets of π(Z). Therefore we only need

35
 to verify that finite dimensional linear subspaces of TVSs are closed, so assume Y = 0.
Fix a vector space isomorphism F : Fm → P Z. With respect to a basis (e1 , . . . , em ) for
m
F the mapping F acts by (a1 , . . . , am ) 7→ aj F (ej ), hence continuous. Let B be the
open unit ball of Fm and S be the boundary of B. K := F (S) is compact and does not
contain 0. By compactness there exists a (balanced) neighborhood V of 0 in X such
that V and K do not intersect. Since F is bijective it follows that U := F −1 (V ∩Z) and
S do not intersect. Since F is linear if follows that U is balanced, hence connected.
Therefore U ⊆ B (otherwise
S U intersects S), or equivalently V ∩ Z ⊆ F (B). Fix
p ∈ Z. Since X = j∈N jV there exists positive integer j such that p also belong
to open jV ⊆ X. Therefore p belongs to the closure of jV ∩ Z ⊆ F (jB) ⊆ F (jB).
However F (jB) is compact, hence closed in X. Therefore p ∈ F (jB) ⊆ Z.)
closed + finite dimensional → closed .

28. Recall the notion of local compactness from general topology [Mun, Theorem 29.2]:
X is locally compact if and only if it has a neighborhood of the origin whose closure
is compact. This happens if and only if the vector space dimension of X is finite.
(Proof. If X is of finite dimension n then by (26) X is homeomorphic to the Euclidean
space Rn , so the closed balls are compact by Heine-Borel theorem. Conversely, let U
be a neighborhood of 0 in X such that U is compact. S By compactness one can find
1
finitely many points x1 , . . . , xn in X such that U ⊆ xj + 2 U . Let Y be the linear
span of {x1 , . . . , xn }. Since U ⊆ Y + 12 U and Y = aY for every T nonzero scalar a, it
follows that U ⊆ Y + 12 Y + 14 U = Y + 41 U , so by repetition U ⊆ j≥1 Y + 2−j U . Since
2−j U constitute a local basis it follows
S that U ⊆ Y . Since Y is closed by (27) it follows
that U ⊆ Y . Therefore Y contains j≥1 jU = X, hence Y = X.)

locally compact ↔ finite dimensional .

3.2 Some famous examples of Frechet spaces

Here are some famous example of Frechet spaces appearing in different branches of math-
ematical analysis. None of these examples are normable, which gives a motivation to
study TVS as a generalization of normed vector spaces.
1. X := C(U ), U ⊆ Rn nonempty open, the vector space of all continuous functions
f : U → C. It is topologized via the separating family of seminorms
pj (f ) = sup{|f (x)| : x ∈ Kj }, j ∈ N,
where Kj is an exhaustion of U by compacts [Lee,
S A.60], namely each Kj is compact
and contained in the interior of Kj+1 and U = Kj . (Different exhaustions induce the
same topology) This is called the topology of uniform convergence on compacts.
(Also equal to the “compact-open topology” [Mun, 46.8].) We will show that:
C(U ) is Frechet, not locally bounded, not normable.
We check these properties step by step:

36
 • A local subbasis is given by Ujk = {pj (f ) < 1/k}, but since p1 ≤ p2 ≤ · · · , a
local basis is given by Uj = {pj (f ) < 1/j}.
• Frechetness. Since the topology is given by a countable separating family of
seminorms it follows that X is locally convex and metrizable. A compatible
2−j p (f −g)
invariant metric is d(f, g) = max 1+pjj(f −g) . For completeness assume a Cauchy
sequence fα in X. This means pj (fα − fβ ) → 0 as α, β → ∞ for every j, so
appears the limit function f : U → C such that fα → f uniformly on compacts.
f is continuous because continuity can be checked locally and on each Kj the
function f is a uniform limit of continuous functions.
• Not locally bounded. We show that no Uj is bounded. A ⊆ X being bounded
means that all pk are bounded on A, namely there exist Mk > 0 such that
|f | < Mk on Kk for every f ∈ A. However every Uj contains (even C ∞ ) functions
with arbitrary large pj+1 [Lee, 2.25].
• Not normable. Because X is not locally bounded.

Exercise: Show that C(R) does not have the Heine-Borel property. (Hint. Consider
the subset {f ∈ C(R) : −1 ≤ f ≤ 1}, and think about the sequence fj = sin(jx).)

2. X := Holo(D), D ⊆ Cm nonempty open, the vector space of all holomorphic (also

called complex analytic) functions D → C. We accept the subspace topology X ⊆
C(D), where D is pretended to be an open subset of R2m in a natural way. We will
show that:

Holo(D) is Frechet, has Heine-Borel property, not locally bounded, not normable.

We check these properties step by step:

• Completeness. By Weierstrass theorem [Ahl, page 176][Hör-SCV, 1.2.5, 2.2.4] X

is a closed subspace of C(D), so complete.
• Heine-Borel property. Let A ⊆ X be a closed and bounded subset. A being
bounded means that there exists Mj > 0 such that |f | < Mj on Kj for every
f ∈ A and every j ∈ N. (Here Kj is an exhaustion of D by compacts.) By Mon-
tel’s compactness theorem [Ahl, pages 224-5][Hör-SCV, 1.2.6, 2.2.5] (which is the
adaption of Arzela-Ascoli theorem to the holomorphic setting) every sequence in
A has a subsequence which converges uniformly on compacts; that limit function
is holomorphic by Weierstrass theorem. Since A is closed it contains the limit
function. This argument shows that A is sequentially compact, hence compact
because X is metrizable.
• Not locally bounded. If X was locally bounded, because it has Heine-Borel prop-
erty, it should have been locally compact, so of finite vector space dimension.
However the monomials z1α1 · · · zm αm
, (α1 , . . . , αm ) ∈ Nm , are in X and linearly
independent. (Here z1 , . . . , zm are standard coordinate functions of Cm .)
• Not normable. Because X is not locally bounded.

37
3. X := E(U ), U ⊆ Rn nonempty open, the vector space of all C ∞ functions U → C. It
is topologized via the separating family of seminorms

pj (f ) = sup{|∂ α f (x)| : x ∈ Kj , |α| ≤ j}, j ∈ N,

where Kj is an exhaustion of U by compacts, and we are using the multi-index nota-

tions (page 3). We will show that:

E(U ) is Frechet, has Heine-Borel property, not locally bounded, not normable.

We check these properties step by step:

• A local basis is given by Uj = {pj (f ) < 1/j}.

• Frechetness. Since the topology is given by a countable separating family of
seminorms it follows that X is locally convex and metrizable. A compatible
2−j p (f −g)
invariant metric is d(f, g) = max 1+pjj(f −g) . For completeness assume a Cauchy
sequence fa in X. This means that pj (fa − fb ) → 0 as a, b → ∞ for each
j, so each ∂ α fa converges uniformly on compacts to some function gα . Since
fa → g0 pointwisely (here 0 means multi-index (0, . . . , 0)), it follows by a simple
application of the mean value theorem for differentiation [Apo-A, 9.13] that each
gα is smooth and equals ∂ α g0 , and that fa → g0 in X.
• Heine-Borel property. Let A ⊆ X be a closed and bounded subset. A being
bounded means that there exists Mj > 0 such that |∂ α f | < Mj on Kj for every
f ∈ A, every α ∈ Nn with |α| ≤ j and every j ∈ N. For every j the inequality
|∂ α f | < Mj valid on Kj for |α| ≤ j, along with a simple application of the mean
value theorem for differentiation, shows that {∂ β f : f ∈ A} is equicontinuous on
Kj−1 for every β ∈ Nn with |β| ≤ j − 1; therefore by Arzela-Ascoli theorem [Fol,
4.43] every sequence in A has a subsequence which converges with respect to pj .
By a Cantor diagonal argument one can deduce that every sequence in A has a
subsequence which converges with respect to every pj , namely in the topology of
X. Since A is closed it contains that limit function. This argument shows that
A is sequentially compact, hence compact because X is metrizable.
• Not locally bounded. If X was locally bounded, because it has Heine-Borel prop-
erty, it should have been locally compact, so of finite vector space dimension.
However the monomials xα1 1 · · · xαnn , (α1 , . . . , αn ) ∈ Nn , are in X and linearly
independent. (Here x1 , . . . , xn are standard coordinate functions of Cn .)
• Not normable. Because it is not locally bounded.

4. X := EK (U ), U ⊆ Rn open subset, K ⊆ U compact, the vector space of all C ∞

functions whose support {x ∈ U : f (x) 6= 0} is contained in K. We accept the subspace
topology X ⊆ E(D). Exactly the same as in the case E(U ) one can show that:

EK (U ), intK 6= ∅, is Frechet, has Heine-Borel property, not locally bounded,

not normable.

38
 That the vector space dimension of EK (U ) is infinite follows from the existence of
smooth bump function [Lee, 2.25]. We assumed that the interior of K is nonempty
just because otherwise EK (U ) = {0}.

5. S, the Schwartz space, the vector space of all C ∞ functions f : Rn → C which,

together
p with all their derivatives, vanish at infinity faster that any power of |x| =
x1 + · · · + x2n , more precisely, supx∈Rn |x|j |∂ α f (x)| < ∞ for every j ∈ N and α ∈ Nn .
2

Some authors call them rapidly decreasing functions functions. This space is
used extensively in Fourier analysis and distribution theory. It is topologized via the
separating family of seminorms

pj (f ) = sup (1 + |x|2 )j |∂ α f (x)| : x ∈ Rn , α ∈ Nn , |α| ≤ j , j ∈ N.



(Replacing (1 + |x|2 )j with |x|j leads to the same topology.)

S is Frechet, has Heine-Borel property, not locally bounded, not normable.

The proof is exactly the same as the one for E(U ). The vector space dimension of S
is infinite because xα exp(−|x|2 ), α ∈ Nn , are linearly independent members of S.

Exercise: Let X be a TVS whose topology is given by a countable separating family

of seminorms pj , j ∈ N. Also assume that p1 ≤ p2 ≤ . . .. (Note that this assumption
holds for all examples of this section.) Give a direct argument that if X is normable
then all pj for sufficiently large j are equivalent. Use this proposition to prove that the
Schwarz space S is not normable. (Hint. Review the proof of 16.)

3.3 An F -space which supports no continuous linear

functional
Let X := Lp ([0, 1]), p ∈ (0, 1), be the space of functions [0, 1] → C which are Lp -integrable
´
with respect to the Lebesgue measure. We topologize it via the metric d(f, g) = |f −g|p .
(Triangle inequality follow from ap + bp ≥ (a + b)p valid for a, b ≥ 0.)

Lp ([0, 1]), p ∈ (0, 1), is F -space, locally bounded, not locally convex (X and ∅ are
the only convex opens), not normable, with no nonzero continuous linear functional.

We check these properties step by step:

• Completeness. Adapt the proof of the completeness of Lp , p ∈ [1, ∞).

• Locally bounded. Ur := {f ∈ X : d(f, 0) < r}, r > 0, is a local basis, and

U1 = r−1/p Ur . It follows that U1 is bounded.

• Not normable. Because it is not locally convex.

39
 • X and ∅ are the only convex opens. Assuming a nonempty convex open V ⊆ X we
show that V = X. We can assume that V contains the origin. Since V is open it
follows that V ⊇ Ur for some r > 0. Fix some f ∈ X. Since 0 < p < 1 one can find
n ∈ N large enough such that np−1 d(f, 0) < r. By ´continuity one can inductively
xj
find x0 = 0 < x1 < . . . < xn−1 < xn = 1 such that xj−1 |f |p = d(f, 0)/n for every
j = 1, . . . , n. Set gj := nf 1(xj−1 ,xj ] . Then
ˆ xj
d(gj , 0) = |f |p np = np−1 d(f, 0) < r,
xj−1

so gj ∈ Ur ⊆ V . On the other hand f = (g1 + · · · + gn )/n. Since gj belongs to

convex V it follows that f ∈ V .

• Let α be a continuous linear functional. B := {a ∈ F : |a| < 1} is open and convex

for every  > 0, so is α−1 (B ) ⊆ X, therefore α−1 (B ) = X. This means that α ≡ 0.
is open and convex, hence equals X be the previous part.

Remark 24. Recall the fundamentals theorems of Banach spaces: Uniform boundedness
principle, Open mapping theorem, Inverse mapping theorem and closed graph theorem
(Theorem 5). They have analogues for TVSs. Statements of the last three are exactly
as before unless “Banach space” is replaced with “F -space” [Rud-FA, 2.12, 2.15]. Here
is the statement for the uniform boundedness principle [Rud-FA, 2.4, 2.6]: A family
Tα : X → Y , α ∈ A, of bounded operators from F -space X to TVS Y is uniformly
equibounded (namely for every bounded subset B ⊆ X there exists a bounded subset
B 0 ⊆ Y such that Tα (B) ⊆ B 0 for every α ∈ A) if and only if it is pointwisely equibounded
(namely {Tα (x) : α ∈ A} is bounded for every x ∈ X). Another good reference is [DS,
chapter 2]. 

40
Chapter 4

Duality theory I: Hahn-Banach

theorem

References: [Rud-FA, chapter 3][Bre, chapter 1].

Theorem 25 (Hahn-Banach). The followings are true:

1- Controlled extension from linear subspaces of real vector spaces. Let X be a real
vector space, Y a linear subspace, p : X → R a sublinear functional (namely, p(x+y) ≤
p(x)+p(y) and p(ax) = ap(x) for every x, y ∈ X and a ≥ 0; These are called subadditivity
and positive homogeneity. For example every seminorm is sublinear.), and f a linear
functional on Y which is dominated by p (namely, f (x) ≤ p(x) for every x ∈ Y .). Then
f can be extended to a linear functional F on X which is still dominated by p.
2- Controlled extension from linear subspaces of real or complex vector spaces. Let X
be a vector space, Y a linear subspace, p : X → [0, ∞) a seminorm, and f a linear func-
tional on Y with is dominated by p (namely, |f (x)| ≤ p(x) for every x ∈ Y ). Then f can
be extended to a linear functional F on X which is still dominated by p.
20 - Continuous extension from linear subspaces of normed vector spaces. Every con-
tinuous linear functional on a linear subspace of a normed vector space can be extended
to a continuous linear functional on the whole space with the same norm.
3- Separation theorem for normed vector spaces. Let X be a normed vector space, Y
a closed linear subspace and x ∈ X \ Y . Then there exists a continuous linear functional
F on X such that F |Y ≡ 0, F (x) = inf y∈Y kx − yk > 0 and kF k = 1. In geometric terms,
in a normed vector space X, a point x and a closed linear subspace Y that are disjoint
can be strictly separated by closed hyperplanes in the sense that there exists a continuous
liner functional F on X and K ∈ R such that ReF (y) < K < ReF (x) for every y ∈ Y .
Specially, X ∗ separate points on X in the sense that for every two distinct points x, y ∈ X
there exits F ∈ X ∗ such that F (x) 6= F (y).
30 - Closure theorem. Let X be a normed vector space, Y a linear subspace and x ∈ X.
Then x ∈ Y if and only if every continuous linear functional on X which vanishes on
Y also vanishes on x. In other words, Y = ⊥ Y ⊥ where the annihilators of subsets
A ⊆ X and B ⊆ X ∗ are defined by A⊥ = {α ∈ X ∗ : α(x) = 0, ∀x ∈ A} and ⊥ B = {x ∈

41
X : α(x) = 0, ∀α ∈ B}. Specially, Y is dense in X if and only if there is no nonzero
continuous functional on X which vanishes on Y (in notations: Y ⊥ = {0}.)
4- Separation theorem for LCTVSs; generalization of (3). Let A and B be two dis-
joint nonempty convex subsets of a TVS X. If A is open then A and B can be separated
by closed hyperplanes in the sense that there exists F ∈ X ∗ and K ∈ R such that
ReF (a) < K ≤ ReF (b) for every a ∈ A and b ∈ B. If A is compact, B is closed and X is
locally convex then A and B can be strictly separated by closed hyperplanes in the
sense that there exists F ∈ X ∗ and K1 , K2 ∈ R such that ReF (a) < K1 < K2 < ReF (b)
for every a ∈ A and b ∈ B. Specially, if X is a LCTVS then X ∗ separate points on X
in the sense that for every two distinct points x, y ∈ X there exits F ∈ X ∗ such that
F (x) 6= F (y).
40 - Closure theorem; generalization of (30 ). The analogues of (30 ) holds if “normed
vector space” is replaced by “LCTVS”.
5- Continuous extension from linear subspaces of LCTVSs. Every continuous linear
functional on a linear subspace of a LCTVS can be extended to a continuous linear func-
tional on the whole space.

A linear subspace Y of a normed vector space X is dense if and only there

is no nozero continuous linear functional on X which vanishes on Y .

Two disjoint nonempty closed convex subsets of a LCTVS, at least

one of them compact, can be strictly separated by closed hyperplanes.

Proof. (1) Consider the set of all pairs (Y1 , f1 ) where f1 is a linear extensions of f to a
linear subspace Y1 ⊆ X containing Y which is dominated by p, and partially order it by
declaring (Y1 , f1 ) ≤ (Y2 , f2 ) if and only if Y1 ⊆ Y2 and
S f1 = f2 |Y1 . In this poset each chain
(Yα , fα ) has an upper bound: (Z, g) where Z = Yα and g|Yα = fα . By Zorn lemma
there is a maximal element (Z, F ). We will show that Z = X. Assuming some x ∈ X \ Z
we need to refute maximality of (Z, F ), namely find a linear extension of F to Z + Rx
which is dominated by p. This is equivalent to finding some b ∈ R such that setting
F (x) = b then we have F (z) + λb = F (z + λx) ≤ p(z + λx) for every z ∈ Z and λ ∈ R.
This happens exactly when
p(z1 − λ1 x) − F (z1 ) p(z2 + λ2 x) − F (z2 )
≤ , ∀z1 , z2 ∈ Z, ∀λ1 , λ2 ∈ R+ ,
−λ1 λ2
which can be rewritten as

λ2 F (z1 ) + λ1 F (z2 ) ≤ p(λ1 z2 + λ1 λ2 x) + p(λ2 z1 − λ1 λ2 x), ∀z1 , z2 ∈ Z, ∀λ1 , λ2 ∈ R+ .

This holds by the sublinearity of p and that it dominates F on Z.

(2) Assuming the real case, note that f (x) ≤ p(x) for every x is equivalent to |f (x)| ≤
p(x) for every x, by replacing
√ x with −x. Therefore (1) can be applied. Now the complex
case. Let f = g + −1h be the √ decomposition
√ of f into real-valued functionals√ g, h.
Then g, h are R-linear, and f ( −1x) = −1f (x) is equivalent to h(x) = −g( −1x).
Since g(x) ≤ |f (x)| ≤ p(x) on Y it follows by (1) that g can be extended √ to an R-linear
√
functional G on X such that G(x) ≤ p(x) on X. Clearly, F (x) = G(x) − −1G( −1x)

42
is a C-linear functional on X which√extends f . It remains to check that |F (x)| ≤ p(x).
√ x, and replacing x with exp( −1θ)x, θ ∈ R, in G(x) ≤ p(x) we have 2G(x) cos θ2 +
Fixing
G(√−1x) sin θ ≤ p(x). Since this is true for every θ it follows that |F (x)| = G(x) +
G( −1x)2 ≤ p(x)2 .
(20 ) Immediate from (2) by taking p to be the norm of X multiplied by kf k.
(3) Since Y is closed and does not contain x it follows that the distance δ := inf y∈Y kx−
yk between x and Y is strictly positive. The linear functional f : Y +Fx → F, y+ax 7→ aδ,
has norm equal to one:

|f (y + ax)| = |a|δ ≤ |a|ky/a + xk ≤ ky + axk, ∀y ∈ Y.

By (20 ), f can be extended to a continuous linear functional F on X with the same norm,
namely 1. F has the other two desired properties.
(30 ) Only if part is trivial. If x ∈ X \ Y then by (3) there exists a continuous linear
functional F on X which vanishes on Y (hence on Y ) but not at x.
(4) We assume F = R, because the complex case is then deduced having in mind the
correspondence between real √ and complex
√ functionals that already appeared in the proof
of (2): f (x) = Ref (x) − −1Ref ( −1x). For the first statement assume A open. Fix
a0 ∈ A, b0 ∈ B, and set x0 := b0 − a0 , C := A − B + x0 . Then C is a convex neighborhood
of the origin in X. Note that:

• C is absorbing because it is a neighborhood of the origin (Section 3.1.(9)).

• Since C is convex and absorbing it follows that for every x ∈ X the set {t > 0 : x ∈
tC} has the property that if it contains t then it contains [t, ∞).

• Just because C is convex and absorbing one can easily check that its gauge func-
tional p = µC : X → [0, ∞), x 7→ inf{t > 0 : x ∈ tC}, is sublinear [Rud-FA,
1.35].

• Since C is also open we have C = {x ∈ X : p(x) < 1} (Section 3.1.(14)).

• A ∩ B = ∅ implies that x0 6∈ C, so p(x0 ) ≥ 1.

Consider the linear functional f : Rx0 → R, tx0 7→ t. f is dominated by p because

if t > 0 then f (tx0 ) = t ≤ tp(x0 ) = p(tx0 ), and if t < 0 then f (tx0 ) = t < 0 ≤ p(tx0 ).
By (2), f can be extended to a linear functional F on X dominated by p. Specially,
F ≤ p < 1 on C, hence by linearity F > −1 on −C, so −1 < F < 1 on the neighborhood
C ∩ −C of 0. The continuity of F follows. For any a ∈ A and b ∈ B we have

F (a) − F (b) + 1 = F (a − b + x0 ) ≤ p(a − b + x0 ) < 1,

because a − b + x0 ∈ C. This gives F (a) < F (b). It follows that F (A) and F (B) are
disjoint convex subsets of R, with F (A) to the left of F (B). On the other hand F (A)
is open because every nonzero continuous linear functional on a TVS is clearly an open
map. Therefore K := sup F (A) works.
For the second statement, first using Section 3.1.(6) find neighborhood U of 0 such
that (A + U ) ∩ B = ∅. Since X is LCTVS one can assume that U is convex. Applying

43
the previous part to separate A + U and B, one finds F ∈ X ∗ such that F (A + U ) and
F (B) are disjoint convex subsets of R with F (A + U ) open and to the left of F (B). We
are done because F (A) is a compact subset of F (A + U ).
(40 ) Only if part is trivial. If x ∈ X \ Y then by (4) there exists a continuous linear
functional F on X such that ReF (x) < K < ReF (z) for every z ∈ Y . Since Y is closed
under scalar multiplication it follows that F |Y ≡ 0 (hence F |Y ≡ 0); however F (x) 6= 0.
(5) Assume continuous linear functional f on linear subspace Y of LCTVS X. Putting
the trivial case f ≡ 0 aside, one can find y ∈ Y with f (y) = 1. By continuity, y
does not belong to the closure of B := Kerf in X, so by (4) there exists F ∈ X ∗ and
K ∈ R such that ReF (y) < K < ReF (b) for every b ∈ B. Since F (B) is a linear
subspace of F, it implies that and F (y) 6= 0 and F |B ≡ 0. such that F (y) = 1 and
F |K ≡ 0. Dividing F by F (y), we may assume that F (y) = 1. Since F and f agree
on the codimension-1 linear subspace B of Y as well as on a single point y not in B, it
follows that F = f on whole Y : For every z ∈ Y , since z − f (z)y ∈ B, it follows that
F (z) = F ((z − f (z)y) + f (z)y) = 0 + f (z)F (y) = f (z). 

Remark 26. Regarding Theorem 25.(4), if X has finite dimension then there are many
elementary proofs in the literature [Web, 2.4.6, 2.4.10][Hör-C, 2.1.11][Rock, 11.4].

Exercise: If x1 , . . . , xn are finitely many linear independent vectors in a normed vec-

tor space X and a1 , . . . , an are arbitrary scalars then there exists a continuous linear
functional F on X such that F (xj ) = aj for every j.
Here are some corollaries:

Theorem 27. Let X, Y be normed vector spaces.

(1) Continuous linear functionals can be used to compute the norm of elements x ∈ X
in the sense that
kxk = sup{| hx, αi | : α ∈ X ∗ , kαk ≤ 1}.
(2) Continuous linear functionals can be used to compute the norm of bounded oper-
ators T ∈ B(X, Y ) in the sense that

kT k = sup{| hT x, βi | : x ∈ X, kxk ≤ 1, β ∈ Y ∗ , kβk ≤ 1}.

(3) kT k = kT ∗ k for every bounded operator T : X → Y .

(4) The mapping X → X ∗∗ , x 7→ x b, given by x
b(α) = α(x) is an isometric isomorphism
onto its range. This map is usually denoted by J and called the natural embedding of
X into X ∗∗ . (Recall that the closure of the range is the completion of X (page 8). Note
that X ∗∗ is always complete so is every closed subspace of it.). X is called reflexive if
this map is surjective.

Proof. (1) ≥ is because |α(x)| ≤ kαkkxk ≤ kxk. ≤ is because by Theorem 25.(3), when
Y = 0 and x 6= 0, there exists α ∈ X ∗ such that α(x) = kxk and kαk = 1.
(2) Immediate by Applying (1) to kT xk in kT k = supkxk≤1 kT xk
(3) Immediate by Applying (2) to kT ∗ k = supkαk≤1 kT ∗ αk = supkαk≤1,kxk≤1 | hx, T ∗ αi |.
(4) The map is clearly linear. That it is an isometry is exactly (1). Every surjective
isometry map clearly have continuous inverse. 

44
Remark 28. There are examples of non-reflexive Banach spaces X that are isometrically
isomorphic to X ∗∗ [Jam].
Example 29. (1) Lp (X, µ) is reflexive if 1 < p < ∞ [Fol, 6.16]. L1 (also L∞ ) is not
reflexive unless it is finite-dimensional [Bre, chapter 5]. (2) We have canonical
P∞ isometric
∗ ∼ 1 1 ∗ ∼ ∞
isomorphisms c0 = l and (l ) = l , both given by coupling (f, g) 7→ j=0 f (j)g(j).
Both are straightforward to check directly [Dou, page 7], but they can be deduced from
big duality theorems [Fol, 7.17 and page 225] and [Fol, 6.15]. Specially, c0 is not reflexive.
See also Example 42. 
Application 30 (Runge’s approximation theorem). For every open D ⊆ C and compact
K ⊆ D the followings are equivalent:
(1; topological condition) K adds no hole to D, in the sense that D \ K has no
component compactly supported in D.
(2; functional analysis condition) O(D) is dense in O(K), in the sense that every
holomorphic function on K can be uniformly approximated on K by holomorphic functions
on D.
(3; function theory condition) K is holomorphically convex in D, in the sense that for
any z ∈ D\K there exists some holomorphic function f on D such that |f (z)| > supK |f |.
Proof. (2 or 3⇒1) Assume (1) fails. Then D \ K has a component O which is compactly
supported in D. Note that ∂O ⊆ K. By the maximum principle
kf kO ≤ kf kK , ∀f ∈ O(D), (4.1)
which contradicts (3) for any z ∈ O. Now let (2) hold. Fix ζ ∈ O. Applying (2) to
f (z) := (z − ζ)−1 ∈ O(K) gives a sequence fn of holomorphic functions on D which
converge uniformly on K to f . Applying (4.1) to fn − fm shows that fn converges
uniformly on O to some limit function F . Note that F is holomorphic on O, continuous
on O, and equals f on ∂O namely (z − ζ)F (z) = 1 on ∂O. This latter identity persists on
O by the maximum principle applied to z 7→ (z − ζ)F (z) − 1. This gives a contradiction
when z = ζ.
(1⇒2) Fix an arbitrary f ∈ O(K). Consider f as an element of the space C(K)
of continuous functions on K equipped with uniform norm. Since the dual of C(K) is
given by (regular Borel) measures, according to the Hahn-Banach theorem, ´ we need to
check that any measure µ on K which is orthogonal to O(D) (namely gdµ = 0 for
all g ∈ O(D)) is also orthogonal to f . Let ψ be a smooth bump function compactly
supported on some neighborhood of K where f is holomorphic on, and ψ equals 1 on
some neighborhood of K. By the Cauchy-Pompeiu formula
ˆ
1 f (ζ)ψζ (ζ)
f (z) = √ dζ ∧ dζ, ∀z ∈ K,
2π −1 ζ −z
√
where dζ ∧ dζ is the Lebesgue measure on C, multiplied by −2 −1. √(Alternatively,
´ one
−1
can find a cycle γ in D \ K such that the Cauchy formula f (z) = (2π −1) γ
f (ζ)(ζ −
−1
z) dζ holds for every z ∈ K.) Therefore, by the Fubini’s theorem:
ˆ ˆ
1
f (z)dµ(z) = √ f (ζ)ψζ (ζ)ϕ(ζ)dζ ∧ dζ,
2π −1

45
 ´
where ϕ(ζ) = (ζ − z)−1 dµ(z). It suffices to show that the function ϕ defined on C \ K
is identically zero. Fix an arbitrary point z ∈ C \ K. Clearly ϕ is holomorphic. It also
vanishes on the unbounded component of C \ K because (ζ − z)−1 is a uniform sum of
monomials z n ∈ O(D) on |ζ| ≥ 2 supw∈K |w|. Let O be an arbitrary bounded component
of C \ K. Because of our topological assumption O´intersects C \ D, so let ζ0 be a point
in the intersection. Then ∂ k ϕ/∂ζ k (ζ0 ) = (−1)k k! (ζ0 − z)−k−1 dµ(z) vanishes because
(ζ0 − z)−k−1 is holomorphic on D. By the identity theorem ϕ vanishes on whole O.
(1 and 2⇒3) Fix z ∈ D\K. Choose a closed disc L centered at z with L ⊆ D\K. The
components of D \ (K ∪ L) are the same as those of D \ K apart from the fact that L has
been removed from exactly one of them. Therefore K ∪ L adds no hole to D. Applying
(2) to the function which is 0 in a neighborhood of K and is 1 in a neighborhood of L
gives f ∈ O(D) such that kf kK < 2−1 and kf − 1kL < 2−1 . This f satisfies (3). 
Remark 31. (1) Theorem 30 is the version of Runge’s approximation theorem which
appeared in [Hör-SCV]. The more famous version says: Given compact K ⊆ C and any
P ⊆ C which contains at least one point in each bounded component of C \ K, every
holomorphic function on K can be uniformly approximated on K by rational functions
with poles in P . This latter version can be proved with exactly the same techniques
[Rud-RCA, 13.6]; An elementary proof is given in [Sar, page 115][Gam, page 344]. (2) The
power series representation of holomorphic functions shows that holomorphic functions
on C can be uniformly approximated on compacts by polynomials. Therefore, for the
special case D = C, one deduces from Theorem 30 that: For K ⊆ C compact, C \ K is
connected if and only if every holomorphic function on K can be uniformly approximated
on K by polynomials. This is another useful version of Runge’s theorem [Rud-RCA, 13.7,
13.8]. 
Here is an application of Hahn-Banach theorem to PDEs. Usually the existence
of Green function for the Laplacian is proved via the solvability of Dirichlet problem;
however Lax [Lax-G] found a detour using Hahn-Banach extension theorem:
Application 32 (Existence of Greens function for smooth domains). Let U ⊆ R2 be a
bounded open with C 2 boundary ∂U . Then: (1) For every (field point) y ∈ U there exists
a function g(x, y) which is harmonic with respect to x ∈ U and continuous up to the
boundary with boundary value (2π)−1 log |x − y|.
(2) The Dirichlet problem “For every f ∈ C(∂U´) find u ∈ C 2 (U ) such that ∆u = 0
on U and u = f on ∂U ” can be solved via u(y) = ∂U f (x)Gn (x, y)dx, where Gn is the
derivative of G(x, y) := g(x, y) − (2π)−1 log |x − y| in the outward normal direction.
Proof. (1) Consider the space C(∂U ) of continuous functions on ∂U equipped with the
uniform norm, let A(∂U ) be the subspace consisting of the boundary values of harmonic
functions in D which are continuous up to the boundary. By maximum principle the
linear functional
αy : A(∂U ) → R, f 7→ f (y),
is bounded by 1, so can be extended to a linear functional on C(∂U ), denoted again by
αy . For every ξ ∈ R2 \ ∂U consider the function kξ defined on ∂U by

kξ (z) = k(z, ξ) = (2π)−1 log |z − ξ|, z ∈ ∂U.

46
We will prove that
g(ξ, y) := αy (kξ ), ξ ∈ R2 \ ∂U,
when restricted to ξ ∈ U , works as our Green function. That g(ξ, y) is harmonic with
respect to ξ is immediate from the two facts that kξ is harmonic with respect to ξ and
αy is a continuous linear functional. On the other hand, for every ξ ∈ R2 \ U , since
kξ ∈ A(∂U ) it follows that g(ξ, y) = kξ (y) = (2π)−1 log |ξ − y|. Therefore, we are left to
show that g(ξ, y) depends continuously on ξ as ξ crosses the boundary ∂U . To show this
let ξ be a point in U close to the boundary, and ξ 0 be its reflection across the boundary in
the sense that ξ + ξ 0 = 2z0 , where z0 is the point on ∂U of smallest distance to ξ. (This
can be proved by the implicit function theorem using C 2 assumption [KP, section 4.4].)
Consider  
0 1 |z − ξ|
g(ξ, y) − g(ξ , y) = αy log .
2π |z − ξ 0 |
Since ∂U is C 1 and compact it follows that |z − ξ|/|z − ξ 0 | → 0, uniformly for all points
z ∈ ∂U , as ξ → ∂U . Since αy is continuous it follows that g(ξ, y) − g(ξ 0 , y) → 0 as
ξ → ∂U .
(2) G satisfies ∆G = δ(x − y) and G|∂U ≡ 0. Green reciprocity formula
ˆ ˆ
uGn − Gun = u∆G − G∆u,
∂U U
´
valid for every G, u ∈ C 2 (U ) reduces to u = ∂U
uGn . 
Here is a simple application of Hahn-Banach separation that will be used later.

Theorem 33 (A variation of Hahn-Banach separation). If B is a convex balanced closed

subset of a LCTVS X and x ∈ X \ B then there exists F ∈ X ∗ such that F (x) > 1 but
|F | ≤ 1 on B.

Proof. By Theorem 25 there exists G ∈ X ∗√such that ReG(x) > K > ReG(y) for some
K ∈ R and every y ∈ B. Let G(x) = r exp( −1θ), r > 0, θ ∈ R. Since B is balanced it
follows that G(B) is balanced,√so it is a closed disk of radius s ∈ (0, r) around the origin
in F. Clearly, F := s−1 exp(− −1θ)G works. 

47
Chapter 5

Duality theory II: Weak and weak

star topologies

References: [Rud-FA, chapter 3][DS, chapter 5][Roy, chapters 14-5][Die].

In the study of spaces of functions there are topologies– different from the original
topology– that are naturally induced by different classes of continuous linear functionals;
these are weak and weak-star topologies to be discussed in this chapter. These topologies
are rarely metrizable (Theorem 41) but are locally convex most of the time (Theorem
34), so Hahn-Banach theorem can be applied to them. We start with the definitions of
these topologies:

1. Recall this from topology [Mun, section 13]: Let X be a set and S be a family of
subsets of X. There is a smallest1 topology on X such that all elements of S are open.
(Smallest in the sense that opens of this topology are opens of all the other topologies
such that all elements of S are open. Sometimes the term weakest is used instead
of “smallest”.) Opens of this topology are arbitrary unions of finite intersections of
elements of S. S is called a subbasis for this topology.

2. Let X be a set and F be a family of maps fα : X → Yα , α ∈ A, from X to topological

spaces Yα . The weak topology on X generated by F is the smallest topology
on X such that all fα are continuous. In other words, it is the topology on X with
subbasis consisting of all fα−1 (Uα ), α ∈ A, Uα ⊆ Yα open. Another description: A net
(xi )i∈I in X converges x ∈ X in this topology if and only if fα (xi ) converges fα (x)
in the topology of Yα for every α ∈ A. When all Yα is Hausdorff one can easily show
that the weak topology on X is Hausdorff if and only if F separate points in X in
the sense that if x, y are two distinct points in X then there exist α ∈ A such that
fα (x) 6= fα (y).

3. Let X be a TVS, with dual space X ∗ , the space of continuous linear functionals on X.
The weak topology on X is the weak topology generated by X ∗ . Another description:
1
On a set X, topology T is called smaller than topology T 0 if every open of T is an open in T 0 .

48
 A net (xi )i∈I in X converges x ∈ X in this topology if and only if α(xi ) → α(x) for
every α ∈ X ∗ . Another description: A local basis at x ∈ X for this topology is given
by {y ∈ X : |αj (x − y)| < , j = 1, . . . , n}, αj ∈ X ∗ , n ∈ N,  > 0.
Exercise: Show that a subset A ⊆ X of a TVS is weakly bounded (namely it is
bounded in the sense of Chapter 3.(8), when we put the weak topology on X) if and
only if each F ∈ X ∗ is bounded on A in the sense that there exists CF > 0 such that
|F (x)| ≤ KF for every x ∈ A.

4. Let X be a TVS. The weak star (or weak∗ , weak∗) topology on X ∗ is the weak
topology generated by {b x : x ∈ X} where x b : X ∗ → F denotes the linear functional
acting by α 7→ α(x). Another description: A net (αi )i∈I in X ∗ converges α ∈ X ∗ in this
topology if and only if αi (x) → α(x) for every x ∈ X. Another description: A local
basis at α ∈ X ∗ for this topology is given by {β ∈ X ∗ : |(β − α)(xj )| < , j = 1, . . . , n},
xj ∈ X, n ∈ N,  > 0.

When X is a normed vector space there are three topologies on X ∗ : weak star topol-
ogy, weak topology and the normed topology (also called strong topology) in increasing
order of strength. (Refer Proposition 35.)

Theorem 34 (When weak topologies are locally convex). (1) If X is a vector space and
F is a vector space of linear functional on X which separate points of X then the weak
topology on X generated by F makes X into a LCTVS whose dual space equals F.
(2) A LCTVS X equipped with the weak topology is a LCTVS whose dual space is X ∗ .
(3) If X is a TVS then the dual of X equipped with the weak-star topology is a LCTVS
whose dual space equals X, where each x ∈ X is seen as a linear functional on X ∗ acting
by α 7→ α(x).

Weakly continuous linear functionals are known if the

weak topology is given by a separating vector space.

Proof. (1) Topology is Hausdorff because F is separating. By linearity of elements of F

translations are homeomorphisms. A local basis at 0 is given by

U = {x ∈ X : |f1 (x)| < , . . . , |fn (x)| < }, n ∈ N, f1 , . . . , fn ∈ F,  > 0.

That vector space operations are continuous are straightforward to check. By linearity,
each U is convex, so X is a LCTVS. It remains to check that X ∗ = F. ⊇ is trivial.
For the other containment assume a linear functionals F on X which is continuous with
respect to the weak topology generated by F. By continuity, for every δ > 0 there exists
fT1 , . . . , fn ∈ F and  > 0 such that |F (x)| < δ whenever |fj (x)| <  for all
Pj. This implies
Kerfj ⊆ KerF , so by linear nullstellensatz (page 21) we have F = Cj fj , Cj ∈ F.
Then F ∈ F because F is a vector space.
(2) Immediate from (1). Note that since X is locally convex X ∗ separates points of
X by Theorem 25.(4).
(3) In (1) replace X by X ∗ and set F := X. Note that by the very definition, X
separate points of X ∗ . 

49
Proposition 35. Let X be a normed space. Then:
(1) Weak and strong topologies on X coincide if and only if the vector space dimension
of X is finite.
(2) Weak and weak-star topologies on X ∗ coincide if and only if X is reflexive.
Proof. (1) Let xj , j = 1, . . . , n be a basis for X. Consider αj ∈ X ∗ ,
P
aj xj 7→ aj .
Then {x ∈ X : |αj (x − y)| < , ∀j = 1, . . . , n},  > 0, y ∈ X, is a basis of opens
for weak topology, and also a basis of opens for the topology induced by l1 norm on
X. Conversely, assume X is of infinite dimension. Then every nonempty basic weak
neighborhood {|αj (x)| < , j = 1, . . . , n} of 0 is unbounded: one can inductively find
infinitely many directions along which all αj vanish. Therefore no such neighborhood is
contained in the open unit ball of the strong topology.
(2) If part is trivial. For only if part, assume continuous linear functional F on X ∗ .
Clearly, F is also continuous with respect to weak topology on X ∗ , hence continuous with
respect to weak-star topology on X, hence of the form x b, x ∈ X, by Theorem 34. 
A Banach space X is called uniformly convex if for every  > 0 there exists δ > 0
such that for every x, y ∈ X we have the following implication:

kxk ≤ 1, kyk ≤ 1, kx − yk >  ⇒ k(x + y)/2k < 1 − δ.

Equivalently, for every sequences xj and yj with kxj k = kyj k = 1 and k(xj + yj )/2k → 1
we have kxj − yj k → 0. Intuitively, if we slide a ruler of length  in the unit ball of X then
its midpoint must stay within a ball of radius 1−δ for some δ > 0. Figuratively speaking,
the unit ball of a uniformly convex space is uniformly free of “flat spots”. This class is
more special that reflexive Banach spaces [Bre, 3.31], but includes Hilbert spaces (an easy
consequence of the parallelogram identity) and Lp spaces, p ∈ (1, ∞) [Bre, chapter 4].
Example: Rn is uniformly convex with l2 norm but not with l1 or l∞ norms.
Theorem 36 (Radon-Riesz). Let xj be a sequence in a normed vector space X which
converges weakly to x ∈ X. Then:
(1) kxk ≤ lim inf kxj k.
(2) If X is uniformly convex then xj strongly converges x if and only if kxk = lim kxj k.
Proof. (1) By Theorem 25 there exists α ∈ X ∗ such that kαk = 1 and α(x) = kxk. Then
kxk = lim α(xj ) ≤ lim inf kαkkxj k = lim inf kxj k.
(2) Only if part is trivial. For the converse, putting the trivial case x = 0 aside, set:

λj := max(kxj k, kxk), yj := xj /λj , y := x/kxk.

Clearly, λj → kxk and yj weakly converges y. By (1) we have kyk ≤ lim inf k(yj + y)/2k.
On the other hand kyk = 1 and kyj k ≤ 1, so in fact k(yj + y)/2k → 1. By uniform
convexity kyj − yk → 0, hence xj strongly converges x. 
Here is an application of the Hahn-Banach Theorem to weak topologies.
Theorem 37 (Mazur). If A is a convex subset of a LCTVS then the norm closure and
weak closure of A coincide. Therefore, A is closed (respectively, dense) if and only if it
is weakly closed (respectively, dense).

50
Proof. Since the weak topology has less closed subsets than the original topology it
follows that the weak closure Aw of A contains the original closure A. To show the
reverse containment assume x 6∈ A. By Theorem 25.(4) there exist F ∈ X ∗ and K ∈ R
such that ReF (x) < K < ReF (y) for every y ∈ A. Then {ξ ∈ X : ReF (ξ) < K} =
F −1 ({z ∈ F : Rez < K}) is a weak neighborhood of x which does not intersect A, hence
x 6∈ Aw . 

Application 38 (Mazur). Let X be a metrizable LCTVS. If xj is a sequence in X that

converges weakly to x ∈ X then there exists a sequence yj in X which converges strongly
to x and each of its terms are convex linear combination of finitely many terms of xj .

Proof. Let C be the convex hull of {xj }, namely the intersection of all convex subsets
of X which contain all xj , or equivalently the set of all convex linear combinations of
finitely many terms of xj . Then x belongs to the weak closure of C, hence to the strong
closure of C by Mazur theorem. 

5.1 Some compactness theorems

Compactness theorems are among the most useful results in analysis. Most important
ones are:

• Bolzano-Weierstrass [Apo-A, 3.24]: Every bounded sequence in Rn has a convergent

subsequence.

• Tychonoff [Fol, 4.22][Mun, 37.3] (for product topology).

• Arzela-Ascoli [Fol, 4.44][Mun, 47.1] (for continuous functions or maps).

• Montel [Hör-SCV, 2.2.5] (for holomorphic functions).

• Rellich [Fol, 9.22][Tay-PDE, chapter 4] (for Sobolev functions).

• Frechet-Kolmogorov [DS, page 298][Bre, 4.26]. (for Lp (Rn ) functions, 1 ≤ p < ∞).

• Riesz [Fol, 2.30-32] (for Lp (X, µ) functions).

• Alaoglu (Theorem 39), Helley (Theorem 44), Kakutani (Theorem 45), Kakutani-
Eberlein-Smulian (Theorem 47), etc. (for normed vector spaces or more generally
topological vector spaces).

Here is a fundamental compactness theorem and one of the main reasons for the
usefulness of weak-star topology. We will crucially use it in Chapter 9.

Theorem 39 (Alaoglu). The closed unit ball of the dual space of any normed vector space
is compact in weak-star topology. More generally, if U is a neighborhood of the origin in
a TVS X then {α ∈ X ∗ : |α(x)| ≤ 1, ∀x ∈ U } is compact in weak-star topology.

The closed unit ball of the dual space of a normed vector space is weak-star compact.

51
Proof. The closed unit ball B of the dual of normed vector space X is the set of linear
elements in
n o Y
α
D := X − → F : |α(x)| ≤ kxk, ∀x ∈ X = Dx , Dx = {a ∈ F : |a| ≤ kxk}.
x∈X

On the other hand the weak-star topology on B and the product topology on D both
coincide with the topology ofQ
pointwise convergence. (For every family Xα of topological
spaces, theQproduct topology Xα is the weak topology
Q generated by canonical projection
maps πα : Xα → Xα ; therefore, a net (xi ) in Xα converges x if and only if πα (xi ) →
πα (x) for every α.) Also, D is compact in product topology by Tychonoff theorem.
Therefore we only need to check that B is closed in D. If (αi ) is a net in B which
converges α ∈ D then for every a, b ∈ F we have

α(ax + by) = (lim αi )(ax + by) = lim(aαi (x) + bαi (y)) = aα(x) + bα(y),

so α is linear, hence α ∈ B. Refer [Rud-FA, 3.15] for the proof of the TVS version. 

Application 40. Every Banach space is (isometrically isomorphic to) a closed linear
subspace of some C(X), X a compact topological space.

Proof. Consider Banach space Y and let X be the closed unit ball of Y ∗ equipped with
the weak-star topology. X is compact by Alaoglu. Define the linear map F : Y → C(X),
F (y)(α) = hy, αi. For every y ∈ Y we have

kF (y)k = sup | hy, αi | = kyk

kαk≤1

by Theorem 27. This shows that F is an isometry. The range of any isometric operator
between Banach spaces is a closed subspace by a standard Cauchy sequences argument.
The inverse of F : Y → RanF is also continuous by inverse mapping theorem. 
Next theorem gather some elementary statements about the fundamental notions of
separability, reflexivity and metrizability:

Theorem 41 (Separability, reflexivity and metrizability). Let X be a normed vector

space.
(1) If X ∗ is separable so is X. (The converse is not true.)
(2) If X is reflexive so is every closed linear subspace of it.
(3) X is reflexive if and only if X ∗ is so.
(4) None of the weak or weak-star topologies on X ∗ are metrizable if vector space
dimension of X is infinite.
(5) The weak topology on the closed unit ball of X generated by a family F ⊆ X ∗ which
is separable and separates points in X, is metrizable. Specially, the weak topology on the
closed unit ball of X is metrizable if X ∗ is separable, and that the weak-star topology on
the closed unit ball of X ∗ is metrizable if X is separable.
(6) If X is reflexive then the weak topology on the closed unit ball of X is metrizable
if X is separable.

52
Proof. (1) Let αj be a dense sequence in X ∗ . For each j find xj ∈ X such that kxj k = 1
and αj (xj ) > 12 kαj k. We claim that the linear span of {xj } is dense in X. If not, there
exists α ∈ X ∗ such that kαk = 1 and α(xj ) = 0 for every j. For every  > 0 one can find
αj with kα − αj k < . This leads to the following contradiction:

1 1
(1 − ) < kαj k < |αj (xj )| = |(α − αj )(xj )| ≤ kα − αj k < .
2 2
The converse is not true: l1 is separable but not l∞ = (l1 )∗ .
(2) Let Y be a closed linear subspace of X, with i : Y ,→ X the inclusion map. Let
F be a continuous linear functional on Y ∗ . Then X ∗ → F, α 7→ F (α ◦ i), is a continuous
linear functional on X ∗ , so there exists x ∈ X such that F (α◦i) = α(x) for every α ∈ X ∗ .
If α ∈ X ∗ vanishes on Y then α(x) = 0. This shows that x ∈⊥ Y ⊥ = Y = Y . Every
β ∈ Y ∗ can be extended to some α ∈ X ∗ , namely β = α ◦ i, so F (β) = α(x) = β(x).
(3) Let X be reflexive, and fix F ∈ X ∗∗∗ . Then the mapping X → F, x 7→ F(b x), is
a continuous linear functional α on X such that F(b x) = α(x). Since every F ∈ X ∗∗ is
of the form x b, x ∈ X, it follows that F(F ) = F (α). This means that X ∗ is reflexive.
Conversely, if X ∗ is reflexive, so is (X ∗ )∗ , so is its closed linear subspace X by (2).
(4,5,6) [Roy, section 15.4]. 
Exercise: Prove that l1 is separable, but not l∞ . (Hint. Use the Cantor diagonal
argument for the second.)

Example 42. X = C([−1, 1]) is not reflexive. Here is a reason. The Dirac unit mass
functionals δx : X → F, f 7→ f (x), x ∈ [−1, 1] constitute an uncountable family of
elements of X ∗ with kδx − δy k = 2 for every two distinct points x, y ∈ [−1, 1]. This shows
that X ∗ is not separable. Since X is clearly separable, if X were reflexive, then Theorem
41.(1) would imply that X ∗ is separable. Another argument.´If X were ´ 1 reflexive then by
∗ 0
Theorem 25.(3), assuming α ∈ C([−1, 1]) given by α(f ) = −1 f − 0 f , one could find
f0 ∈ C([−1, 1]) such that kf0 k = 1 and α(f0 ) = kαk. This is absurd because kαk = 2
and |α(f )| < kf k for every f ∈ C([−1, 1]) \ {0}. More generally, one can prove that if Y
is a compact Hausdorff space then C(Y ) is reflexive (respectively, separable) if and only
if Y is finite (respectively, second countable) [Roy, pages 302, 251]. 
√
Example 43 (von Neumann). Let A ⊆ l2 be the set of all jej , j = 1, 2, . . ., where
e0 , e1 , e2 , . . . is the standard orthonormal basis of l2 = l2 ({0, 1, 2, . . .}). Then, the origin
is in the weak closure of A, but no sequence in A weakly converges the origin. To verify
the first statement, we need to check that A intersects with the basic neighborhood

{y ∈ l2 : | hxk , yi | < , ∀k = 1, . . . , n},

xk = (xk1 , xk2 , . . .) ∈ l2 , k = 1, . . . , n,  > 0,
√
there exists positive integer√j such that j |xkj | <  for every k.
or that, equivalently, P
This is true because i |xki |2 < ∞ implies that i |xki | <  for every k and almost all
i. For the second
√ statement, contrapositively, assume a sequence jn ofP positive integers
such that an := jn ejn weakly converges 0. Then, han , xi → 0 for x := (j + 1)−2 ej ∈ l2
implies that jn → ∞. However, Theorem 52.(7) implies that an is strongly bounded. 

53
 Recall that a topological space is sequentially compact if every sequence has a
convergent subsequence. In metrizable spaces, this notion coincides with the usual notion
of compactness (every open cover has a finite subcover), but in general neither implies the
other [Kel, page 138]. Note that Alaoglu Theorem does not say that the closed unit ball
of the dual of a normed vector space is sequentially compact with respect to the weak-star
topology. (Example: αj : l∞ → F, (x0 , x1 , . . .) 7→ xj , is a sequence in the closed unit ball
of (l∞ )∗ but has no weak-star convergent subsequence.); however we have:

Theorem 44 (Helley selection principle). (1) The closed unit ball of the dual space of
a separable normed vector space X is sequentially compact in weak-star topology; more
explicitly, every bounded sequence Fj in X ∗ has a subsequence Fjn and F ∈ X ∗ such that
Fjn (x) → F (x) for every x ∈ X. More generally, if U is a neighborhood of the origin
in a separable TVS X then every sequence Fj in X ∗ which is equibounded on U (namely
there exists C > 0 such that |Fj (x)| ≤ C for every j and x ∈ U ) has a subsequence which
converges in weak-star topology.
(2) The closed unit ball of a reflexive normed vector space X (for example a Hilbert
space or an Lp (X, µ) space, 1 < p < ∞; no separability is assumed.) is sequentially
compact in weak topology; more explicitly, every bounded sequence in X has a weakly
convergence subsequence.

Equibounded sequences of continuous linear functional on separable

normed vector spaces have pointwisely convergent subsequences.

Bounded sequences in reflexive normed vector spaces have weakly convergence subsequences.

Proof. (1) TVS case. One can assume C = 1. Set K := {F ∈ X ∗ : |F (x)| ≤ 1, ∀x ∈ U }.

We are supposed to show that K is sequentially compact in weak-star topology, however
we know by Alaoglu theorem that K is compact in weak-star topology compact, so we are
done by proving that K is metrizable in weak-star topology. Let xk be a dense sequence of
P −ktopology T of K coincides with the topology
points in X. We assert that the (weak-star)
Td induced by the metric d(F, G) = 2 |F (xk ) − G(xk )|. Fixing F , each summand
of d is T -continuous
P −k+1with respect to G and the series converges uniformly (because it is
dominated by 2 ), so d is T -continuous, hence Td ⊆ T . The identity map K → K,
where the source is equipped with topology T and the target with Td , is a bijective
continuous map from a compact space to a Hausdorff one, hence a homeomorphism. We
proved Td = T .
Normed vector space case. We give a direct argument avoiding Alaoglu. Let xk be a
dense sequence of points in X. Let Fj be a sequence in X ∗ , which is bounded by C > 0.
For each k the scalar sequence hxk , Fj i is bounded in F, so has a convergent subsequence
by Bolzano-Weierstrass theorem. By Cantor diagonal argument, after passing to a sub-
sequence, we can assume that for each k the sequence hxk , Fj i converges. For every  > 0
and x ∈ X, choosing xl with kx − xl k < , the estimation

|hx, Fj − Fk i| ≤ | hx − xl , Fj − Fk i | + | hxl , Fj − Fk i | ≤ 2C + | hxl , Fj − Fk i |,

shows that hx, Fj i converges for every x ∈ X.

54
 (2) Let xj be a bounded sequence in a reflexive normed vector space X. The closed
linear span Y of {xj } is reflexive (Theorem 41.(2)) and separable (because F is separable).
Y ∗ is also separable by Theorem 41.(1). x bj is a bounded sequence of continuous linear
functional on Y ∗ , so by (1), has a subsequence x bjn which converges pointwisely to some
yb, y ∈ Y . This latter statement, since every functional in X ∗ restricts to a functional in
Y ∗ , means that xjn weakly converges y.
Here is a direct argument for Hilbert spaces. Let xj be a sequence in a Hilbert space
X, which is bounded by C > 0. For every k ∈ N the scalar sequence hxk , xj i is bounded in
F, so has a convergent subsequence by Bolzano-Weierstrass theorem. By Cantor diagonal
argument, after passing to a subsequence, we can assume that for every k the sequence
hxk , xj i converges, or equivalently, hy, xj i converges for every y in the linear space Y of
{xj : j ∈ N}. For every  > 0 and y 0 ∈ Y , choosing y ∈ Y with ky − y 0 k < , the
estimation

|hy 0 , xj − xk i| ≤ | hy 0 − y, xj − xk i | + | hy, xj − xk i | ≤ 2C + | hy, xj − xk i |,

shows that hy, xj i converges for every y ∈ Y . Therefore y 7→ lim hy, xj i is a well-defined
bounded linear functional on Hilbert space Y . By Riesz representation theorem there
exists x ∈ Y such that lim hy, xj i = hy, xi for every y ∈ Y . Since the same equality
⊥ ⊥
trivially holds for every y ∈ Y it follows that the equality holds for every y ∈ Y + Y =
X. This means that xj weakly converges x. 
Application 45 (Banach-Saks-Kakutani). Every bounded sequence in a uniformly convex
Banach space has a subsequence whose arithmetic means is strongly convergent.

Bounded sequences in uniformly convex Banach spaces

have strongly Césaro convergent subsequences.

Proof. For general case refer [Die, chapter 8]. We prove it for Hilbert spaces. Let xj
be a sequence in Hilbert space X, bounded by C > 0. By Helley selection principle
(Theorem 44) we can assume that xj converges weakly to some x ∈ X. Replacing xj
by xj − x we can assume x = 0. After passing to a subsequence one can assume that
| hxj , xk i | ≤ 1/(j − 1) for j ≥ 2 and 1 ≤ k < j. Then for every positive integer n we have

n 2 n
!
1X X X
xj = n−2 kxj k2 + 2Re hxj , xk i ≤
n j=1 j=1 1≤k<j≤n
!
X C 2 n + 2(n − 1)
n−2 C 2n + 2(j − 1)−1 = ,
1≤k<j≤n
n2

which approaches 0 as n → ∞. 
Application 46 (Hardy functions). Suppose p ∈ (1, ∞), D ⊆ C the open unit disk, dθ
the Lebesgue measure on unit circle T := ∂D. If f is a harmonic function on D such that
ˆ  √ p
sup f re −1θ dθ < ∞, (5.1)
0<r<1

55
(notation: f ∈ hp (D)) then there exists f ∗ ∈ Lp (T) such that f the Poisson extension of
f ∗ in the sense that
 √  ˆ  √ 
f re −1θ = f ∗ e −1ϕ Pr (θ − ϕ)dϕ,

where
√ 2
−1t
X √ 1−r 2 1 − re
Pr (t) = r|j| e −1jt
= = √ 2.
j∈Z
1 − 2r cos t + r2 1 − re −1t
Conversely, if f : D → C is the Poisson extension of some f ∗ ∈ Lp (T) then f ∈ hp (D).

Proof. Let f ∈ hp (D). For any √ r ∈ (0, 1) let fr : √

T → C be the restriction of f to the
circle {|z| = r}, namely fr (exp( −1θ) = f (r exp( −1θ). The assumption (5.1) exactly
says that {fr : r ∈ (0, 1)} is a bounded set in Lp (T). Fix some sequence rj ∈ (0, 1) which
approaches 1. By Helley selection principle, after passing to
∗ p
´ ∗ rj → 1 and
´ a subsequence,
frj weakly converges some f ∈ L (T) in the sense that lim frj P dθ = f P dθ for every
P ∈ Lq (T) = (Lp (T))∗ . On the other hand each f (rj −) is a continuous function on D
which is harmonic on D, so equals the Poisson extension of its boundary values [Ahl, page
169]. We have
 √   √  ˆ  √ 
−1θ −1θ
f re = lim f rj re = lim frj e −1ϕ Pr (θ − ϕ)dϕ =
ˆ  √ 
∗
f e −1ϕ Pr (θ − ϕ)dϕ.

For the converse (and to know what happens for p = 1 or Cm instead of the complex
plane) refer [Rud-RCA, 11.16, 11.30][Rud-SCV, 4.3.3][Kra, chapter 8]. 
Although weak topologies are far from being metrizable but we have the following
surprising result:

Theorem 47 (Eberlein-Smulian-Kakutani). The closed unit ball B of a Banach space X

is compact if and only if B is weakly sequentially compact if and only if X is reflexive.

The closed unit ball of a Banach space X is weakly compact if and

only it is weakly sequentially compact if and only if X is reflexive.

Every bounded sequence in a uniformly convex Banach space has a subsequence whose
arithmetic means are strongly convergent.
Proof. [Roy, page 304][DS, chapter 5][Die, chapter 4]. The deepest part is to deduce
weak compactness from sequentially weak compactness, a miracle. 

Application 48. Let X be a reflexive Banach space, Y ⊆ X a closed convex subset and
x ∈ X a point. Then Y has a point of shortest distance to x.

56
Proof. We can assume x = 0 and 0 6∈ Y . Set δ := inf y∈Y kyk, and let yj be a sequence
in Y such that kyj k → δ. Neglecting finitely many terms one can assume that all yj
lies in the intersection of Y with the closed ball B of radius 2δ around the origin in X.
B is weakly sequentially compact by Eberlein-Smulian theorem, so after passing to a
subsequence, one can assume that yj weakly converges y ∈ B. Since Y ∩ B is closed and
convex it is weakly closed by Mazur theorem, therefore y ∈ Y . By Theorem 25 there
exists α ∈ X ∗ such that kαk = 1 and α(y) = kyk. Then:

δ ≤ kyk = α(y) = lim α(yj ) ≤ lim kαkkyj k = lim kyj k = δ.

So y is a point in Y of shortest distance to x. 

Example 49. This example shows that the conclusion of Application 48 fails if reflexivity
assumption is dropped. Supposed the space X ´:= C([−1, ´ 11]; R), and let Y be the closed
0
linear subspace consisting of all y ∈ X with −1 y = 0 y = 0. (Note that X is not
´0 ´1 ´0
reflexive by Example 42.) Fix x ∈ X with −1 x = 1 = − 0 x. Since −1 x − y = 1 it
follows that max[−1,0] x − y ≥ 1 and equality happens exactly when x − y is constantly
1 on [−1, 0]. Similarly, min[0,1] x − y ≤ −1 and equality happens exactly when x − y is
constantly −11 on [0, 1]. Since these two equality condition can not hold simultaneously
it follows that kx − ykX > 1. However one can really find elements y in Y such that
kx − ykX is sufficiently closed to 1. 

57
Chapter 6

Duality theory III: Krein-Milman

theorem

References: [Rud-FA, 3.22-25].

Assuming a vector space X, a subset S ⊆ X and a point x ∈ X, x is called an

extreme point of S if x is not an internal point of any line interval whose end points
are in S, except when both end points are in x, more precisely, for every y, z ∈ S and
every λ ∈ (0, 1) if λy + (1 − λ)z ∈ S then y = z = x.

Theorem 50 (Krein-Milman). Let X be a TVS on which X ∗ separates points (for exam-

ple a LCTVS or the dual of a TVS equipped with weak-star topology), and K a nonempty
compact subset. Then K has at least one extreme point. If K is also convex then K is
the closed convex hull of the set of its extreme points.

Proof. [Rud-FA, 3.23]. Uses Zorn lemma together with a variation of Hahn-Banach
separation theorem. 

Application 51 (Stone-Weierstrass theorem). Let X be a compact Hausdorff space.

(1) Let C(X; R) be the real algebra of continuous real-valued functions on X. A
subalgebra A ⊆ C(X; R) (subalgebra means being closed under vector space operations and
the pointwise multiplication) which contains the constant function 1 and separates points
(namely for every two distinct points x, y ∈ X there exists f ∈ A such that f (x) 6= f (y))
is dense.
(2) Let C(X) be the complex algebra of continuous functions on compact topological
space X. A subalgebra A ⊆ C(X) (subalgebra means being closed under vector space
operations and the pointwise multiplication) which contains the constant function 1, is
closed under the pointwise conjugation, and separates points (namely for every two dis-
tinct points x, y ∈ X there exists f ∈ A such that f (x) 6= f (y)) is dense.

Proof. (1) Recall that the dual of C(X; R) is canonically identified with the space M (X)
of finite signed Borel measures on X (Theorem 53). If, by contradiction, A is not dense
in C(X; R)´ then by Theorem 25 the set K of measures µ ∈ M (X) which ´ vanishes on A
(namely f dµ = 0 for every f ∈ A) and their total variation kµk := d|µ| is ≤ 1 has

58
a nonzero element. K is clearly convex and weak-star compact by Alaoglu theorem. By
Krein-Milman theorem K has a extreme point ν with kνk = 1. The main observation is
that every g ∈ A with values in (0, 1) gives the following convex representation of ν in
terms of other elements in K:
ˆ
gν (1 − g)ν
ν = t + (1 − t) , t := kgνk = gd|ν| ∈ (0, 1).
t 1−t

That gν/t and (1 − g)ν/(1 − t) vanish on A is because A is closed under multiplication.

Since ν is an extreme point it follows that ν = gν/t, hence g has ´the same value at all
points of the support of ν, namely the points x ∈ X such that V |dν| > 0 for every
open neighborhood V of x. From this we can easily deduce that the support of ν consists
of only one point: If x, y are two distinct points in the support of ν, since A separate
points one can find g ∈ A such that g(x) 6= g(y); by adding a large enough constant to
g and then multiplying by a small enough positive real number one can assume that g
has values in (0, 1), and this is a contradiction. Therefore, ν equals the Dirac unit mass
measure ±δx at some´ x ∈ X. Since the constant function one belongs to A, we get the
contradiction 0 = 1dν = ±1.
(2) Apply (1) to the algebra A0 of real parts of elements of A. Note that A√ 0
contains
the imaginary parts of element of A because A is closed under multiplication by −1. 
Krein-Milman theorem has many other applications, among them:

• Gelfand-Raikov theorem: A locally compact group has enough irreducible unitary

representations to separate points of the group. [Fol-AHA, 3.24].

• Bochner theorem: In a locally compact abelian group G the Fourier transform pro-
vides a one-to-one correspondence between continuous positive-definite1 functions ϕ
on G which are normalized ϕ(1) = 1 and probability measures µ on the Pontryagin
´
b namely ϕ(g) = b ξ(g)dµ(ξ) for every g ∈ G. [Fol-AHA, 4.19]
dual group G, G

´
1
f (x)f (y)ϕ(y −1 x)dydx ≥ 0 for every f ∈ L1 (G).

59
Chapter 7

Duality theory IV: Summary of

results

References: [Rud-FA, chapters 3,4].

The most important slogan of functional analysis is that (continuous linear) functionals
can be used to detect many notions in functions spaces. In the following theorem we
gather several instances of this phenomenon.

Theorem 52. Continuous linear functionals can be used for:

1. Separating convex subsets. Let A and B be two disjoint nonempty convex subsets of
a TVS X. If A is open then A and B can be separated by closed hyperplanes in the
sense that there exists F ∈ X ∗ and K ∈ R such that ReF (a) < K ≤ ReF (b) for every
a ∈ A and b ∈ B. If A is compact, B is closed and X is locally convex then A and B
can be strictly separated by closed hyperplanes in the sense that there exists F ∈ X ∗
and K1 , K2 ∈ R such that ReF (a) < K1 < K2 < ReF (b) for every a ∈ A and b ∈ B.
Specially, if X is a LCTVS then X ∗ separate points in X in the sense that for every
two distinct points x, y in X there exits F ∈ X ∗ such that F (x) 6= F (y).

2. Computing the closure of linear subspaces. If Y is a linear subspace of a LCTVS X

and Z is a linear subspace of X ∗ then ⊥ Y ⊥ gives the (norm) closure of Y in X and
⊥ ⊥
Z gives the weak-star closure of Z in X ∗ . (⊥ is defined on page 42.) In words:
x ∈ X belongs to the (norm) closure of Y in X exactly when every continuous linear
functional on X which vanishes on Y also vanishes on x; and α ∈ X ∗ belongs to the
weak-star closure of Z in X ∗ exactly when α kills every x ∈ X which is also killed by
all members of Z.

3. Proving density. If Y is a linear subspace of a LCTVS X and Z is a linear subspace of

X ∗ then Y is dense in X if and only if there is no nonzero continuous linear functional
on X which vanishes on whole Y (in notations: Y ⊥ = {0}); and Z is dense in X ∗
if and only if the origin is the only common zero of all elements of Z (in notations:
⊥
Z = {0}).

60
 4. Computing the closure of convex subsets; Mazur. If A is a convex subset of a LCTVS
then the (original) closure and weak closure of A coincide.

5. Computing the norms of elements. If x is an element of a normed vector space X then

kxk = sup{| hx, αi | : α ∈ X ∗ , kαk ≤ 1}.

6. Computing the norms of operators. If T : X → Y is a bounded operator between

normed vector spaces then

kT k = sup{| hT x, αi | : x ∈ X, kxk ≤ 1, α ∈ Y ∗ , kαk ≤ 1}.

7. Proving boundedness of subsets. A subset of a LCTVS is bounded if and only if it is

weakly bounded.

8. Computing quotients. If X is a normed vector space and Y a closed linear subspace

then we have canonical isometric isomorphisms:
∼
=
(X/Y )∗ −
→ Y ⊥, α 7→ (x 7→ α(x + Y )),
∼
=
→ Y ∗,
X ∗ /Y ⊥ − α + Y ⊥ 7→ (y 7→ α(y)).

9. Proving continuity of linear maps. If T : X → Y is a linear map between normed

vector spaces then T is continuous if and only if T ∗ (Y ∗ ) ⊆ X ∗ (namely β ◦ T is con-
tinuous for every continuous linear functional β on Y ) if and only if T is weak-to-weak
continuous (namely continuous if both X and Y are equipped with weak topologies).

10. Computing kernels of operators. If T : X → Y is a bounded operator between normed

vector then
KerT = ⊥ RanT ∗ , KerT ∗ = Ran⊥T,
w∗
RanT ∗ = Ker⊥
T, RanT = ⊥ KerT ∗ .

11. Proving that operators are injective or range-dense. If T : X → Y is a bounded oper-

ator between normed vector spaces then:

(a) T is range-dense if and only if T ∗ is injective.

(b) T is injective if and only if T ∗ is weak-star range dense.

12. Proving surjectivity of operators. If T : X → Y is a bounded operator between Banach

spaces, then the followings are equivalent:

(a) T is surjective.
(b) T is open. Equivalently, there exists r > 0 such that T X1 ⊇ Yr , where Xr denotes
the open ball of radius r in X around the origin, and similarly for Yr . More
concretely, for every y ∈ Y there exists x ∈ X with T x = y and kxk ≤ r−1 kyk.
(c) There exists r > 0 such that T X1 ⊇ Yr .

61
 (d) T ∗ is bounded from below; equivalently, T ∗ is injective and closed-range. (Theo-
rem 8.)

13. Proving that operators are closed-range. If T : X → Y is a bounded operator between

Banach spaces then the followings are equivalent:

(a) T is range-closed.
(b) There exists C > 0 such that for every y ∈ RanT there exists x ∈ X with T x = y
and kxk ≤ Ckyk.
(c) T ∗ is weak-star range-closed.
(d) T ∗ is range-closed.
(e) kT xk ≥ C inf{kx − ξk : ξ ∈ KetT } for some C > 0 and every x ∈ X.

Proof. (1) Proved in Theorem 25.(4).

(2) Y ⊆ ⊥ Y ⊥ and Z w∗ ⊆ ⊥ Z ⊥ are clear. If x ∈ X \ Y then by Theorem 25.(4)
there exists a continuous linear functional F on X which vanishes on Y (hence on Y )
but not at x, namely x ∈ / ⊥ Y ⊥ . If F ∈ X ∗ \ Z w∗ then by Theorem 25.(4) there exists a
weak-star-continuous linear functional α on X ∗ which vanishes on Z w∗ (hence on Z) but
not at F . By Theorem 34, α is of the form x b, x ∈ X, acting by F 7→ F (x). We have
⊥ ⊥
proved that x b 6∈ Z .
(3) Immediate from (2).
(4) Proved in Theorem 37.
(5, 6) Proved in Theorem 27.
(7) Only if part is trivial. The converse is a deep theorem whose proof needs Alaoglu,
Hahn-Banach separation theorem and Baire category theorem [Rud-FA, 3.18]. (It also
follows from a theorem of Mackey on dual systems [Tre, 36.2][MV, pages 248-9].) Here we
prove the special case when X is a normed vector space. Let S ⊆ X be a weakly bounded
subset of normed vector space X. This means that for each α ∈ X ∗ there exists Cα > 0
such that | hs, αi | ≤ Cα for every s ∈ S. In terms of the natural embedding X → X ∗∗ ,
x 7→ x b(α) = hx, αi, this exactly means that the family of bounded linear functionals
b, x
sb : X ∗ → F, s ∈ S, is pointwisely equibounded; hence it is uniformly equibounded by the
uniform boundedness principle (Theorem 5.(1)): There exists C > 0 such that kb sk < C
for every s ∈ S. Since kb sk = ksk (Theorem 27) this means that S is bounded.
(8) The first map is clearly well-defined and linear. Its set-theoretic inverse is given
by β 7→ (x + Y 7→ β(x)). Just using definitions it is straightforward to check that both
these maps are isometry. In regard to the second identification, the map given is clearly
well-defined and linear. Its set-theoretic inverse is: f ∈ Y ∗ is mapped to F + Y ⊥ , where
F is any extension of f to X. (There is at least one by Theorem 25.(20 ).) Clearly,
kF + Y ⊥ k = inf g∈Y ⊥ kF + gk ≥ kf k. The reverse inequality is immediate from 25.(20 ) by
choosing F with kF k = kf k.
(9) If T is continuous and β ∈ Y ∗ then clearly T ∗ β = β ◦ T is continuous. If T ∗ (Y ∗ ) ⊆
X ∗ , β ∈ Y ∗ and xα is a net in X weakly converging x then hxα , T ∗ βi → hx, T ∗ βi, or
equivalently, hT xα , βi → hT x, β ◦ T i, namely T xα weakly converges T x, and this means
that T is weak-to-weak continuous. Finally, assume that T is weak-to-weak continuous
but not continuous. Then the image of the unit ball under T is not bounded, so is not

62
weakly bounded by (7). This means that there exists β ∈ Y ∗ such that {| hT x, βi | : kxk ≤
1} is not bounded. To get a contradiction it suffices to show that T ∗ β ∈ X ∗ . Let xα be a
net in X which weakly converges x. Then by our assumption T xα weakly converges T x.
Therefore β ◦ T xα converges β ◦ T x. This means that T ∗ β is weakly continuous on X.
T ∗ β ∈ X ∗ by Theorem 34.
(10)

x ∈ KerT ↔ T x = 0 ↔ hT x, βi = 0, ∀β ∈ Y ∗ ↔ hx, T ∗ βi = 0, ∀β ∈ Y ∗ ↔ x ∈⊥ RanT ∗ .

Theorem 25

β ∈ KerT ∗ ↔ T ∗ β = 0 ↔ hx, T ∗ βi = 0, ∀x ∈ X ↔ hT x, βi = 0, ∀x ∈ X ↔ β ∈ Ran⊥

T.

This proves the first two equations. The last two follow from the first two and (2).
(11) Immediate from (10).
(12) (12a)⇔(12b)⇔(12c) was proved during the proof of the open mapping theorem
(Theorem 5.(2)).
(12b)⇔(12d) Suppose that T X1 ⊇ Yr for some r > 0. Then for every β ∈ Y ∗ we have
kT ∗ βk = sup hx, T ∗ βi = sup hT x, βi ≥ sup hy, βi = rkβk.
x∈X1 x∈X1 y∈Yr

Therefore T ∗ is bounded from below. Conversely, suppose that kT ∗ βk ≥ rkβk for some
r > 0 and every β ∈ Y ∗ . We assert that T X1 ⊇ rY1 . Fix y0 6∈ T (X1 ). By Theorem 33
there exists β ∈ Y ∗ such that | hy0 , βi | > 1 and | hy, βi | ≤ 1 for all y ∈ T (X1 ). Then
kT ∗ βk = sup | hx, T ∗ βi | = sup | hT x, βi | ≤ 1,
x∈X1 x∈X1

hence
r < r| hy0 , βi | ≤ rky0 kkβk ≤ ky0 kkT ∗ βk ≤ ky0 k,
namely y0 6∈ rY1 .
(13) All parts are proved by applying (12) to appropriate restriction maps.
(13a)⇔(13b) If RanT is closed then the restriction S : X → RanT of T to its range
is a surjective continuous map between Banach space, hence an open map by (12). This
means that there exists C > 0 such that S(Xr ) ⊇ YCr ∩ RanT for every r > 0, where Xr
is the open ball in X of radius r around the origin, and similarly for Yr . This is exactly
(13b). The converse is via a famous Cauchy sequence argument: If T xj → y then T xj
is Cauchy, and by (13b) one can assume that xj is bounded and Cauchy, so xj → x;
therefore y = T x.
w∗
(13a,13b)⇒(13c) Since RanT ∗ = Ker⊥ ⊥
T by (10), we need to show that KerT ⊆ RanT ∗ .
In other words, fixing α ∈ Ker⊥ T we need to find F ∈ Y
∗
such that α = F ◦ T . Since
⊥
α ∈ KerT it follows that the linear functional f : RanT → F, T x 7→ α(x), is well-defined.
f is continuous by (13b), hence can be extended continuously to some F ∈ Y ∗ . Clearly,
α = F ◦ T.
(13c)⇒(13d) Trivial.
(13d)⇒(13a) Let S : X → Z be the restriction of T to Z := RanT . Note that S ∗
is injective by (11). Every f ∈ Z ∗ can be extended to some F ∈ Y ∗ by Hahn-Banach
theorem, so
hx, T ∗ F i = hT x, F i = hT x, f i = hx, S ∗ f i , ∀x ∈ X,

63
hence RanS ∗ = RanT ∗ is closed. Therefore S ∗ : Z ∗ → RanS ∗ is open by (12). This means
that there exists r > 0 such that kS ∗ f k ≥ rkf k for every f ∈ Z ∗ . (Note that S ∗ is
injective.) Therefore S is surjective by (12), hence RanT = RanS = Z = RanT is closed.
(13a)⇔(13e) Consider the naturally defined operator S : X/KerT → RanT , x +
KerT 7→ T (x). Note that RanS = RanT . Clearly: T is range-closed if and only if S is
range-closed if and only if S is surjective if and only if S is bounded from below. This
last sentence is exactly the condition in (13e). 
We have developed some duality. To put them into action in concrete areas of analysis
we need the computations of dual spaces:

Theorem 53. (1) The dual of a Hilbert space X is isometrically isomorphic to X via
X → X ∗ , x 7→ hx, −i.
(2) The dual of Lp (X, µ), 1 < p < ∞, (X, µ) measure´ space, is isometrically isomor-
phic to Lq (X, µ), 1/p + 1/q = 1, via Lq → (Lp )∗ , f 7→ f − dµ. The same is true for
p = 1 if the measure is σ-finite.
(3) The dual of Bergman space Lpa (D) := Lp (D) ∩ {holomorphic}, 1 ≤ p < ∞, D an
open subset of Cm equipped with Lebesgue measure,
´ is isometrically isomorphic to Lqa (D),
1/p + 1/q = 1, via Lqa (D) → Lpa (D)∗ , f 7→ f − dµ.
(4) The dual of C(X), X compact Hausdorff space, is isometrically isomorphic to
M (X) (the vector space of complex´ Radon measures, normed by total variation kµk =
|µ|(X)) via M (X) → C(X)∗ , µ 7→ −dµ.1
(40 ) The dual of C([0, 1]) is isometrically isomorphic to N BV ([0, 1]) (the
Pnvector space
of functions [0, 1] → C which are of bounded variation kf kN BV := sup j=1 |f (xj ) −
f (xj−1 )| < ∞, supremum taken over all 0 = x0 < x1 < · · · < xn−1 < xn = 1, which vanish
at 0 and which are continuous ´ from the left on (0, 1)) via Riemann-Stieltjes integration
∗
BV ([0, 1]) → C(X) , ϕ 7→ −dϕ.

Proof. (1) Theorem 11.

(2) [Fol, 6.15].
(3) [Zhu-FT, 2.12].
(4) [Fol, 7.17][Rud-RCA, 6.19][DS, pages 262-5 and 258].
(40 ) [Fol, 3.29][Dou, 1.37]. 
Dual of L∞ (X, µ) and many other function spaces is given in [DS, pages 375-9]. Dual
of M (X) is hard to describe [DS, footnote of page 374][Kpl].

1
The real version is: The dual of C(X; R) is isometrically isomorphic to the vector space of all signed
Radon measures. Refer [Fol, Chapter 7] for the definition of Radon measures.

64
Chapter 8

Compact and Fredholm operators

References: [Dou, chapter 5][Rud-FA, chapter 4][Con-FA, chapter 11].

One of the most useful notions in many areas of geometry and algebra is deformation.
This chapter develops a deformation (or homotopy) theory for operators. The role of
infinitesimals is played by the so-called compact operators, and Fredholm operators are
those which are invertible up to addition by an infinitesimal. (These ideas eventually lead
to the K-theory for operator algebras [Bla].) Here is another motivation for compact
operators. If one thinks of functional analysis as a generalization of linear algebra to
infinite dimensions, then the immediate class of operators to be studied after finite rank
operators (those whose range is of finite dimension) is the class of those operators which
are norm limits of finite ranks; these are compact operators (at least on Hilbert spaces).
Theorem 54. For a bounded operator T on a Banach space X, the followings are equiv-
alent:
(1) The image of any bounded subset of X under T has compact closure. Equivalently,
if xj is a bounded sequence in X then T xj has a convergent subsequence.
(2) T is weak-to-strong continuous in the sense that if xj is a weakly convergent
sequence in X then T xj is convergent.
In case any of the conditions above holds, T is called a compact operator.
Theorem 55. For a bounded operator T on a Hilbert space X the followings are equiva-
lent:
(1) T is compact.
(2) T is the norm limit of a sequence of finite rank operators.
(3; Schmidt (or canonical)
PN representation) T can be represented as a norm con-
vergent series of the form n=1 λn hϕn , −i ψn , where N ∈ N ∪ {∞}, λn ∈ R, λn → 0,
and {ϕn }, {ψn } are orthonormal sets.
(T) RanT contains no closed infinite dimensional linear subspace.
Proposition 56 (Compacts as infinitesimals). If T is a compact operator on a separable
Hilbert space X, then for every  > 0, there exists a finite dimensional linear subspace
Y ⊆ X such that the norm of the restriction of T to Y ⊥ (namely kP T P k where P is the
orthogonal projection onto Y ⊥ ) is less than .

65
Example 57. Here are the most famous examples of compact operators:

1. A diagonal operator l2 (N) → l2 (N), (xj ) 7→ (λj xj ) is compact if and only if λj → 0.

2. All linear maps between finite dimensional normed vector spaces are compact. Ac-
cording to Theorem 2.(5), the identity map on a normed vector space X is compact if
and only if X is of finite vector space dimension.

3. Pseudodifferential operators of negative order on closed smooth manifolds are compact.

(Closed means compact and without boundary.)

4. Hilbert-Schmidt and trace class operators are compact [Arv-S, pages 70-74]. 

Theorem 58 (Riesz’s theory for compact operators). Let T be a compact operator on

an infinite dimensional Banach space X. Then:
(1) σ(T ) accumulates only at origin, so specially it is countable and contains 0.
(2) Suppose λ ∈ σ(T )\{0}. Then λ is an eigenvalue of T of finite multiplicity (namely
dim KerT −λ < ∞); λ (and also λ) is an eigenvalue of T ∗ with the same multiplicity. More
generally, not only the ordinary eigenspace KerT −λ but also the generalized eigenspace
{x ∈ X : (T − λ)n x = 0, ∃n ∈ N} is finite dimensional, and of the same dim as {x ∈ X :
(T ∗ − λ)n x = 0, ∃n ∈ N}.
If, furthermore, T is normal then:
(3) H has an orthonormal basis of eigenvectors of T , and eigenspaces corresponding
to distinct eigenvalues are orthogonal to each other.
If, furthermore, T is self-adjoint then:
(4) Eigenvalues are real and in the closed interval from −kT k to kT k, with at least
one of these endpoints being an eigenvalue.

Proof. [Rud-FA, pages 103-111]. 

Theorem 59. For a bounded operator T on a Banach space X, the followings are equiv-
alent:
(1) T is invertible modulo compacts, namely, there exists a bounded operator S on X
such that T S − 1 and ST − 1 are both compact. S is called a parametrix for T .
(2) The kernel KerT and cokernel CokerT = X/RanT of T are both finite-dimensional.
If so, then T is range-closed.
In case any of the conditions above holds, T is called a Fredholm operator of index
ind(T ) := dim KerT − dim CokerT .

Example 60. Here are the most famous examples of Fredholm operators:

1. An invertible operator plus a compact is Fredholm of index 0.

2. The unilateral forward shift operator is Fredholm of index −1.

3. Elliptic pseudodifferential operators on closed smooth manifolds are Fredholm. Their

index is given by the Atiyah-Singer index theorem [Tay-PDE, chapters 7,10][AS, AS-I].


66
 Recall from linear algebra that a square matrix A is either invertible (det A 6= 0) or it
is neither injective nor surjective (det A = 0). In other words, either Ax = b has a unique
solution for every b, or Ax = 0 has a nonzero solution; in this latter case, Ax = b has a
solution if and only if b ∈ RanA = Ker⊥ A∗ , and if Ax0 = b then the whole solution space
of Ax = b is x0 + KerA . Here is a generalization:

Theorem 61 (Fredholm alternative). Let T be a bounded operator on a Hilbert space H

which is the sum of an invertible operator (for example, the identity map) and a compact.
Then:
(1) T is a Fredholm operator of index zero. Conversely, every Fredholm operator of
index zero is an invertible plus a compact.
(2) T is range-closed (hence RanA = Ker⊥ A∗ ) and it is injective if and only if it is
surjective.
(3) Either Ax = b has a unique solution for every b or Ax = 0 has a nonzero finite
dimensional solution space N := ker A; in this latter case Ax = b has a solution if and
only if b is orthogonal to the finite dimensional space KerA∗ , and if Ax0 = b then the
whole solution space of Ax = b is x0 + N .

Proof. (1)
(2) Immediate from (1).
(3) Immediate from (1,2). 

Theorem 62. Two Fredholm operators on a Hilbert space are homotopic if and only if
they have the same index.

Here is a surprising application of the Fredholm alternative:

Application 63 (Dirichlet problem; Hilbert). [RSz] or [Fol-PDE].

67
Chapter 9

Spectral theory I: Commutative

Banach and C* algebras

References: [Rud-FA, chapters 10-11][Dou, chapters 2,4][Fol-AHA, chapter 1].

A (unital) Banach algebra (or B-algebra) is a Banach space A which is also a

C-algebra with (multiplicative) identity element 1 such that k1k = 1 and kxyk ≤ kxkkyk
for every x, y ∈ A. A (unital) C*-algebra is a Banach algebra A, equipped with a
conjugate-linear map A → A, x 7→ x∗ , called involution, such that

kx∗ k = kxk, x∗∗ = x, (xy)∗ = y ∗ x∗ , kxx∗ k = kxk2 ,

for every x, y ∈ A.

Example 64. Here are the most important examples of B and C* algebras.

1. Let G be a locally compact Hausdorff topological group, and dx denotes its left Haar
measure [Fol, 11.8]. Then the Banach space
´ L1 (G) is a Banach algebra if multiplication
is defined via convolution (f ∗g)(x) = G f (xy −1 )g(y)dy. (Recall the Young inequality:
Suppose 1 ≤ p, q, r ≤ ∞ such that 1/p+1/q = 1+1/r. Then the convolution of f ∈ Lp
with g ∈ Lq is a almost everywhere defined function h in Lr satisfying khkr ≤ kf kp kgkq .
The proof in [Fol, 8.9] for G = Rn works for the general case.) For more about these
algebras refer [Kan].

2. C(X), X a compact Hausdorff space, is a C*-algebra with involution given by pointwise

conjugation f ∗ (x) = f (x). Every commutative C*-algebra is of this form (Theorem
72).

3. B(X), X a Hilbert space, is a C*-algebra with involution given by the adjoint of

operators (Theorem 12).

4. A(K), K ⊆ C compact, the set of continuous functions K → C which are holomorphic

in the interior of K is a closed subalgebra of C(K), so a Banach algebra. If K is the
closed unit disk then A(K) is called the disc algebra.

68
5. H ∞

6. Multiplier algebras. [Kan, 1.4] 

Theorem 65 (Fundamental lemma). Let A be a Banach algebraPand x ∈ A.

(1) If kxk < 1 then 1 − x is invertible, with inverse given by j≥0 xj .
(2) If kxk < |λ|, λ ∈ C, then λ − x is invertible, with inverse given by j≥0 λ−j−1 xj .
P

(3) The set of invertible elements of A is open, and the map x 7→ x−1 is continuous
on it.

Let A be a Banach algebra. A multiplicative linear functional on A is a linear

functional ϕ : A → C such that ϕ(1) = 1 and ϕ(xy) = ϕ(x)ϕ(y) for every x, y ∈ A. From
this one can deduce that kϕk = 1. Proof. kϕk ≥ 1 because ϕ(1) = 1. By contradiction
assume x ∈ A such that |ϕ(x)| > kxk. Since x can be written (uniquely) as λ1 − y
where λ ∈ C and y ∈ Kerϕ it follows that 1 > |1 − z| where z = y/λ ∈ Kerϕ . By the
fundamental lemma there exists w ∈ X which zw = 1. This leads to the contradiction
1 = ϕ(1) = ϕ(zw) = ϕ(z)ϕ(w) = 0. Q.E.D. The set MA of all multiplicative linear
functionals on A, quipped with the weak star topology, is called the maximal ideal
space of A. Note that MA is a closed subset of the closed unit ball of the dual A∗ of
A, so compact by Alaoglu theorem. Let x be an element of Banach algebra A. The
spectrum of x, denoted by σ(x) or more precisely by σA (x), is the set of all λ ∈ C such
that λ − x is invertible. The complement of the spectrum of x is called the resolvent
set of x. The spectral radius of x, denoted by ρ(x) or more precisely by ρA (x), is the
supremum of all |λ| such that λ ∈ σ(x).

Theorem 66 (Gelfand). Let A be a Banach algebra with maximal ideal space MA . Then:
(1) The spectrum of x ∈ A is nonempty and compact.
(1; Gelfand-Mazur) If A is a division algebra (namely every nonzero element in A is
invertible) then there is a unique isometric isomorphism A → C.
(1; Gelfand-Beurling formula) The spectral radius of x is not larger than kxk. In fact,
it is given by lim kxj k1/j .
(1) The Gelfand transform X → C(MX ) is a C-algebra homomorphism which pre-
serves ∗ and is contractive in the sense that kbxk ≤ kxk for every x ∈ A.
(1; Spectral mapping theorem) The spectrum of each f (x), x ∈ A, f an entire function,
equals the image of σ(x) under f .
If A is commutative then:
(1) ϕ 7→ Kerϕ provides a one-to-one correspondence between element of MA and the
maximal ideals of A. (A nonempty subset I ⊆ A is an ideal if it is closed under subtraction
and multiplication by elements of A.) The inverse is given by A → A/m, x 7→ x + m.
Note that maximal ideals are closed.
(1) MA is nonempty.
(1) x ∈ A is invertible if and only if its Gelfand transform x
b ∈ C(MA ) is invertible.
(1) The spectrum and spectral radius of each x ∈ A can be computed by its Gelfand
transform x b ∈ C(MA ) in the following way:

σA (x) = σMA (b
x) = (range of x
b), ρA (x) = ρMA (b
x) = (uniform norm of x
b).

69
Example 67 (Spectrum). Here are the most important examples of spectrums of ele-
ments of Banach algebras:

1. The spectrum of a square matrix, as a linear map between finite dimensional vector
spaces, is the set of its eigenvalues.

2. The spectrum of f ∈ C(X) equals its range.

3. The spectrum of the unilateral forward shift (Example 13) is the closed unit disc.

4. The spectrum of f ∈ L∞ (X, µ) equals its essential range namely the set of all λ ∈ C
such that the inverse image of any disc of radius  > 0 around λ has strictly positive
measure.

5. The spectrum of the multiplication operator Mf ∈ B(L2 (X, µ)) associated to the
bounded symbol f ∈ L∞ (X, µ) equals the essential range of f .

6. The spectrum of the Toeplitz operator Tf ∈ B(H 2 (S1 )) associated to the continuous
symbol f ∈ C(S1 ) is given by the range of f union with the set of all z ∈ C such that
the winding number of f around z is nonzero.

7. The max ideal space of A(K) (Example 73) is K. [Kan]

8. L∞ (X) has the fame of having very huge and complicated max ideal space. The reason
is that its elements are not true functions, but just equivalence classes of functions.

9. Let X be a completely regular Hausdorff space. The maximal ideal space of BC(X)
(Example 1) can be identified with the Stone-Čech compactification β(X) of X [Mun,
section 38]. 

Example 68 (Empty maxiaml ideal spaces and spectra). (1) There are Banach algebras
with empty maximal ideal space.
(2) There are unbounded operators with empty spectrum. 

Example 69. ThePmaximal ideal space of X := l1 (Z) (multiplication is given by convo-

1
lution (f ∗ g)(j) = k∈Z f (kP− j)g(k)) jcan be identified with the unit circle via S → MX ,
z 7→ ϕz , given by ϕz (f ) = j∈Z f (j)z for every f ∈ X. 

Application 70 (Wiener). Let f be a nowhere-zero continuous funtion on the unit circle.

If the Fourier series of f is absolutely convergent then so is the Fourier series of 1/f .

Theorem 71 (Šilov). Let A be a Banach algebra, B ⊆ A a closed subalgebra, and x ∈ B.

Then σB (x) is obtained from σA (x) by filling some (possibly no) holes in σA (x), namely
by taking union of σA (x) with some of the bounded components of its complement.

Theorem 72 (Gelfand-Naimark). (1) Every commutative unital C*-algebra is C*-isomorphic

to some C(X), X a compact topological space. More specifically, for every commutative
unital C*-algebra A the Gelfand map A → C(MA ), a 7→ b a, ba(ϕ) = ϕ(a), is a C*-
isomorphism.

70
 (2) For every compact Hausdorff space X the Gelfand map X → MC(X) , x 7→ x b,
x
b(f ) = f (x), is a homeomorphism.
(3) The category of commutative unital C*-algebras and C*-homomorphisms between
them is equivalent to the category of compact Hausdorff spaces and continuous maps
between them, via Gelfand maps:

X → MC(X) , x 7→ x
b, x
b(f ) = f (x),

A → C(MA ), a 7→ b
a, a(ϕ) = ϕ(a).
b
(4) The category of commutative C*-algebras and C*-homomorphisms between them
is equivalent to the category of locally compact Hausdorff spaces and proper continuous
maps between them via Gelfand maps.

A von Neumann algebra (or W*-algebra) is a C*-subalgebra of some B(X), X

a Hilbert space, which is closed under weak operator topology.

Example 73. Here are the most important examples of W* algebras.

1. L∞ (X, µ) is a W*-algebra with involution given by pointwise conjugation. Every

commutative W*-algebra is of this form [Dav, II.2.9].

2. Double commutant. 

71
Chapter 10

Spectral theory II: Bounded normal

operators

References: [Rud-FA, chapters 4,12][Con-FA, chapters 2,9][Fol-AHA, chapter 1].

Theorem 74 (Spectral theorem; bounded normal operators). Let T be a bounded normal

operator on a separable Hilbert space H. Then:

1. Multiplication operator version. T is unitarily equivalent to a multiplication operator;

more precisely, there exists a σ-finite measure space (X, µ), an essentially bounded
Borel measurable function f ∈ L∞ (X, µ) and a unitary operator U : H → L2 (X, µ)
such that T = U −1 Mf U , where Mf acts on L2 (X, µ) by g 7→ f g. If H is separable
then µ can be assumed finite.1

2. Direct integral decomposition version There exists a unique σ-finite measure µ on σ(T )
´⊕
and a unitary operator U from H to the direct integral σ(T ) Hλ dµ(λ) such that T =
U −1 Mλ U , where Mλ acts on the direct integral by mapping the section s(λ) to the
´section
⊕
λs(λ). Uniqueness is in the sense that if another representation is given by
σ(T )
Hλ dµ0 (λ), after modifying µ and µ0 to make multiplicity functions λ 7→ dim Hλ
0

and λ 7→ dim Hλ0 nowhere zero, it is the case that µ and µ0 are mutually absolutely
continuous and the multiplicity functions are almost everywhere the same.

3. Projection-valued measure version. There exists a unique projection-valued measure P

defined on the Borel sigma
´ algebra of the spectrum σ(T ) of T with values in projections
on H such that T = σ(T ) λdP (λ).

4. Continuous functional calculus. The maximal ideal space of the (commutative) C*-
algebra CT∗ generated by T is homeomorphic to σ(T ), hence the Gelfand transform
CT∗ → C(σ(T )) is a ∗-isomorphic isomorphism. In other words, there exists a unique
map Φ : C(σ(T )) → B(H) with the following properties:
1
Neither X is the spectrum of T , nor f is the coordinate function (on C); however both of these are
true if T is cyclic (or multiplicity-free).

72
 (a) Φ preserves the C-algebra structure, conjugation and norm.
(b) The identity map is mapped to T .
(c) The spectrum of each Φ(f ) equals the image of σ(T ) under f . This is called
spectral mapping theorem.
(d) Each Φ(f ) is a nonnegative operator if f is a nonnegative function.
(e) If fj is a sequence of functions converging in C(σ(T )) to f then Φ(fj ) converges
Φ(f ) in norm.

5. Bounded Borel functional calculus. There exists a positive regular Borel measure µ
supported on σ(T ) and a ∗-isomorphic isomorphism Ψ from the von Neumann algebra
WT∗ generated by T to L∞ (σ(T ), µ) which extends the Gelfand map CT∗ → C(σ(T ))
of (4). In other words, there exists a unique map Ψ : L∞ (σ(T ), µ) → B(H) with the
following properties:

(a) Ψ preserves the C-algebra structure, conjugation.

(b) Ψ is contractive namely kΨ(f )k ≤ kf k for every L∞ (σ(T ), µ).
(c) The identity map is mapped to T .
(d) Each Φ(f ) is a nonnegative operator if f is a nonnegative function.
(e) If fj is a sequence of functions converging almost everywhere to f and kfj k is
bounded then Ψ(fj ) converges Ψ(f ) in the strong operator topology.

Theorem 75 (Spectral multiplicity theorem; bounded normal operators). Let Tj , j =

1, 2, be bounded normal operators on separable Hilbert spaces Hj .
(1) Assume a direct integral representations for Tj as Theorem 74.(2) with measure
µj chosen such that the Hilbert space dimension of (Hj )λ is nonzero for µj -almost every
λ. Then T1 and T2 are unitarily equivalent if and only if they have the same spectrum,
measures µ1 and µ2 are mutually absolutely continuous, and the multiplicity functions
λ 7→ dim(Hj )λ are almost everywhere the same.
(2) Assume a projection-values measure representations for Tj as Theorem 74.(3) with
projection-valued measure Pj . Then T1 and T2 are unitarily equivalent if and only if they
have the same spectrum and P1 = P2 .

73
Chapter 11

Spectral theory III: Unbounded

normal operators

References: [Rud-FA, chapter 13][Wei, chapters 4-5][dOl, chapters 1-2].

The applications of functional analysis to partial differential equations is through the

language of unbounded operators, which we develop in this chapter.

11.1 Unbounded operators

Let X and Y be Hilbert spaces over C. By an unbounded operator A : X → Y we just
mean a C-linear map A : DomA → Y defined on some linear subspace DomA ⊆ X. (So
every bounded (= continuous) operator is an unbounded operator in this terminology!)
A is called densely defined if DomA is dense in X.
1. A is called closed if the graph GA = {(f, Af ) : f ∈ DomA } of A is closed in X × Y .
Equivalently, for every sequence fj in DomA such that fj converges to f ∈ X and Afj
converges to g we must have f ∈ DomA and Af = g. (The closed graph theorem says
that an unbounded operator defined on whole X is closed if and only it is continuous,
but when DomA 6= X neither of the notions of closedness and continuity implies the
other.)
2. If A is densely defined then the adjoint of A, denoted by A∗ , is the unbounded
operator A∗ : Y → X defined as follows: DomA∗ consists of all g ∈ Y such that
hAh, gi is continuous with respect to h ∈ DomA , namely | hAh, gi | ≤ Ckhk for some
positive constant C. If so then the functional DomA → C mapping h to hAh, gi has a
unique continuous extension to X by Hahn-Banach theorem, so by Riesz representation
theorem there exists a unique f ∈ X such that hAh, gi = hh, f i, and we set A∗ g = f .
Equivalently, A∗ can be characterized by GA∗ = (JGA )⊥ where J(f, g) = (g, −f ).
3. If A is densely defined then A∗ is closed.
(Proof. GA∗ = (JGA )⊥ and the orthogonal complement of every subset of a Hilbert
space is closed.)

74
4. If A is densely defined and closed then so is A∗ , and we have A∗∗ = A.
(Proof. Since J 2 = −id and J commutes with the operations of closure and orthogonal
complement when applied to subspaces it follows that JGA⊥∗ = −GA = GA . To show
that A∗ is densely defined assume g ∈ Dom⊥ ⊥
A∗ . Since (0, g) ∈ JGA∗ = GA it follows
that g = 0. This shows that DomA∗ is dense in Y . Finally, GA∗∗ = JGA⊥∗ = JJGA⊥⊥ =
−GA = GA shows that A∗∗ = A.)

5. If A is densely defined then Ran⊥ A = KerA∗ . If A is densely defined and closed then
Ran⊥A∗ = KerA , so KerA is closed.
(Proof. The first assertion is immediate from definition. Replacing A by A∗ gives the
second.)

6. Staying in the framework of Zermelo-Frankel set theory (not using the axiom of choice)
one can not construct a noncontinuous unbounded operator X → Y which is defined on
whole X [Wri][Fol, page 179]. In other words, all concrete noncontinuous unbounded
operators are partially defined. The most important examples of unbounded operators
are differential operators, specially the d-bar operator in our case.

11.2 Spectral theorem

Theorem 76 (Spectral theorem; unbounded normal operators).

11.3 Closed range theorem for unbounded operators

Recall that in finite dimensional linear analysis (namely linear algebra) we have RanA =
Ker⊥A∗ for every matrix A, so that Au = f has a solution if and only if hf, gi = 0 for every
g with A∗ g = 0. In infinite dimenional linear analysis (namely functional analysis) we
have only RanA = Ker⊥ A∗ for densely defined closed operators A. The following theorem
says how to deal with the closure in the left hand, and gives an if and only if condition
for the solvability of Au = f .
Theorem 77 (Closed range theorem for unbounded operators). Let A : X → Y be a
densely defined closed unbounded operator between Hilbert spaces. Then:
(1) For every f ∈ Y , there exists u ∈ X with Au = f if and only if | hf, gi | ≤ CkA∗ gk
for every g ∈ DomA∗ and some C ≥ 0.
(2) For every closed subspace F ⊆ Y which F ⊇ RanA , we have F = RanA if and
only if kgk ≤ CkA∗ gk for every g ∈ DomA∗ ∩ F and some C ≥ 0.
(3) RanA is closed if and only if kgk ≤ CkA∗ gk for every g ∈ DomA∗ ∩ RanA and
some C ≥ 0.
(4) RanA is closed if and only if RanA∗ is closed.
Proof. (1) We only prove the if part because the other direction trivial. In accordance with
the general philosophy of the duality theory in functional analysis (namely understand-
ing a linear space through linear functionals living on it) one observes that our desired
u is exactly the element of X which represents the anti-linear functional RanA∗ → C

75
mapping A∗ g to hf, gi. This functional is well-defined and bounded by C according to
our hypothesis. By Hahn-Banach theorem it can be extended to a linear functional on
whole X with the same bounded operator norm. (Another way: First extend by conti-
nuity to RanA∗ and then extend to whole X by declaring the functional to vanish on the
orthogonal complement of RanA∗ .) If u ∈ X is the vector that represents this extended
functional according to the Riesz representation theorem then hf, gi = hu, A∗ gi for every
g ∈ DomA∗ . It then follows by the very definition of the adjoint that u ∈ DomA∗∗ and
A∗∗ u = f . Since A is densely defined and closed it follows that A∗∗ = A, and we are
done.
(2) For the if part, fixing arbitrary f ∈ F and g ∈ DomA∗ , according to (1) we need
to show that | hf, gi | ≤ CkA∗ gk for some C ≥ 0. Let g = g 0 + g 00 , g 0 ∈ F , g 00 ∈ F ⊥ , be
the orthogonal decomposition of g. Since F ⊇ RanA it follows that F ⊥ ⊆ Ran⊥ A = KerA∗ ,
00 0 ∗ 0 ∗
hence we deduce g ∈ KerA∗ , g ∈ DomA∗ and A g = A g. By applying our hypothesis
to g 0 we have
| hf, gi | = | hf, g 0 i | ≤ kf kkg 0 k ≤ Ckf kkA∗ gk.
Only if part. If for some g ∈ DomA∗ ∩ F we have A∗ g = 0, then g = Af ∈ RanA = F
for some f ∈ DomA and A∗ Af = 0, hence kgk2 = hAf, Af i = hf, A∗ Af i = 0, therefore
g = 0. As a result we need only show that

G := {g/kA∗ gk : g ∈ DomA∗ ∩ F, A∗ g 6= 0}

is bounded as a subset of the Hilbert space F . For every h = Af ∈ RanA = F the set
{hh, gi : g ∈ G} is a bounded subset of C with bound khk. This means that G is weakly
bounded in F . It is famous that weakly bounded subsets of Hilbert spaces are bounded.
(This is immediate from the uniform boundedness principle [Fol, 5.13]. Refer [Jos-RS,
page 85] for a direct proof. Compare [Rud-FA, 3.18].)
(3) In (2) let F be the closure of the range of A.
(4) It suffices to prove the only if part because the other direction can be deduced
from this one by replacing A with A∗ . Let RanA be closed. Then (3) gives

kgk ≤ CkA∗ gk, ∀g ∈ G, G := DomA∗ ∩ RanA . (11.1)

This inequality combined with a straightforward Cauchy sequence argument shows that
A∗ restricted to G has closed range. (Details: Assume a sequence gj ∈ G such that
A∗ gj converges to f ∈ X. Since A∗ gj is Cauchy it follows from (11.1) that gj is also
Cauchy, hence convergent to some g ∈ Y . Since A∗ has closed graph it follows that g ∈ G
and A∗ g = f .) However the range of A∗ |G equals the range of A∗ because A∗ kills the
orthogonal complement of RanA . This proves the only if part. 

76
Chapter 12

Implicit function theorem for

Banach spaces

References: [Jos-A, chapter 10]

One of the most useful observations in advanced calculus is that:

1. If real variable y depends smoothly on real variable x, y(x0 ) = y0 and dy/dx(x0 ) 6= 0,
then x depends smoothly on y locally around y0 .

2. If F (x, y) = 0 is an implicit smooth equation between real variables x and y,

F (x0 , y0 ) = 0 and ∂F/∂y(x0 , y0 ) 6= 0 then y is given by a smooth function y =
y(x) around x0 . For example, the equation of circle x2 + y 2 = 1 can be√ solved
locally around each √ of its points with y > 0 (respectively, y < 0) by y = 1 − x2
(respectively, y = − 1 − x2 ).
More generally we have:
Theorem 78 (Inverse and implicit function theorems). (1) Assuming a C k map F :
U → Rn , k ∈ {0, 1, . . . , ∞}, defined on an open U ⊆ Rn , if the n × n Jacobian matrix of
F is invertible at a point x0 ∈ U then F is a C k diffeomorphism (namely, it is bijective
and with C k inverse) on some neighborhood of x0 .
(2) Assuming the natural splitting Rm+n = Rm × Rn coordinated by (x, y), a C k map
F : U → Rn , k ∈ {0, 1, . . . , ∞}, defined on some open of Rm+n , and a point (x0 , y0 ) ∈ U
such that F (x0 , y0 ) = 0, if the n × n Jacobian matrix ∂F/∂y is invertible at (x0 , y0 )
then there exist neighborhoods V ⊆ Rm around x0 and W ⊆ Rn around y0 such that for
every x ∈ V there exists a unique y = f (x) ∈ W such that F (x, y) = 0. Furthermore,
f : V → W is C k .
These two statements are equivalent: Applying (1) to (x, y) 7→ (x, F (x, y)) gives (2).
Applying (2) to x 7→ y − F (x) gives (1). Here is the sketch of the proof of (1). Without
loss of generality one can assume x0 = 0, F (x0 ) = 0 and the Jacobian matrix of F at x0
is the identity matrix. The intuition here is that F behaves like identity maps around the
origin. Local injectivity is immediate from the mean value theorem for differentiation,

77
but to prove local surjectivity we need to use the Banach fixed point theorem: A map
T : X → X on a Banach space X which is contractive (namely kT x − T yk ≤ akx − yk
for some 0 < a < 1 and every x, y ∈ X) has a unique fixed point.
In this short chapter we formulate and prove an infinite-dimensional analogue of The-
orem 78. First we need to make sense of differentiability in infinite-dimensional spaces.
Let X and Y be normed vector spaces. A map F : U → Y defined on an open U ⊆ X is
said to be differentiable at x ∈ U if there exists a continuous linear map T : X → Y
such that kF (x+ξ)−F (x)−T (ξ)k/kξk → 0 as ξ → 0. If so, then T is unique, denoted by
dF (x), and called the total (or Frechet) derivative of F at x. F is called differentiable
on U if it is differentiable at every point of U . If F is differentiable on U then it is called
twice differentiable at x ∈ U if U → B(X; Y ), x 7→ dF (x), is differentiable at x. If so,
the total derivative of this latter map is denoted by d2 F (x); it is initially an element of
B(X; B(X; Y )), but can be naturally identified as a bilinear map X × X → Y bounded
in the sense that kd2 F (x)(ξ, ξ 0 )k ≤ Ckξkkξ 0 k for some C > 0 and every ξ, ξ 0 ∈ X. Higher
order derivatives are defined inductively. If F is differentiable on U then it is called C 1
if U → B(X; Y ), x 7→ dF (x), is continuous. If F is twice differentiable on U then it is
called C 2 if U → B(X; B(X; Y )), x 7→ d2 F (x), is continuous. C k , k ∈ {0, 1, . . . , ∞}, is
defined inductively.

Theorem 79 (Inverse and implicit function theorems for Banach spaces). (1) Let X and
Y be Banach spaces. Assuming a C 1 map F : U → Y defined on an open U ⊆ X, if the
Frechet derivative dF : X → Y is invertible (as a map between Banach spaces) at a point
x0 ∈ U then F is a C 1 diffeomorphism on some neighborhood of x0 , namely there exist
opens x0 ∈ V ⊆ U and W = F (V ) ⊆ Y such that the restriction F : V → W is bijective
and with C 1 inverse.
(2) Let X, Y and Z be Banach spaces, and let (x, y) coordinate X × Y . Assuming
a C 1 map F : U → Z defined on open U ⊆ X × Y and a point (x0 , y0 ) ∈ U such that
F (x0 , y0 ) = 0, if ∂F/∂y : Y → Z is invertible at (x0 , y0 ) then there are neighborhoods
V ⊆ X around x0 and W ⊆ Y around y0 such that for every x ∈ V there exists a unique
y = f (x) ∈ W such that F (x, y) = 0. Furthermore, f : V → W is C 1 .

Proof. (1)
(2) Apply (1) to (x, y) 7→ (x, F (x, y)). 
For more about implicit function theorems refer [KP]. Richard Hamilton (the inventor
of the Ricci flow) developed an implicit function theorem for special classes of Frechet
spaces appearing in differential geometry [Ham].

78
Chapter 13

GNS construction

References: [Con-FA, chapter 8].

Theorem 80 (Gelfand-Naimark-Segal). Every C ∗ -algebra is ∗-isomorphic to a closed

C ∗ -subalgebra of some B(X), X a Hilbert space.

Proof. 
With the same ideas one can prove a generalization of GNS, the so-called Stinespring’s
dilation theorem which characterizes completely positive maps from C*-algebras into
B(X) in terms of ∗-homomorphisms into B(Y ), Y some other Hilbert space [Pau, chapter
4].

79
Chapter 14

Semigroups

References: [Rud-FA, chapter 13][Lax, chapter 34][Bre, chapter 7].

Recall that the first-order linear constant-coefficient system of ordinary differential equa-
tions dx/dt = Ax, A a square matrix of complex numbers, can be solved using matrix
exponentials by x(t) = exp(At)x(0). In this chapter we generalize this to infinitely many
equations.
A (one-parameter) semigroup T of operators on a Banach space X is a family
Tt : X → X, 0 ≤ t < ∞, of bounded operators on X such that T0 = 1 and Tt+s = Tt ◦ Ts
for every t, s ≥ 0. The infinitesimal generator of T is the unbounded operator A :
DomA → X defined by A(x) = limt→0+ (Tt x − x)/t. T is called strongly continuous if
kTt x − xk → 0 as t → 0+ for every x ∈ X. Here is a basic fact:

Theorem 81. Let T be a strongly continuous semigroup in a Banach space X with the
infinitesimal generator A. Then:
(1) A is densely defined and closed.
(1) kTt k ≤ M exp(ωt) for some ω ∈ R and M > 0.
(1) dTt x/dt = ATt = Tt A for every x ∈ DomA .
(1) DomA = X if and only if Tt → 0 in the norm topology as t → 0+ if and only if A
is bounded and Tt = exp(tA).

Theorem 82 (Hille-Yoshida). (1) An unbounded operator A on a Banach space X gen-

erates a strongly continuous semigroup T with kTt k ≤ M exp(ωt) for some ω ∈ R and
M > 0 if and only if A is densely defined and closed, every λ > ω belongs to the resolvent
set of A, and k(λ − A)−n k ≤ M (λ − ω)−n for every positive integer n.
(2) An unbounded operator A on a Banach space X generates a strongly continuous
semigroup T with kTt k ≤ M for some M > 0 if and only if A is densely defined and
closed, every λ > 0 belongs to the resolvent set of A, and k(λ − A)−1 k ≤ M (λ − ω)−1 .

Theorem 83 (Stone). If U is a strongly continuous semigroup of unitary operators on

a Hilbert
√ space X then there exists a self-adjoint unbounded operator H on X such Ut =
exp( −1Ht).

80
References

[Ahl] Ahlfors, L., Complex analysis, Third edition, McGraw-Hill, 1979. 37, 56

[Alp] Alpay, D., (Editor), Operator theory, Springer, 2015. 5

[Apo-C] Apostol, T., Calculus, Second edition, Volumes I-II, John Wiley and Sons, 1967.
20

[Apo-A] Apostol, T., Mathematical analysis, Second edition, Addison Wesley Publishing
Co., 1974. 15, 24, 26, 38, 51

[Arv-C] Arveson, W., An invitation to C*-algebras, Springer, 1976.

[Arv-S] Arveson, W., A short course on spectral theory, Springer, 2002. 66

[AS] Atiyah, M., Singer, I., The index of elliptic operators on compact manifolds, Bull.
Amer. Math. Soc. 69 (1963), 422–433. 66

[AS-I] Atiyah, M., Singer, I., The index of elliptic operators. I, Ann. of Math. (2) 87
(1968), 484–530. 66

[Bla] Blackadar, B., K-Theory for operator algebras, Second edition, Cambridge Univer-
sity Press, 1998. 65

[BS] Böttcher, A., Silbermann, B., Analysis of Toeplitz operators, Second edition,
Springer, 2006. 5

[BG] Boutet de Monvel, L., Guillemin, V., The spectral theory of Toeplitz operators,
Princeton University Press, 1981. 5

[Bre] Brezis, H., Functional analysis, Sobolev spaces and partial differential equations,
Springer, 2011. 6, 41, 45, 50, 51, 80

[Brw] Brown, A., A version of multiplicity theory, Topics in operator theory, Pages 129–
160, Math. Surveys, no. 13, Amer. Math. Soc., 1974.

[Con-FA] Conway, J., A course in functional analysis, Second edition, Springer, 1990.
65, 72, 79

[Con-OT] Conway, J., A course in operator theory, American Mathematical Society,

2000.

81
[Dau] Daubechies, I., Ten lectures on wavelets, Society for Industrial and Applied Math-
ematics, 1992. 27

[Dav] Davidson, K., C*-algebras by example, American Mathematical Society, 1996. 71

[dOl] de Oliveira, C., Intermediate spectral theory and quantum dynamics, Birkhäuser,
2009. 74

[DiB] DiBenedetto, E., Real analysis, Second edition, Birkhäuser, 2016. 13

[Die] Diestel, J., Sequences and series in Banach spaces, Springer, 1984. 48, 55, 56

[Die] Dieudonné, J., Treatise on analysis, Volumes I–VIII, Academic Press, 1969–1993.

[Dou] Douglas, R., Banach algebra techniques in operator theory, Second edition,
Springer, 1998. 5, 6, 24, 25, 45, 64, 65, 68

[DS] Dunford, N., Schwartz, J., Linear operators, Interscience Publishers, 1963. 6, 9, 20,
40, 48, 51, 56, 64

[DS] Duren, P., Schuster, A., Bergman spaces, American Mathematical Society, 2004.
16, 26

[Fol-PDE] Folland, G., Introduction to partial differential equations, Second edition,

Princeton University Press, 1995. 67

[Fol] Folland, G., Real analysis: Modern techniques and their applications, Second edi-
tion, John Wiley and Sons, 1999. 6, 8, 9, 14, 23, 24, 26, 27, 38, 45, 51, 64, 68, 75,
76

[Fol-AHA] Folland, G., A course in abstract harmonic analysis, Second edition, CRC
Press, 2016. 59, 68, 72

[Gam] Gamelin, T., Complex analysis, Springer, 2001. 46

[Grf] Grafakos, L., Classical Fourier analysis, Second edition, Springer, 2008. 5, 13, 15

[Hall] Hall, B., Quantum theory for mathematicians, Springer, 2013. 19

[Hal-P] Halmos, P., A Hilbert space problem book, Springer, 1974. 8

[Hal-S] Halmos, P., Introduction to Hilbert space and the theory of spectral multiplicity,
AMS Chelsea Publishing, 1998. 8

[Ham] Hamilton, R., The inverse function theorem of Nash and Moser, Bulletin of the
American Mathematical Society 7(1982), 65–222. 78

[Har] Hardy, G., Weierstrass’s nondifferentiable function, Trans. Amer. Math. Soc. 17
(1916), 301–325. 13

[HK] Hoffman, K., Kunze, R., Linear algebra, Second edition, Prentice Hall, 1971. 21,
28

82
[Hör-SCV] Hörmander, L., An introduction to complex analysis in several variables, Third
edition, North-Holland Publishing Co., 1990. 37, 46, 51

[Hör-C] Hörmander, L., Notions of convexity, Birkhäuser, 1994. 44

[Jam] James, R., A non-reflexive Banach space isometric with its second conjugate space,
Proc. Nat. Acad. Sci. U.S.A. 37 (1951), 174–177. 45

[Jos-A] Jost, J., Postmodern analysis, Third edition, Springer, 2005. 77

[Jos-RS] Jost, J., Compact Riemann surfaces, Third edition, Springer, 2006. 76

[Kan] Kaniuth, E., A course in commutative Banach algebras, Springer, 2009. 68, 69, 70

[Kpl] Kaplan, S., On the second dual of the space of continuous functions, Trans. Amer.
Math. Soc. 86 (1957), 70–90. 64

[Kpy] Kaplansky, E., Set theory and metric spaces, Allyn and Bacon, 1972. 8

[Kel] Kelley, J., General topology, Springer, 1991. 54

[Kol] Kolmogorov, A., Une série de Fourier-Lebesgue divergente partout, C. R. Acad. Sci.
Paris 183 (1926), 1327–1328. 15

[Kra] Krantz, S., Function theory of several complex variables, Second edition, American
Mathematical Society, 2001. 56

[KP] Krantz, S., Parks, H., The implicit function theorem. History, theory, and applica-
tions, Reprint of the 2003 edition, Birkhäuser, 2013. 47, 78

[Lax-G] Lax, P., On the existence of Green’s function, Proc. Amer. Math. Soc. 3 (1952),
526–531. 46

[Lax] Lax, P., Functional analysis, John Wiley and Sons, 2002. 5, 9, 22, 80

[Lee] Lee, J., Introduction to smooth manifolds, Second edition, Springer, 2013. 3, 36,
37, 39

[LT] Lindenstrauss, J., Tzafriri, L., Classical Banach spaces, Volumes I–II, Springer,
1977–9. 9

[MV] Meise, R., Vogt, D., Introduction to functional analysis, Oxford University Press,
1997. 62

[Mul] Müller, V., Spectral theory of linear operators and spectral systems in Banach al-
gebras, Second edition, Birkhäuser, 2007.

[Mun] Munkres, J., General topology, Second edition, Prentice Hall, 2000. 8, 34, 36, 48,
51, 70

[Nik] Nikol’skiĭ, N., Treatise on the shift operator, Springer, 1986. 5

83
[Pau] Paulsen, V., Completely bounded maps and operator algebras, Cambridge University
Press, 2002. 79

[Pea] Pearcy, C., (Editor), Topics in operator theory, American Mathematical Society,
1979. 5

[Ped] Pedersen, G., Analysis now, Springer, 1989. 9

[RS] Reed, M., Simon, B., Methods of modern mathematical physics I. Functional anal-
ysis, Second edition, Academic Press, 1980.

[RSz] Riesz, F., Sz.-Nagy, B., Functional analysis, Reprint of the 1955 original, Dover
Publications, 1990. 67

[Rock] Rockafellar, R., Convex analysis, Princeton University Press, 1970. 44

[Roh] Rohn, J., Computing the norm kAk∞,1 is NP-hard, Linear and Multilinear Algebra
47 (2000), 195–204. 7

[Roy] Royden, H., Fitzpatrick, P., Real analysis, Fourth edition, Prentice Hall, 2010. 48,
53, 56

[Rud-PMA] Rudin, W., Principles of mathematical analysis, Third edition, McGraw-Hill,

1976.

[Rud-SCV] Rudin, W., Function theory in the unit ball of Cn , Springer, 1980. 56

[Rud-RCA] Rudin, W., Real and complex analysis, Third edition, McGraw-Hill, 1987. 5,
46, 56, 64

[Rud-FA] Rudin, W., Functional analysis, Second edition, McGraw-Hill, 1991. 21, 30,
32, 33, 34, 35, 40, 41, 43, 48, 52, 58, 60, 62, 65, 66, 68, 72, 74, 76, 80

[Sar] Sarason, D., Notes on complex function theory, Hindustan Book Agency, 1994. 46

[SFBK] Sz.-Nagy, B., Foiaş, C., Bercovici, H, Kérchy, L., Harmonic analysis of operators
on Hilbert space, Second edition, Springer, 2010. 5

[Shu] Shubin, M., Pseudodifferential operators and spectral theory, Second edition,
Springer, 2001. 5

[Tay-PDO] Taylor, M., Pseudodifferential operators, Princeton University Press, 1981. 5

[Tay-PDE] Taylor, M., Partial differential equations, Second edition, Volumes I-III,
Springer, 2011. 51, 66

[Tre] Tréves, F., Topological vector spaces, distributions and kernels, Academic Press,
1967. 62

[Upm] Upmeier, H., Toeplitz operators and index theory in several complex variables,
Birkhäuser, 1996. 5

84
[Web] Webster, R., Convexity, Oxford University Press, 1994. 44

[Wei] Weidmann, J., Linear operators on Hilbert spaces, Springer, 1980. 8, 18, 19, 74

[Wri] Wright, M., All operators on a Hilbert space are bounded, Bull. Amer. Math. Soc.
79 (1973), 1247–1250. 75

[Zhu-OT] Zhu, K., Operator theory in function spaces, Marcel Dekker Inc., 1990. 5

[Zhu-FT] Zhu, K., Spaces of holomorphic functions in the unit ball, Springer, 2005. 8,
26, 64

[Zyg] Zygmund, A., Trigonometric series, Third edition, Volumes I-II, Cambridge Uni-
versity Press, 2002.

15

85