Download as pdf or txt
Download as pdf or txt
You are on page 1of 134

FUNCTIONAL ANALYSIS

(MTH3C13)

SELF LEARNING MATERIAL

III SEMESTER (Core Course)

M.Sc. MATHEMATICS
(2019 Admission)

UNIVERSITY OF CALICUT
SCHOOL OF DISTANCE EDUCATION
CALICUT UNIVERSITY P.O.
MALAPPURAM - 673 635, KERALA

190563
School of Distance Education
University of Calicut
Self Learning Material
III Semester (Core Course)
M.Sc. Mathematics
(2019 Admission)

MTH3C13 : FUNCTIONAL ANALYSIS

Prepared by:
Dr. VINOD KUMAR P.
Associate Professor
Department of Mathematics
T.M. Government College, Tirur.

Scrutinized by:
Dr. BIJUMON R.
AssociateProfessor
Assistant Professor &&Head
Head
Department of Mathematics
Mahatma Gandhi College, Iritty, Kannur.

DISCLAIMER

"The author(s) shall be solely responsible


for the content and views
expressed in this book".
MTH3C13 : FUNCTIONAL ANALYSIS

Contents
1 Introduction 3
2 Normed Spaces 7

2.1 Basics of Linear Algebra . . . . . . . . . . . . . . 8

2.2 Normed Spaces : Definition and Examples . . . . 17

2.3 Banach Spaces . . . . . . . . . . . . . . . . . . . 35


3 Hilbert Spaces 47

3.1 Basic Notions . . . . . . . . . . . . . . . . . . . . 48


3.2 Orthonormal Systems . . . . . . . . . . . . . . . . 56
3.3 Orthogonal Bases . . . . . . . . . . . . . . . . . . 62
3.4 Orthogonal Decomposition . . . . . . . . . . . . . 69
3.5 Linear Functionals . . . . . . . . . . . . . . . . . 75
3.6 Bounded Linear Functionals . . . . . . . . . . . . 77

4 The Dual Space of a Normed Space 84


4.1 The Hahn-Banach Theorem and its Consequences. 85
4.2 Reflexive Spaces . . . . . . . . . . . . . . . . . . . 90
MTH3C13 : FUNCTIONAL ANALYSIS

5 Bounded Linear Operators 97


5.1 The Space of Bounded Linear Operators . . . . . 98
5.2 Compact Operators . . . . . . . . . . . . . . . . . 109
5.3 Dual Operators . . . . . . . . . . . . . . . . . . . 116
5.4 Finite Rank Operators . . . . . . . . . . . . . . . 121
5.5 Different Notions of Convergences in L(X) . . . . 124
5.6 Invertible Operators . . . . . . . . . . . . . . . . 127

References.........................................131

The only way to learn Mathematics is


to do Mathematics!
Paul Halmos
MTH3C13 : FUNCTIONAL ANALYSIS

Chapter 1
Introduction

Functional Analysis is one of the highest abstractions in the


main stream of Mathematics where generalization and abstrac-
tion not only became the most fashionable, but side by side it
became unreasonably effective and unpredictably relevant. It is
in this hard core area, one sees the unity and the largest inter-
play of the different branches of Mathematics. It is not possible
to imagine any future development of Mathematics or its ap-
plications which will not bear on it the impact of the thought
MTH3C13 : FUNCTIONAL ANALYSIS

processes of this area of human knowledge. ’Functional Anal-


ysis’ may be thought of as the study of certain topological -
algebraic structures and of the methods by which knowledge of
these structures can be applied to analytic problems. It can be
considered as a rare blend of the three important branches of
Mathematics, namely, algebra, analysis and topology.
The idea of Functional Analsis was first introduced in 1887
by Volterra in his work on Integral equations. Early contribu-
tions to Functional analysis were made by J. Hadamard around
1903, and he also suggested the name Functional Analysis. The
axioms for a normed space as is stated here goes back to the
work of F. Riesz on integral equations. The beauty of a normed
space, which is the basic structure considered in functional anal-
ysis, is that one can talk of unit balls and spheres even when
there is no obvious Euclidean context. In fact, the purpose of
functional analysis is to introduce a geometric context where
there was apparently none earlier. It is the geometric context
in analysis which helps us talk about approximations, limits,
maximum, optimization etc., and manipulate with them as we
would do in a Euclidean space. The construct Banach space,
normed spaces which are complete, is one of the most impor-
tant objects of study of 20th century Mathematics and has al-

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

ready permeated several areas of applications. It is an important


tool for scientific investigations in Mathematical physics, elec-
trical engineering, and optimization as Euclidean geometry and
Trigonometry are for surveying.
The notion of inner product, which is a generalization of the
dot product of vectors in Euclidean spaces, helps to introduce
orthogonality in arbitrary vector spaces. Inner product spaces
which are complete, are known as Hilbert spaces. The axioms
for abstract Hilbert space were first enunciated by John Von
Neumann (1903 - 1957) in 1927, in his monumental exposition of
the mathematical formulation of Quantum Mechanics. A large
area of current research interest is centered around the theory
of operators on Hilbert spaces.
The concept of completeness which is crucial in Banach and
Hilbert spaces, is important in the following philosophical sense.
Analysis is the study of limit processes, where we are not talking
of numbers alone, but functions in general. In fact, what we
study could be applied to complex phenomena, available in real
world or in scientific investigations. A scientific investigation
make a finite number (though large) of observations and usually
wants to conclude something about the ideal situation, which

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

is beyond the finite number of experiments. Thus from the


finite to the infinite, there is a process of approximations. It
is a limit process of phenomena, not of numbers. Functional
analysis attempts to give a model for this limit process and in
many surprising instances, it has not only succeeded in giving a
model, but it has come out with such approximations which are
workable. In every one of these manipulations with the process
of going from the finite to infinite, one has to be, in some sense
assured the existence of the limiting case, within the frame -
work of the space under consideration. This is the requirement
of “completeness”.
Functional Analysis plays an increasing role in the applied
sciences as well as in Mathematics itself. As Functional Analysis
is a blend of algebra, analysis and topology, for enjoying Func-
tional Analysis, the student should have rather a good back-
ground in these areas.
In Module I (Chapter 2), we recall some basics of linear al-
gebra, and introduce normed spaces. In Module II (Chapter 3),
we study about Hilbert spaces, and bounded linear functionals.
The dual space of a normed space is introduced in Module III
(Chapters 4 and 5). Also, we study about bounded linear op-
erators and realize it as a normed space endowed with operator
norm and consider different types of operators.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

Chapter 2
Normed Spaces

Students are advised to revise the basic ideas of vector spaces


and linear transformations that they have studied in their Linear
Algebra course.
Linear algebra is the branch of Mathematics concerned with
the study of vectors, vector spaces (or linear spaces) and lin-
ear transformations (or linear mappings) between vector spaces.
The ideas behind the abstract notion of a vector space occurred
in many concrete examples during the nineteenth century and
earlier. It was Descartes and Fermat who first discussed the

7
MTH3C13 : FUNCTIONAL ANALYSIS

2.1. Basics of Linear Algebra 8

vector spaces R2 and R3 in much the way that are presented to-
day. In the study of vector spaces, we generalize the geometric
concept of a vector as a line segment of given length and direc-
tion advantageously in an abstract way. The modern definition
of a vector space seems to be due to the Italian mathematician
Giuseppe Peano. In Functional Analysis, we study about linear
spaces endowed with some additional structures that helps to
study about continuity and convergence, and some aspects of
geometry.
In what follows, F will denote either R, the field of real
numbers or C, the field of complex numbers.

2.1 Basics of Linear Algebra

Recall that a linear space (or a vector space) over F is a


non empty set E along with a function + : E × E → E, called
addition and a function . : F × E → E, called scalar multi-
plication, such that

• (E, +) is an abelian group,

• 1. k.(x + y) = k.x + k.y

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.1. Basics of Linear Algebra 9

2. (k + l).x = k.x + l.x


3. (kl)x = k.(lx)
4. 1.x = x, for all x, y ∈ E, and for all k ∈ F.

The space Fn of all n−tuples of elements of F is an example


for a linear space over the field F.

Exercise 2.1. Verify that Fn is a linear space over the field F.

A nonempty subset E1 of a linear space E is said to be


subspace of E if kx + y ∈ E1 whenever x, y ∈ E1 , and k ∈ F.
i.e., E1 is closed relative to the linear space operations in E, and
in this case we write E1 ,→ E.
The set consisting of only the zero vector of E is a subspace of
E, called the zero subspace of E, and is denoted by 0.
A subspace E1 of a linear space E is said to be a proper if
6 E.
E1 =

Example 2.1.1. Let s be the set of all scalar sequences (ai )∞


i=1 ,
where ai ∈ F. Then s is a linear space with respect to the
addition of sequences and multiplication with real (or complex)
numbers.
The set `∞ of all bounded scalar sequences is a subspace of s.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.1. Basics of Linear Algebra 10

Also, since any convergent sequence is bounded, the set c of


all convergent sequences is a subspace of `∞ , and the space
c0 of all sequences that converges to zero, is a subspace of c.
Let s∗ be the set of all sequences with finite support (i.e., the
set of all sequences with all but finite zero elements.) Then, a
sequence (ai )∞ ∗
i=1 ∈ s if and only if there exists N ∈ N such that
ai = 0 for all i > N. Any such sequence has only finitely many
non zero terms and hence converges to 0. So s∗ is a subspace of
c0 . (Some authors denotes the space s∗ as c00 .)
Thus, we have s∗ ⊆ c0 ⊆ c ⊆ `∞ ⊆ s.

Exercise 2.2. Show that the inclusions s∗ ⊆ c0 ⊆ c ⊆ `∞ ⊆ s


are proper.

If E1 is a subset of a linear space E, the linear span of E1 ,


denoted by span E1 is the smallest subspace of E that contains
the set E1 . Equivalently, it is the intersection of all subspaces of
E containing E1 .
If E1 is non empty, span E1 is the set of all (finite) linear
combinations of elements of E1 . i.e.,
span E1 = {k1 x1 +k2 x2 +...+kn xn |k1 , k2 , ...kn ∈ F, x1 , x2 , ..., xn ∈
E1 }.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.1. Basics of Linear Algebra 11

If span E1 = E, we say that E1 spans E.


A set of vectors x1 , x2 , ..., xn of a linear space E is said to be
linearly dependent if there exists numbers a1 , a2 , ..., an , not
all of them are zeros, such that a1 x1 + a2 x2 + ... + an xn = 0.
It is called linearly independent if it is not linearly de-
pendent.
Thus, a set of vectors x1 , x2 , ...xn of a linear space E is linearly
independent if ni=1 ai xi = 0 implies that all ai = 0.
P

A subset B of a linear space E is called a basis for E if span


B = E and B is linearly independent.
Equivalently; a basis is a maximal linearly independent set.
If B is a basis for the linear space E, any element of E, can
be expressed uniquely as a linear combination of elements of B
(basis vectors).
If a linear space E has a basis consisting of a finite number
of elements, then E is said to be finite dimensional and the
number of elements in a basis for E is called the dimension of
E.
The space {0} is said to have zero dimension.
If a linear space contains an infinite linearly independent subset,
then it is said to be infinite dimensional.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.1. Basics of Linear Algebra 12

Example 2.1.2. The set {e1 , e2 , ..., en }, in Fn with,


ei = (0, 0, ..., 0, 1, 0, ...0), where 1 occurs only at ith place, is a
basis for the space Fn of all n−tuples of elements of F, so that
dimension of Fn is n.

Example 2.1.3. Let C[0, 1] be the set of all continuous func-


tions defined on the interval [0, 1]. Then C[0, 1] is a linear space
with respect to the operations point wise addition and scalar
multiplication. This space is infinite dimensional since the func-
tions pn (t) = tn , n ∈ N are linearly independent.

Let E1 and E2 be linear spaces over F. A linear map from


E1 to E2 is a function A : E1 → E2 such that A(ax + by) =
aA(x) + bA(y) for all x, y ∈ E1 and a, b ∈ F.
We often write Ax instead of A(x).
Two important sets are associated with such a linear trans-
formation; namely
ker(A) = {x ∈ E1 | Ax = 0}, called the kernel of A, and,
ImA = {Ax | x ∈ E1 }, called the image of A.
It is easy to check that kerA is a subspace of E1 , and Im(A)
is a subspace of E2 . Also, A is one-to-one if and only if kerA = 0.
(Here 0 stands for the zero subspace of E1 .)

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.1. Basics of Linear Algebra 13

A linear map A : E1 → E2 is an isomorphism if kerA = 0


and ImA = E2 . i.e., if and only if A is one-to-one and onto.
In this case, A is invertible with A−1 : E2 → E1 , which is also
linear.

Exercise 2.3. Show that any finite dimensional vector space is


isomorphic to Fn for some n ∈ N.

If E is a linear space over F and if E1 is a subspace of E, we


can define a new linear space; the quotient space of E with
respect to E1 , denoted by E/E1 as follows:

E/E1 = {[x] | x ∈ E},

where [x] = x + E1 = {x + y | y ∈ E1 }, called cosets of


E relative to the subspace E1 . The vector space operations in
E/E1 are [x] + [y] = [x + y] and k[x] = [kx], for x, y ∈ E, k ∈ F.
The identity element for addition is [0] = {0 + y | y ∈ E1 } = E1 .
Note that x ∈ [x] for all x ∈ E, since x = x + 0 and 0 ∈ E1 ,
as it is a subspace. If x ∈ E1 , then [x] = [0] = E1 . Also, if
x, y ∈ E, then the two cosets [x], [y] either coincides or they are
disjoint sets.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.1. Basics of Linear Algebra 14

To see this, let z ∈ [x] ∩ [y]. Then z = x + y1 and z = y + y2


for some y1 , y2 ∈ E1 . Therefore, z − x and z − y belongs to E1 ,
and y − x = (z − x) − (z − y) ∈ E1 . If a ∈ [x], then a − x ∈ E1 ,
so that (a − x) − (y − x) = a − y ∈ E1 and a = y + (a − y) ∈ [y].
Thus [x] ⊆ [y]. Similarly, we get [y] ⊆ [x]. Hence, if [x], [y] are
not disjoint, then [x] = [y].

Exercise 2.4. Show that [x] = [y] if and only if x − y ∈ E1 .

The dimension of the quotient space E/E1 is called the codi-


mension of E1 with respect to E and is denoted by codimE E1 .

Example 2.1.4. Consider the quotient space c/c0 . (See Ex-


ample 2.1.1) For any sequence x = (xn ) in c, we can write
x + c0 = a(1 + c0 ), where a = lim xn and 1 is the constant
n→∞
sequence with all terms as 1. Thus, [x] = a[1] for any x ∈ c, so
that {[1]} forms a basis for c/c0 and hence codimc c0 = 1.

A set of vectors x1 , x2 , ..., xn of a linear space E is said to


be linearly independent relative to the subspace E1 if
Pn
ai xi ∈ E1 implies ai = 0 for all i. Note that the set of vectors
i=1
x1 , x2 , ..., xn is linearly independent if and only if it is linearly
independent relative to the zero subspace.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.1. Basics of Linear Algebra 15

We now use this concept to derive a characterization of dimen-


sion of E/E1 .

Lemma 2.1.5. Let E be a linear space over F and E1 be a


subspace of E. Then, the dimension of E/E1 is n if and only
if there exists a set of vectors x1 , x2 , ..., xn linearly independent
relative to E1 such that for every x ∈ E, there exists a unique
set of numbers a1 , a2 , ..., an and a unique vector y ∈ E1 such that
n
P
x= ai xi + y. Moreover, the cosets [x1 ], [x2 ], ..., [xn ] forms a
i=1
basis for E/E1 .

Proof. Suppose that dimension of E/E1 is n. Choose a basis


[x1 ], [x2 ], ..., [xn ] for E/E1 . We claim that the set of vectors
x1 , x2 , ..., xn is linearly independent relative to E1 .
Pn
Let ai xi ∈ E1 . Then,
i=1

n
X n
X
ai [xi ] = [ ai xi ] = [0] ⇒ ai = 0, ∀i.
i=1 i=1

Hence, the vectors x1 , x2 , ..., xn are linearly independent relative


to E1 . Also, for any [x] ∈ E/E1 ,, there are numbers a1 , a2 , ..., an
Pn Pn Pn
such that [x] = ai [xi ] = [ ai xi ], implies x = ai xi + y for
i=1 i=1 i=1

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.1. Basics of Linear Algebra 16

some y ∈ E1 .
To see that this representation is unique, suppose we also have
Pn Pn
x= bi xi + z with z ∈ E1 . Then (ai − bi )xi = z − y ∈ E1 .
i=1 i=1
Since the vectors x1 , x2 , ..., xn are linearly independent relative
to E1 , this shows that ai = bi , ∀i and y = z.
For the converse part, assume that for each x ∈ E, we have
Pn
a unique expression x = ai xi + y with y ∈ E1 , and with the
i=1
set of vectors x1 , x2 , ..., xn are linearly independent relative to
E1 . Note that
n
X n
X n
X
ai [xi ] = [ ai xi ] = [0] ⇒ ai xi ∈ E1 ⇒ ai = 0, ∀i.
i=1 i=1 i=1

This implies that [x1 ], [x2 ], ..., [xn ] are linearly independent in
Pn
E/E1 . Also, for any [x] ∈ E/E1 , by assumption, x = ai x i + y
i=1
with y ∈ E1 . Thus,
n
X n
X n
X
x− ai xi ∈ E1 ⇒ [x] = [ ai xi ] = ai [xi ].
i=1 i=1 i=1

⇒ {[x1 ], [x2 ], ..., [xn ]} is a basis for E/E1 and dimE/E1 = n.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.2. Normed Spaces : Definition and Examples 17

2.2 Normed Spaces : Definition and


Examples

In this session, we introduce the concept of normed spaces


which is fundamental to the development of the theory of Ba-
nach spaces, which plays a key role in functional analysis. The
notion of normed spaces can be thought of as a generalization of
the n-dimensional space Fn with the Euclidean length given by
n 1
( x2i ) 2 . The concept of norm in a linear space can be used to
P
j=1
define a metric and hence the notion of convergence in a linear
space, which is necessary, for example to give a meaning to the
sum of an infinite series.
Also we shall see that every normed space is a metric space
and therefore a topological space; thus it is natural that the
topological concepts such as open subset, closed subset, limit,
closure, denseness, compactness, separability, etc., make sense
on a normed space. Among the class of normed spaces, the most
important ones are the complete normed spaces which are
universally known as Banach spaces.
Recall that a metric d on a nonempty set E is a function
d : E × E → R such that for all x, y, z ∈ E,

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.2. Normed Spaces : Definition and Examples 18

• d(x, y) ≥ 0 and d(x, y) = 0 if and only if x = y

• d(x, y) = d(y, x)

• d(x, y) ≤ d(x, z) + d(z, y) ( Triangle inequality .)

A metric space is a non empty set E along with a metric on


it.
For example, the modulus or absolute value | · | defines a
metric on F, namely d(x, y) = |x − y|, for x, y ∈ F.

Definition 2.2.1. Let E be a linear space over F. A norm on


E is a function ||.|| from E to R such that for all x, y ∈ E and
for all λ ∈ F,

1. ||x|| ≥ 0 and ||x|| = 0 if and only if x = 0. - [Positive


definiteness]

2. ||λx|| = |λ| ||x|| - [ Homogeneity of norm]

3. ||x + y|| ≤ ||x|| + ||y|| - [Triangle inequality]

A normed space is a linear space with a norm on it.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.2. Normed Spaces : Definition and Examples 19

We denote a linear space E equipped with a norm || · || as


X, i.e., X = (E, || · ||).
In a normed space X, we define a metric by d(x, y) = ||x−y||
for x, y ∈ X.

Exercise 2.5. Verify that d is a metric on X.

Definition 2.2.2. Let || · ||1 and || · ||2 be any two norms on a


linear space E. Then the norm, ||·||1 is said to be stronger than
|| · ||2 if there exists a constant C > 0 such that ||x||2 ≤ C ||x||1
for all x ∈ E.
The norms || · ||1 and || · ||2 are said to be equivalent if there
exist constants C, c > 0 such that c ||x||2 ≤ ||x||1 ≤ C ||x||2 for
all x ∈ E.

Exercise 2.6. Prove that any subspace Y of a normed space X


is again a normed space with the inherited norm (obtained by
restricting the norm in X to the space Y ).

Example 2.2.3. On the linear space `∞ of all bounded scalar


sequences, define || · ||∞ as the supremum of the absolute values
of the terms of the sequences; i.e., ||x||∞ = sup |ai | for any
i
x = (ai )∞
i=1 ∈ `∞ . It is easy to verify that this is a norm on `∞ .

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.2. Normed Spaces : Definition and Examples 20

By Exercise 2.6, it follows that s∗ , c0 , c are all normed spaces


with this norm.

Example 2.2.4. Consider C[0, 1], the space of all continuous


functions defined on the interval [0, 1]. For f ∈ C[0, 1], define
||f ||∞ = sup |f (t)|. This makes C[0, 1] a normed space. The
t∈[0,1]
space of all polynomial functions P [0, 1] is a subspace of this
normed space.

Exercise 2.7. Verify that ||f ||∞ = sup |f (t)| is a norm on the
t∈[0,1]
space C[0, 1].

Before listing more examples for normed spaces, we now


state and prove two important inequalities: Hölder’s inequal-
ity and Minkowski’s inequality.

Theorem 2.2.5. Hölder’s inequality for Sequences


For every sequence of scalars (ai ) and (bi ) and for 1 < p < ∞,
we have
X X X 1 X 1
| ak b k | ≤ |ak bk | ≤ ( |ak |p ) p ( |bk |q ) q ,

where p1 + 1q = 1. (The case of this inequality with p = q = 2 is


known as the Cauchy-Schwartz inequality.)

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.2. Normed Spaces : Definition and Examples 21

Proof. If is enough to prove this inequality for every finite set of


numbers (an )N N
n=1 and (bn )n=1 , because by letting N → ∞ in the
finite case, we see that the above inequality holds for countable
sets too.
p
From p1 + 1q = 1, we get p + q = pq ⇒ p = (p − 1)q and p−1 = q,
1
subtracting 1 from both sides, we obtain p−1 = q − 1.
1 1
Set ci = |ai |/( |aj |p ) p and di = |bi |/( |bj |q ) q , then cpi =
P P P

1, dqi = 1, and to prove the Hölder’s inequality, we have to


P

ci di ≤ 1. Consider the function y = xp−1 and in-


P
show that
tegrate it with respect to x from x = 0 to x = ci . This integral
represents the area S1 in the figure 2.1. Similarly, integrating
the inverse x = y 1/(p−1) = y q−1 from y = 0 to y = di , we get the
area S2 in the figure 2.1.
Also, from the figure 2.1, it is clear that the sum of the areas
S1 + S2 always exceeds the area of the rectangle,namely ci di .
Rc Rd
Thus, ci di ≤ S1 + S2 = 0 i xp−1 dx + 0 i y q−1 dy = p1 cpi + 1q dqi ,
with equality only if di = cp−1
i .
Taking sum of these inequalities on i, we get ci di ≤ p1 cpi +
P P
1
P q P p P q
q
di ≤ 1, since ci = 1 and di = 1. This completes the
proof.
1
Remark 2.2.6. By formally defining ( |bk |q ) q to be sup |bk |
P

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.2. Normed Spaces : Definition and Examples 22

Figure 2.1: ci di ≤ S1 + S2

for the case q = ∞, we can interpret the Hölder’s inequality for


all p with 1 ≤ p ≤ ∞.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.2. Normed Spaces : Definition and Examples 23

Theorem 2.2.7. Hölder’s inequality for Functions


Let 1 < p < ∞ and let q be such that p1 + 1q = 1. Then, for
all functions f, g on an interval [a, b] such that the integrals
Rb Rb Rb
a
|f (t)|p dt, a |g(t)|q dt and a |f (t)g(t)|dt exist ( as Riemann
integrals), we have
Z b Z b Z b
p 1/p
|f (t)g(t)|dt ≤ ( |f (t)| dt) ( |g(t)|q dt)1/q .
a a a

Rb
Remark 2.2.8. If a f (t)g(t)dt also exists, then from the in-
Rb Rb
equality, | a f (t)g(t)dt| ≤ a |f (t)g(t)|dt, it follows that
Z b Z b Z b
p 1/p
| f (t)g(t)dt| ≤ ( |f (t)| dt) ( |g(t)|q dt)1/q .
a a a

Proof. Consider a partition a = x0 < x1 < ... < xn = b of the


interval [a, b] into n equal subintervals of length ∆x = b−a
n
. We

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.2. Normed Spaces : Definition and Examples 24

have, by using the Hölder’s inequality for sequences,


n n
X X 1 1
|f (xi )g(xi )|∆x = |f (xi )g(xi )|(∆x) p + q
1=1 1=1
n
X 1 1
= (|f (xi )|p ∆x) p (|g(xi )|q ∆x) q
1=1
n n
X 1 X 1
≤ ( |f (xi )|p ∆x) p ( |g(xi )|q ∆x) q
1=1 1=1

As n → ∞, the sums will be replaced by the integrals, and we


get the desired inequality.

Using the Hölder’s inequality, we now prove the Minkowski’s


inequality.

Theorem 2.2.9. Minkowski’s inequality for Sequences


For every sequence of scalars a = (ai ) and b = (bi ) and for
1 ≤ p ≤ ∞, we have

ka + bkp ≤ kakp + kbkp ,

1
|ai |p ) p for 1 < p < ∞, and kak∞ = sup |ai |.
P
where, kakp = (
i

Proof. The case p = 1 follows from the triangle inequality |ai +

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.2. Normed Spaces : Definition and Examples 25

bi | ≤ |ai | + |bi | for all ai , bi ∈ F. The case p = ∞ also fol-


lows similarly from the subadditivity of the maximum (supre-
mum)function.
To prove the case 1 < p < ∞, we have,
X
ka + bkpp = |ai + bi |p
X
≤ (|ai + bi |)p−1 (|ai | + |bi |)
X X
= (|ai + bi |)p−1 |ai | + (|ai + bi |)p−1 |bi |
X 1 X 1
≤ ( (|ai + bi |)p ) q ( (|ai |)p ) p
X 1 X 1
+( (|ai + bi |)p ) q ( (|bi |)p ) p
X 1
= ( (|ai + bi |)p ) q (kakp + kbp k)

where the first inequality follows from the triangle inequality


of modulus and the last inequality is obtained by applying the
Hölder’s inequality, and noting that (p − 1)q = p. Thus,
X X 1
ka + bkpp = |ai + bi |p ≤ ( (|ai + bi |)p ) q (kakp + kbp k).

1
(|ai + bi |)p ) q , we get
P
Dividing both sides by (

(ka + bkpp )1−1/q ≤ kakp + kbp k ⇒ ka + bkp ≤ kakp + kbp k,

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.2. Normed Spaces : Definition and Examples 26

1
since 1 − q
= p1 .

Theorem 2.2.10. Minkowski’s inequality for Functions


Let 1 < p < ∞ and let q be such that p1 + 1q = 1. Then, for
all functions f, g on an interval [a, b] such that the integrals
Rb p
Rb p
Rb
a
|f (x)| dx, a
|g(x)| dx exist, the integral a
|f (x) + g(x)|p dx
exists too, and it satisfies
Z b Z b Z b
p 1/p p 1/p
( |f (x)+g(x)| dx) ≤ ( |f (x)| dx) +( |g(x)|p dx)1/p .
a a a

Proof. Consider a partition a = x0 < x1 < ... < xn = b that


divides the interval [a, b] into n equal subintervals of length
∆x = b−an
. We have, by using the Minkowski’s inequality for
sequences,

Xn Xn n
X
p 1/p p 1/p
( |f (xi )+g(xi )| ∆x) ≤ ( |f (xi )| ∆x) +( |g(xi )|p ∆x)1/p
i=1 i=1 1=1

As n → ∞, the sums will be replaced by the integrals, and we


get the desired inequality.

Example 2.2.11. For 1 ≤ p < ∞, let `np denotes the linear


space of all n-tuples of real numbers equipped with the norm
n
kxkp = ( |xi |p )1/p , where x = (x1 , x2 , ..., xn ), i.e., `np = (Rn , k ·
P
i=1

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.2. Normed Spaces : Definition and Examples 27

kp ). Here, the triangle inequality of the norm follows from the


Minkowski’s inequality. Similarly, `n∞ denotes the normed space
(Rn , k · k∞ ), where kxk∞ = max |xi |.
i

Also, for 1 ≤ p < ∞, let `p denotes the linear space of all


p- summable sequences of real numbers (i.e., all sequences x =

|xi |p < ∞.) equipped with the
P
(x1 , x2 , ..., xn , ....), for which
i=1

p 1/p
. i.e., `p = (R∞ , k · kp ). The facts that
P
norm kxkp = ( |xi | )
i=1
this space `p is closed under addition and the triangle inequality
of the norm k · kp follows from the Minkowski’s inequality.

Exercise 2.8. If 1 ≤ p < q < ∞, prove that `p ( `q .

The norm on a linear space induces a metric and that helps


us to introduce topological and geometric notions.
We say that a sequence (xn ) in a normed space X converges
to a point x ∈ X if and only if kxn − xk → 0, as n → ∞.
From the triangle inequality of norm, it follows that norm is
a continuous function, i.e., if xn → x in X, then kxn k → kxk.
Using triangle inequality of norm, one have | kxk − kyk | ≤ kx −
yk, and this in fact, shows that norm is uniformly continuous.
Let X be a normed space, x0 ∈ X and r > 0. The set

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.2. Normed Spaces : Definition and Examples 28

Dr (x0 ) = {x | kx − x0 k < r} is said to be the open ball with


center at x = x0 and of radius r.
A set O ⊆ X is said to be open if and only if for every
x ∈ O, ∃r > 0 such that Dr (x) ⊆ O.
A subset F of X is said to be closed if for every sequence
xn ∈ F that converges to some x ∈ X, implies that x ∈ F.

Lemma 2.2.12. A set O in a normed space X is open if and


only if its complement F = Oc is closed.

Proof. Let O be an open set in X. Let F = Oc and let (xn ) be


a sequence in F and xn → x ∈ X. Suppose x ∈ O. Since O is
open, ∃r > 0 such that the open ball Dr (x) ⊆ O. Since xn → x,
for large n, kxn − xk < r ⇒ xn ∈ O, which is a contradiction,
since xn ∈ F = Oc . This proves that x ∈ F ⇒ F is closed.
Conversely assume that F is closed and O = F c . Let x ∈ O.
We claim that ∃ > 0 such that D (x) ⊆ O. If not, for each
decreasing sequence n > 0, converging to 0, Dn ∩ F is non-
empty, so that ∃xn ∈ F such that kxn − xk < n ⇒ xn → x.
Since F is closed, we get x ∈ F. This contradiction shows that
D (x) ⊆ O. Since x ∈ O is arbitrary, this implies O is open.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.2. Normed Spaces : Definition and Examples 29

Let X be a normed space and A ⊆ X. We define the closure


of A as the set consisting of all x ∈ X such that there exists a
sequence (xn ) in A with xn → x, and is denoted by A. For any
x ∈ A, we have the constant sequence xn = x, ∀n, converging to
x, so that A ⊆ A. Note that A is closed if and only if A = A.
We say that A is a dense set in X, if A = X.

Exercise 2.9. In any normed space X, show that the empty


set and the whole set X are both open and closed.

Exercise 2.10. In any normed space X, show that union of any


arbitrary collection of open sets and intersection of any finite
collection of open sets are open.

Exercise 2.11. Show that A is the intersection of all closed sets


containing A.

Let X be a normed space and x, y ∈ X. Then the set of all


points {λx + (1 − λ)y | 0 ≤ λ ≤ 1} is geometrically the line
segment joining x and y. We call this set an interval and is
denoted by I[x, y].
A subset M of a linear space E is said to be convex if
and only if for every two points x, y ∈ M, the interval I[x, y] is
contained in M.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.2. Normed Spaces : Definition and Examples 30

Exercise 2.12. Show that intersection of a family of convex


sets in a linear space is convex.

If E is a normed space, the set

D(E) = {x ∈ E | kxk ≤ 1},

is called the unit ball of the space E. Since kxk = k − xk, we


see that x ∈ D(E) if and only if −x ∈ D(E), i.e., the unit ball
is symmetric with respect to the origin.
The unit ball D(E) is a convex set. To see this, let x, y ∈
D(E). Then for any 0 ≤ λ ≤ 1,

kλx + (1 − λ)yk ≤ λkxk + (1 − λ)kyk ≤ λ + (1 − λ) = 1.

Therefore, I[x, y] ⊆ D(E). ⇒ D(E) is convex.


The set S(E) = {x ∈ E | kxk = 1 is called the unit sphere
of E.
A subset M of a normed space E is said to be bounded if
there exists C > 0 such that kxk ≤ C, ∀x ∈ M. Equivalently,
M ⊂ C · D(E).
In Exercise 2.6, we have noted that any subspace of a normed

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.2. Normed Spaces : Definition and Examples 31

space is again a normed space. Now we will discuss how to define


a norm on a quotient space.
Lemma 2.2.13. Let X be a normed space and let X0 be a
closed subspace of X. For any [x] ∈ X/X0 , define k[x]k =
inf kx − yk. This is a norm on X/X0 , called the quotient norm.
y∈X0
(Equivalently,k[x]k = inf kx + yk, since X0 is a subspace.)
y∈X0

Proof. Clearly, k[x]k ≥ 0, and if [x] = 0, then x ∈ X0 , so that

k[x]k = inf kx − yk = inf kzk = 0.


y∈X0 z∈X0

Now, if k[x]k = 0, then inf kx − yk = 0, so that for any n ∈ N,


y∈X0
∃yn ∈ X0 with kx − yn k < n1⇒ yn → x.
Since X0 is closed, this implies x ∈ X0 , so that [x] = 0.
Also, for any number λ,

kλ[x]k = k[λx]k
= inf kλx − yk
y∈X0
= |λ| inf kx − (y/λ)k
y∈X0
= |λ| inf kx − zk
z∈X0
= |λ| k[x]k.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.2. Normed Spaces : Definition and Examples 32

Now it remains to prove the triangle inequality.


Let [x], [y] ∈ X/X0 and  > 0.
By definition of quotient norm, ∃z1 , z2 ∈ X0 such that kx+z1 k ≤
k[x]k +  and ky + z2 k ≤ k[y]k + . Therefore,

k[x] + [y]k = inf kx + y + zk


z∈X0
≤ kx + y + z1 + z2 k
≤ kx + z1 k + ky + z2 k
≤ k[x]k + k[y]k + 2.

Since  > 0 is arbitrary, this proves the triangle inequality. Thus,


the quotient space X/X0 is a normed space with this norm.

A function p on a linear space E is called a seminorm if it


satisfies the properties of the norm except that it may be zero
for non-zero vectors.
More precisely, a seminorm is a function p : E → R+ ∪ {0} such
that

• p(0) = 0,

• p(λx) = |λ|p(x),

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.2. Normed Spaces : Definition and Examples 33

• p(x + y) ≤ p(x) + p(y),

for all x, y ∈ E and all λ ∈ F.


Suppose p is a seminorm on a linear space E. Let E0 be its
kernel, i.e., E0 = {x ∈ E | p(x) = 0}. Let x, y ∈ E0 and λ ∈ F.
Then p(λx + y) ≤ |λ|p(x) + p(y) = 0. ⇒ E0 is a subspace of
E.
Also, note that p(x + y) is independent of y ∈ E0 , i.e., p(x +
y1 ) = p(x + y2 ) for any y1 , y2 ∈ E0 .
To see this,

p(x + y1 ) = p(x + y2 − y2 + y1 )
≤ p(x + y2 ) + p(y1 − y2 )
= p(x + y2 )

Similarly,

p(x + y2 ) = p(x + y2 − y1 + y1 )
≤ p(x + y1 ) + p(y2 − y1 )
= p(x + y1 ).

Thus, p can be regarded as a function on cosets, i.e., p([x]) =

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.2. Normed Spaces : Definition and Examples 34

p(x) and it does not depend on the representative of the coset.


So, a seminorm p on E defines naturally a norm on E/E0 .
It is easy to see that the space of all continuous functions
Cp [a, b] on an interval [a, b] is a normed space with norm kf kp =
Rb
( a |f (x)|p dx)1/p , where 1 ≤ p < ∞.
Rb
Exercise 2.13. Show that for 1 ≤ p < ∞, kf kp = ( a |f (x)|p dx)1/p
defines a norm on the space of all continuous functions on an
interval [a, b].

Now consider the set,


Z b
Rp [a, b] = {f : [a, b] → F | |f (x)|p dx)1/p < ∞}.
a

Rb
For f ∈ Rp [a, b] define kf kp = ( a |f (x)|p dx)1/p . Then this sat-
isfies all the requirements for a norm except, kf kp = 0 ⇒ f = 0.
Rb
In fact, a |f (x)|p dx = 0 only implies f = 0 a.e. on [a, b]. So,
k · kp defines a seminorm on Rp [a, b]. By passing to a quotient
space; quotient with respect to the set of zeroes of k · k, we get
a normed space, denoted by Lp [a, b], but its elements are not
functions, but every element of this space is a class of functions.
Any two representatives, say f1 , f2 , of the same class, will differ
only on a “nonessential set” i.e., on a set of measure zero. We

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.3. Banach Spaces 35

often regard elements of this space as functions, by selecting a


convenient representative of each coset.

2.3 Banach Spaces

A sequence (xn ) in a normed space X is called a Cauchy se-


quence if and only if ∀ε > 0, ∃N ∈ N such that kxn − xm k <
ε, ∀n, m > N. i.e., kxn − xm k → 0 as n > m → ∞.
If (xn ) is a Cauchy sequence in a normed space X, then
(kxn k) is a Cauchy sequence in R and hence lim kxn k exists.
n→∞
(This follows from the inequality | kxk − kyk | ≤ kx − yk and
from the completeness of R. )
Recall that a sequence (xn ) in a normed space X is bounded
if there exists M > 0 such that kxn k ≤ M, ∀n ∈ N. Since every
convergent sequence of real numbers is bounded, it follows that
any Cauchy sequence in X is bounded.

Definition 2.3.1. A normed space X is called complete if and


only if every Cauchy sequence (xn ) in X converges to an element
x of the space X.
A complete normed space is called a Banach space.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.3. Banach Spaces 36

Example 2.3.2. From our studies of real and complex analysis


in degree classes, we recall and note that the spaces R and C
are Banach spaces under the norm ‘modulus’. Similarly the
n
spaces Rn and Cn with euclidean norm kxk = ( |xk |2 )1/2 ,
P
k=1
where x = (x1 , x2 , ..., xn ), are Banach spaces.
The space (Q, | · |) is not complete, for instance, the finite

decimal truncations 1, 1.4, 1.41, 1.414, 1.4142, ... of 2 is a
Cauchy sequence, but not convergent in Q.

Example 2.3.3. Consider the space C[a, b] of all continuous,


complex valued functions equipped with the norm (See Exam-
ple 2.2.4)

kf k∞ = sup |f (t)|, for f ∈ C[a, b].


t∈[a,b]

Let (fn ) be a Cauchy sequence in C[a, b] and ε > 0.


Then ∃N (ε) ∈ N such that kfn − fm k∞ < ε, ∀n > m > N (ε).
⇒ |fn (t) − fm (t)| < ε, ∀n > m > N (ε).
⇒ (fn (t)) is a Cauchy sequence of numbers and hence converges,
for any t ∈ [a, b].
Let f (t) = lim fn (t), ∀t ∈ [a, b].
n→∞

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.3. Banach Spaces 37

Then, letting m → ∞ in kfn − fm k∞ < ε, we get

kfn − f k∞ < ε, ∀n > N (ε).

Note that convergence with respect to this norm is the uniform


convergence. Since uniform limit of a sequence of continuous
functions is continuous, it follows that f ∈ C[a, b]. This shows
that (C[a, b], k · k∞ ) is a complete normed space.

Example 2.3.4. For 1 ≤ p < ∞, consider the linear space `p of


all p- summable sequences of real (or complex) numbers, (i.e.,

|xi |p < ∞,)
P
all sequences x = (x1 , x2 , ..., xn , ....), for which
i=1

p 1/p
. i.e., `p = (R∞ , k ·
P
equipped with the norm kxkp = ( |xi | )
i=1
kp ). (See Example 2.2.11)
Let xn = (xnm )∞
m=1 be a Cauchy sequence in `p and ε > 0.
(Here, each x is a sequence, xn = (xn1 , xn2 , xn3 , ...). )
n

Then ∃N (ε) ∈ N such that kxn − xk kp < ε, ∀n > k > N (ε).


Thus, ∀n > k > N (ε)

X
|xnm − xkm |p < εp (2.1)
m=1

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.3. Banach Spaces 38

⇒ |xnm − xkm | < ε, ∀n > k > N (ε) for each m.


This shows that for each m, (xnm ) is a Cauchy sequence in
R(or C) and hence converges.
Let xm = lim (xnm ), for m = 1, 2, ... and x = (xm )∞ m=1 .
n→∞
We claim that x ∈ `p and xn → x in `p as n → ∞.
M M M
|xm |p = ( lim |xnm |p ) = lim |xm |p ≤
P P P
For each M ∈ N,
m=1 m=1 n→∞ n→∞ m=1
sup kxn kpp < C < ∞, for some C > 0, since Cauchy sequences
n
are bounded. Since this is true for any M ∈ N, letting M → ∞,

|xm |p < ∞ ⇒ x ∈ `p .
P
we get
m=1

From equation 2.1, for any M ∈ N, we have, for all suffi-


ciently large n, k,

M
X
|xnm − xkm |p < εp
m=1

Letting k → ∞ first and then M → ∞, we get



X
|xnm − xkm |p ≤ εp
m=1

This means, kxn − xkp ≤ ε for large values of n.


⇒ xn → x in `p .

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.3. Banach Spaces 39

Thus, (`p , k · kp ) is a Banach space.

Exercise 2.14. Show that `np = (Rn , k · kp ), 1 ≤ p < ∞ is a


Banach space.

Exercise 2.15. Show that every finite dimensional normed lin-


ear space is a Banach space.

Exercise 2.16. Show that the space (`∞ , k · k∞ ) (See Exam-


ple 2.2.3) is a Banach space.

Suppose that E be a Banach space and E1 be a subspace


of E. Then E1 is a Banach space if and only if E1 is
closed.
To see this, first assume that E1 is a Banach space with the
inherited norm. If (xn ) is a sequence in E1 that converges to
some x ∈ E, then (xn ) is a Cauchy sequence in E1 . Since E1 is
complete, x ∈ E1 . This implies, E1 is closed.
Conversely assume that E1 is closed. If (xn ) is a Cauchy
sequence in E1 ⊆ E, it converges to some x ∈ E, since E is
complete. But E1 is closed, so that x ∈ E1 . Thus, E1 is a
Banach space.
From Exercise 2.16, it follows that the subspaces c0 and c
are Banach spaces, as they are closed in `∞ .

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.3. Banach Spaces 40

To see c0 is closed in `∞ , let (x(m) ) be a sequence in c0 , ε > 0


and suppose kx(m) − xk∞ ≤ ε/2 for all m ≥ N for some N ∈ N.
(N ) (N )
Since x(N ) ∈ c0 , lim xk = 0, so that |xk | < ε/2 for all k > N1
k→∞
for some N1 ∈ N. Hence for all k > N1 ,

(N ) (N ) (N )
|xk | ≤ |xk − xk | + |xk | ≤ kx − x(N ) k∞ + |xk | < ε.

This implies that lim xk = 0. ⇒ x = (xk ) ∈ c0 .


k→∞

Exercise 2.17. Show that c is a closed subspace of `∞ .

Exercise 2.18. Show that the space of all polynomial functions


(P [0, 1], k · k∞ ) is not a Banach space.

Exercise 2.19. Let X be a Banach space and E be a closed


subspace of X. Show that the quotient space X/E is a Banach
space.

Example 2.3.5. Consider the subspace s∗ of `∞ consisting of


sequences with all but finite zero terms, as a subspace of `∞ .(See
Example 2.1.1). The sequence (a(n) ) defined by,
a(n) = (1, 1/2, 1/3, ...., 1/n, 0, 0, ....) for n ∈ N,
is a sequence in s∗ and (a(n) ) converges to the sequence (1/n),
but it does not belongs to s∗ . Thus, s∗ is not closed, and hence
not a Banach space.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.3. Banach Spaces 41

Definition 2.3.6. Let E, F be normed spaces. We say that


a linear map T : E → F is an isometry into if and only if
kT xkF = kxkE , for all x ∈ E.

Next, we show that any normed space can be densely em-


bedded in a Banach space.

Theorem 2.3.7. Let E be a normed space. There exists a com-


plete normed space Ê and a linear map T : E → Ê such that
kT xk = kxk for all x ∈ E, (isometry into); and ImT (:= T E)
is a dense set in Ê (i.e., T E = Ê.)
We call such a space Ê the completion of E.

Proof. Let E be the space of all Cauchy sequences x = (xi )∞i=1


in E. Define p(x) = lim kxi k, where x ∈ E. Note that p is well-
i→∞
defined (Since, if (xn ) is a Cauchy sequence in a normed space
X, then (kxn k) is a Cauchy sequence in R and hence lim kxn k
n→∞
exists.) and is a seminorm on E.
Let N = {x | p(x) = 0}. Then, N is the subspace of all
sequences in E that converges to 0. Also, p defines a norm on the
quotient space Ê = E/N, by the same formula p(x) = lim kxi k,
i→∞
for any representative x of an equivalence class [x] = x + N ∈
E/N.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.3. Banach Spaces 42

Define T : E → Ê by T (x) = [x], where x is the constant


sequence (x, x, x, ....). Note that a constant sequence is a Cauchy
sequence and p(x) = kxk. Clearly, T is a linear map and kT xk =
kxk.
Now, we show that T E is dense in Ê. Let x = (xn ) ∈ E.
For any ε > 0, ∃N ∈ N such that kxn − xm k ≤ ε for all m >
n ≥ N. For such an n > N, define xn as the constant sequence
(xn , xn , xn , ...). Then xn ∈ E, and T xn = [xn ]. ⇒ [xn ] ∈ T E.
Also, d([x], [xn ]) = p([x] − [xn ]) = p(x − xn ) = lim kxm − xn k ≤
m→∞
ε. Since ε > 0 is arbitrary, this shows that any [x] ∈ Ê can be
approximated by elements of T E. This proves that T E is dense
in Ê.
Next, we prove that Ê is complete. Consider a Cauchy se-
quence ([x(n) ]) in Ê.
Then p([x(n) ] − [x(m) ]) = p(x(n) − x(m) ) → 0 as n ≥ m → ∞.
Choose εn > 0 and converging to 0, then since T E is dense in
Ê, we get xn ∈ E such that p(x(n) − T xn ) < εn . We claim that

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.3. Banach Spaces 43

x0 = (xn ) is a Cauchy sequence in E. We have,

kxn − xm k = p(T xn − T xm )
≤ p(T xn − x(n) ) + p(x(n) − x(m) ) + p(x(m) − T xm )
→ 0 as n ≥ m → ∞.

Therefore, x0 = (xn ) ∈ E.
Also,

p([x(0) ] − [x(n) ]) = p(x(n) − x(0) )


≤ p(x(n) − T xn ) + p(T xn − x(0) )
→ 0 as n → ∞.

Thus, [x(0) ] = lim [x(n) ]. Hence, any Cauchy sequence ([x(n) ]) in


n→∞
Ê converges in Ê. This completes the proof.

The completion of a normed space is unique in the following


sense.

Theorem 2.3.8. Let E be a normed space, and Xi , i = 1, 2 be


complete normed spaces. Let Ti : E → Xi be isometries into
dense subspaces of Xi for i = 1, 2. Then the natural isometry
T2 ◦ T1−1 : T1 E → T2 E can be extended to an isometry between

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.3. Banach Spaces 44

X1 and X2 .

Proof. Let y ∈ X1 . Since T1 E is dense in X1 , there exists a


sequence (xn ) in E such that T1 xn → y. Now consider the se-
quence (T2 xn ) in X2 . Since Ti ’s are isometries, we have

kT2 xn − T2 xm k = kT2 (xn − xm )k = kxn − xm k = kT1 xn − T1 xm k,

by linearity of Ti .
This shows that (T2 xn ) is a Cauchy sequence in X2 , and hence
converges, say z = lim T2 xn .
n→∞
Define T : X1 → X2 by T y = z.
Note that T is well-defined, since if (un ) is another sequence in
E such that T1 un converges to y, then T2 un also converges to z.
This follows from,

kT2 xn − T2 un k = kT2 (xn − un )k


= kxn − un k
= kT1 (xn − un )k
= kT1 xn − T1 un k → 0 as n → ∞.

Clearly, T is linear. We claim that T is a an isometry from X1

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.3. Banach Spaces 45

onto X2 . Since norm is continuous, we have

kyk = lim kT1 xn k = lim kxn k = lim kT2 xn k = kT yk.


n→∞ n→∞ n→∞

This implies that T is an isometry.


Now it remains to prove ImT = X2 . Let z ∈ X2 . Since T2 E
is dense in X2 , we can choose a sequence xn ∈ E such that
lim T2 xn = z.
n→∞
Then as above, (T1 xn ) is a Cauchy sequence in X1 and let y =
lim T1 xn . Then, T y = z. This completes the proof.
n→∞

Example 2.3.9. Consider the normed space of all continuous


functions Cp [a, b] on an interval [a, b] equipped with the norm
Rb
kf kp = ( a |f (x)|p dx)1/p , where 1 ≤ p < ∞. It can be shown
that Cp [a, b] is dense in the Banach space Lp [a, b] and it is the
completion of Cp [a, b].

Exercise 2.20. Show that Cp [0, 1] is not a Banach space with


Lp -norm, 1 ≤ p < ∞.(Hint: On [0, 1], define a sequence of

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

2.3. Banach Spaces 46

functions un for n ≥ 3, by


 0 0 ≤ t < 12 − n1
n 1
un (t) = nt − 2
+1 2
− n1 ≤ t ≤ 12

1
1 <t≤1

2

Show that (un ) is a Cauchy sequence in Cp [0, 1] and kun −ukp →


0, where
(
0 0 ≤ t < 12
u(t) = 1
1 2
≤t≤1

but u ∈
/ Cp [0, 1].)

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

Chapter 3
Hilbert Spaces

In this module, we introduce a very important mathematical


abstract structure, called Hilbert spaces, which were devel-
oped by the eminent German mathematician David Hilbert
(1862-1943). A Hilbert space is the abstraction of the finite di-
mensional Euclidean spaces of Geometry. The greatest advan-
tage of a Hilbert space is its underlying concept of orthogonality.
In order to explore the properties of Hilbert spaces, first we in-
troduce the notion of inner product in linear spaces, which is a

47
MTH3C13 : FUNCTIONAL ANALYSIS

3.1. Basic Notions 48

natural generalization of the concept of dot product in R2 .

3.1 Basic Notions

We begin with the precise definition of inner product.

Definition 3.1.1. Let X be a linear space over C. An inner


product ( or a scalar product) on X is a complex-valued func-
tion from X × X → C such that for all x, y, z ∈ X and a, b ∈ C,
we have

• Linearity with respect to the first argument :

hax + by, zi = ahx, zi + bhy, zi;

• Conjugate-Symmetry:

hx, yi = hy, xi;

• Non-negativeness:

hx, xi ≥ 0, and hx, xi = 0 if and only if x = 0.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.1. Basic Notions 49

A linear space with an inner product is called an inner


product space.

Remark 3.1.2. Conjugate-symmetry of the inner product shows


that it is conjugate-linear (or semi-linear ) with respect to the
second argument; i.e., hx, ay + bzi = ahx, yi + bhx, zi, for all
x, y, z ∈ X and a, b ∈ C.
To see this, note that

hx, ay + bzi = hay + bz, xi


= ahy, xi + bhz, xi
= a hy, xi + b hz, xi
= a hx, yi + b hx, zi.

Theorem 3.1.3. Cauchy-Schwartz Inequality


Let X be a linear space with an inner product h·, ·i. Then for
any x, y ∈ X,

|hx, yi| ≤ hx, xi1/2 · hy, yi1/2 .

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.1. Basic Notions 50

Proof. Let p(x) = hx, xi1/2 . Then for any number λ, we have:

hx − λy, x − λyi = hx, x − λyi − λhy, x − λyi


= hx, xi − λhx, yi − λhy, xi + λλhy, yi
= p(x)2 − λhy, xi − λhy, xi + |λ|2 p(y)2
= p(x)2 − 2 Re(λhy, xi) + |λ|2 p(y)2

so that
p(x)2 − 2 Re(λhy, xi) + |λ|2 p(y)2 ≥ 0.

If hx, yi = 0, the inequality is trivial. So, assume that hx, yi =


6 0.
2
Substitute λ = p(x) /hy, xi in the above inequality, we get

0 ≤ −p(x)2 + p(x)4 p(y)2 /|hy, xi|2 ⇒ |hx, yi|2 ≤ p(x)2 · p(y)2 ,

which implies the desired inequality.

Remark 3.1.4. Equality holds in Cauchy-Schwartz inequality


if and only if the set {x, y} is linearly dependent. i.e., x = λy
for some number λ.

Now we note that an inner product on a linear space X


induces a norm on X in a natural way, namely p(x) = kxk =
hx, xi1/2 for x ∈ X. Indeed, p(λx) = |λ| p(x), and to prove the

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.1. Basic Notions 51

triangle inequality, note that

Rehx, yi ≤ |hx, yi| ≤ p(x)p(y),

by the Cauchy-Schwartz inequality. Therefore,

p(x + y)2 = hx + y, x + yi
= p(x)2 − 2 Rehx, yi + p(y)2
≤ (p(x) + p(y))2

so that p(x + y) ≤ p(x) + p(y).


Thus every inner product space is a normed space and this
norm is called “hilbertian norm” which will be denoted hereafter
by kxk instead of p(x).
If an inner product space X is complete with respect to this
norm, then X is said to be a Hilbert space. We often use H
to denote a Hilbert space.

Example 3.1.5. For x = (a1 , a2 , ..., an ) and y = (b1 , b2 , ..., bn )


n
in Cn , define hx, yi =
P
ai bi . This defines an inner product
i=1
on Cn , called the standard inner product on Cn . The norm on
Cn induced by this inner product is just the Euclidean norm;

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.1. Basic Notions 52

n
kxk = hx, xi1/2 = ( |ai |2 )1/2 . Under this norm, Cn is complete.
P
i=1
(See Example 2.3.2). Thus, Cn with the standard inner product
is a Hilbert space.

Example 3.1.6. For elements x = (ai )∞ ∞


i=1 and y = (bi )i=1 in `2 ,

P
define hx, yi = ai bi . This is well defined, since
i=1

v v
∞ u∞ u∞
X uX uX
| ai bi | ≤ t |ai |2 · t |bi |2 < ∞
i=1 i=1 i=1

using Hölder’s inequality (See Theorem 2.2.5.) One can easily


verify that this defines an inner product on `2 . The norm induced

by this inner product is given by kxk = hx, xi1/2 = ( | ai |2 )1/2 ,
P
i=1
which is nothing but k · k2 on `2 . From Example 2.3.4, we have
seen that `2 with this norm is complete, hence `2 with this inner
product is a Hilbert space.

Example 3.1.7. Consider the space of all continuous functions


C[a, b] on an interval [a, b]. For f, g ∈ C[a, b], define
Z b
hf, gi = f (t)g(t)dt.
a

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.1. Basic Notions 53

From Hölder’s inequality for functions (See Theorem 2.2.7), we


see that the above integral converges absolutely and it can be
verified that this defines an inner product on C[a, b].
The norm induced by this inner product is
Z b
1/2
kf k = hf, f i =( |f (x)|2 dx)1/2 = kf k2 ,
a

and from Example 2.3.9, C2 [a, b] = (C[a, b], k·k2 ) is not closed in
the Banach space L2 [a, b]. Hence C2 [a, b] with the above defined
inner product is not a Hilbert space.

Using Cauchy-Schwartz inequality (Theorem 3.1.3), we now


show that the inner product is continuous in both the variables;
i.e., hxn , yn i → hx, yi when xn → x and yn → y. Indeed, we
have

|hxn , yn i − hx, yi| = |hxn , yn i − hxn , yi + hxn , yi − hx, yi|


≤ |hxn , yn − yi| + |hxn − x, yi|
≤ kxn kkyn − yk + kxn − xkkyk
→ 0, since(kxn k) is a bounded sequence.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.1. Basic Notions 54

Also, for any x, y in an inner product space X, we have

kx + yk2 + kx − yk2 = 2(kxk2 + kyk2 ).

This is known as the parallelogram law, and to see this:

kx + yk2 + kx − yk2 = hx + y, x + yi + hx − y, x − yi
= 2(hx, xi + hy, yi)
= 2(kxk2 + kyk2 ).

Using inner product, we introduce the notion of orthogonality


in linear spaces.
We say that two vectors x and y are orthogonal to each
other, denoted by x⊥y, if hx, yi = 0.
If x⊥y, then kx + yk2 = kxk2 + kyk2 . This is known as the
Pythagorean theorem. Indeed, we have

kx + yk2 = hx + y, x + yi
= hx, xi + hx, yi + hy, xi + hy, yi
= kxk2 + kyk2 .

Exercise 3.1. Show that for any x, y in an inner product space

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.1. Basic Notions 55

X,

1
hx, yi = (kx + yk2 − kx − yk2 + ikx + iyk2 − kx − iyk2 ).
4

Remark 3.1.8. We noted that every inner product space is a


normed space. It can be shown that a normed space is an inner
product space if and only if the norm satisfies the parallelogram
law. In other words, a norm k · k on a linear space X is induced
by an inner product h·, ·i on it if and only if the norm satisfies the
parallelogram law. If it is so, then h·, ·i is given by the expression
in the above exercise 3.1. This result is due to Jordan and
von Neumann (1935). In view of this, we see that a Hilbert
space is a complete normed space in which the norm satisfies
the parallelogram law. In particular, every Hilbert space is a
Banach space, and a Banach space is a Hilbert space only if the
norm satisfies the parallelogram law. Using this idea, one can
infer that among the sequence spaces (`p , k · kp ); 1 ≤ p ≤ ∞,
(`2 , k · k2 ) is the only Hilbert space.

Example 3.1.9. Consider the space C2 [a, b]. From Example 3.1.7,
we note that L2 [a, b] is the completion of the normed space
C2 [a, b]. Any element x of L2 [a, b] can be regarded as a function
Rb
x(t) such that the Lebesgue integral a |x(t)|2 dt < ∞. Using

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.2. Orthonormal Systems 56

continuity of inner product, we can extend the inner product on


C2 [a, b] to the space L2 [a, b] and indeed, for any x, y ∈ L2 [a, b],
Rb
we have hx, yi = a x(t)y(t)dt. The space L2 [a, b] with this inner
product is a very important Hilbert space.

Exercise 3.2. If {ei }n1 are pairwise orthogonal and normalized


in a Hilbert space H (i.e., kei k = 1 for 1 ≤ i ≤ n), then for any
numbers αi , show that k n1 αi ei k = ( n1 |αi |2 )1/2 . (Hint: Use
P P

Pythagorean theorem.)

3.2 Orthonormal Systems

Recall that two elements x and y in an inner product space are


orthogonal (to each other) if hx, yi = 0. A set of vectors {ei }i≥1
in an inner product space is said to be an orthonormal system
if the vectors {ei } are pair-wise orthogonal and kei k = 1, for all i.
More precisely, we have hei , ej i = δij where δij is the Kronecker’s
delta function,
(
1 if i = j
δij =
0 6 j
if i =

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.2. Orthonormal Systems 57

The set {e1 , e2 , ..., en },

ei = (0, 0, ..., 0, 1, 0, ...0),

where 1 occurs only at ith place, is an orthonormal system in


the Hilbert space Cn with the standard inner product hx, yi =
n
ai bi , for x = (a1 , a2 , ..., an ) and y = (b1 , b2 , ..., bn ) in Cn . One
P
i=1
can easily check that hei , ej i = δij .

Exercise 3.3. Show that the set of vectors { √12π eint }n=∞
n=−∞ is
an orthonormal system in the Hilbert space L2 [−π, π] with the

inner product hf, gi = −π f (t)g(t)dt.

Exercise 3.4. Show that the set of vectors { √12π , cos√ nx , sin
√ nx
π π
for n = 1, 2, ...} is an orthonormal system in L2 [−π, π].

We now state and prove Bessel’s inequality.

Theorem 3.2.1. For any orthonormal system {ei }i≥1 ⊂ H, and


for every x ∈ H, we have
X
|hx, ei i|2 ≤ kxk2 .
i≥1

Proof. First we assume that the given orthonormal system {ei }i≥1

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.2. Orthonormal Systems 58

n
P
is finite. Consider yn = hx, ei iei . Then, by Cauchy-Schwartz
i=1
inequality,

|hyn , xi| ≤ kxk · kyn k

Also,
n
X
hyn , xi = h hx, ei iei , xi
i=1
n
X
= hx, ei ihei , xi
i=1
n
X
= hx, ei ihx, ei i
i=1
n
X
= |hx, ei i|2
i=1

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.2. Orthonormal Systems 59

and,

kyn k2 = hyn , yn i
Xn n
X
= h hx, ei iei , hx, ej iej i
i=1 j=1
n
X Xn
= hx, ei ihei , hx, ej iej i
i=1 j=1
Xn n
X
= hx, ei i hx, ej ihei , ej i
i=1 j=1
Xn
= hx, ei ihx, ei i
i=1
Xn
= |hx, ei i|2
i=1

Hence,
n
X
|hx, ei i|2 = |hyn , xi|
i=1
≤ kxk · kyn k
Xn
= kxk · ( |hx, ei i|2 )1/2
i=1

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.2. Orthonormal Systems 60

which proves the desired inequality for a finite set {ei }n1 .
If the given orthonormal system {ei }i≥1 is infinite, then let
n ∞
|hx, ei i|2 ≤ kxk2 , we get |hx, ei i|2 ≤ kxk2 .
P P
n → ∞ in
i=1 i=1

Corollary 3.2.2. For any x ∈ H and any orthonormal system



{ei }∞
P
1 , there exists y ∈ H such that y = hx, ei iei .
i=1

Proof. From the infinite case of the equality in Exercise 3.2, and
using Bessel’s inequality (Theorem 3.2.1), we note that yn =
Pn
hx, ei iei is a Cauchy sequence in the Hilbert space H. Due
i=1
n
P
to the completeness of H, the series yn = hx, ei iei converges
i=1
to some element y ∈ H. Thus, there exists y ∈ H such that
P∞
y = hx, ei iei .
i=1

A system of vectors {xi }i≥1 in a normed space X is called a


complete system if the linear span
n
X
{ αi xi : for all scalars αi , for all n ∈ N}
i=1

of the vectors {xi }i≥1 is a dense set in X.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.2. Orthonormal Systems 61

Example 3.2.3. From the Weierstrass approximation theorem


in real analysis, the set of polynomials is dense in the space
C[0, 1] so that the system {tn }n≥0 is complete in the space
C[0, 1]. Since C[0, 1] is dense in L2 [0, 1], it follows that the sys-
tem {tn }n≥0 is complete in the space L2 [0, 1] too.
Also, from the Weierstrass theorem, the set of trigonometric
polynomials is dense in the space C[−π, π]. Hence it follows that
the sets {1, cos nx, sin nx : n = 1, 2, 3, ...} and {eint }∞
n=−∞ are
complete systems in L2 [−π, π].

Lemma 3.2.4. If {fi } is a complete system in a Hilbert space


H and x ⊥ fi for all i, then x = 0.

Proof. Since {fi } is a complete system in H, linear span of {fi }


is dense in H and since x ⊥ fi for all i, we have x ⊥ span{fi }
for all x ∈ H. Hence for any x ∈ H, there exists a sequence
xn converging to x such that xn ⊥ x for all n ∈ N, so that
kxk2 = hx, xi = limhxn , xi = 0, implies x = 0.

Exercise 3.5. If {fi } is a system of vectors in a Hilbert space


H, with the property that x ⊥ fi for all i, implies x = 0, then
show that {fi } is a complete system.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.3. Orthogonal Bases 62

3.3 Orthogonal Bases

Given a linearly independent set of vectors {xi } in an inner


product space, there is an inductive procedure, called Gram -
Schmidt orthonormalization process, to construct a set of
orthonormal vectors {ei } such that span{xi }n1 = span{ei }n1 for
every n = 1, 2, 3, ... We set e1 = x1 /kx1 k and define inductively
for n = 2, 3, ...,
en = (xn − yn )/kxn − yn k where yn = n−1
P
i=1 hxn , ei iei .
This definition makes sense since we always have kxn − yn k = 6 0,
due to the linear independence of {xi }. (Verify it!)
This sequence {ei }∞ 1 is an orthonormal system, i.e., hei , ej i = δij ,
such that span{xi }n1 = span{ei }n1 for every n = 1, 2, 3, ....
We know that the set of rational numbers is a countable
dense subset of the set of real numbers. We now introduce a
special class of normed spaces with such a property as follows.

Definition 3.3.1. A normed space X is called a separable


space if there exists a dense countable set in X.

We now present a characterization of separable Hilbert spaces.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.3. Orthogonal Bases 63

Theorem 3.3.2. A Hilbert space H is separable if and only if


there exists a complete orthonormal system {ei }i≥1 in H.

Proof. First, assume that H is separable. Then by definition,


there exists a countable subset Y = {yi }i≥1 , which is dense in
H. We now inductively construct a linearly independent subset
{xi }i≥1 of Y such that for every n, span{xi }n1 = span{yi }N 1
n
for
some Nn ∈ N. For this, we choose x1 as the first non-zero element
of Y = {yi }i≥1 , and then inductively, if {xi }n1 are chosen, we
take xn+1 as the first yi , i > Nn , which is linearly independent
of {xi }n1 . Clearly, span{xi }i≥1 is also dense in H. If H is of
finite dimension, say dimH = n, then this procedure will stop
after a finite number of steps, producing a basis {xi }n1 . If dimH
is infinite, this process will continue indefinitely. In any case,
by applying Gram-Schmidt orthogonalization procedure to the
system {xi }i≥1 , we find a complete orthonormal system {ei }i≥1
in H. This completes one direction of the proof.
For the other direction, if {ei }i≥1 is complete in H, then the
P
set of all finite sums αi ei with rational coefficients αi forms
a countable dense subset of H, and hence H is separable.

Recall the definition of basis for a vector space. We now

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.3. Orthogonal Bases 64

define the concept of a basis for a normed space.

Definition 3.3.3. A sequence {xi }∞ i=1 is called a basis of a


normed space X if for every x ∈ X, there exists a unique
P
series i≥1 ai xi that converges to x.

Note the difference; our usual notion of basis requires that


any element of the space can be expressed as a unique finite
linear combination of basis vectors; but here we allow infinite
sums too.
Next, we prove that any complete orthonormal system in a
Hilbert space is a basis for it.

Theorem 3.3.4. A complete orthonormal system {ei }∞


i=1 in a
Hilbert space H is a basis in H.

Proof. Let {ei }∞ i=1 be a complete orthonormal system in a Hilbert


space H and x ∈ H. By the Bessel’s inequality (Theorem 3.2.1),
P∞
|hx, ei i|2 < ∞. Hence by Corollary 3.2.2, the element y =
Pi=1

i=1 hx, ei iei exists in H.
Also, it is easy to check that hy − x, ei i = 0 for all i, implies
that y − x ⊥ ei for every i. Therefore, by Lemma 3.2.4, we get
y − x = 0 or y = x. Thus, x = ∞
P
i=1 hx, ei iei .

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.3. Orthogonal Bases 65

Now let x = ∞
P
ai ei , then hx, ei i = ai , showing that the
P∞i=1
expression x = i=1 hx, ei iei is unique.

Combining the results in Theorem 3.3.2 and Theorem 3.3.4


we get the following corollary.

Corollary 3.3.5. Every separable Hilbert space has an orthonor-


mal basis.

One can easily check that the system of vectors described


in exercises 3.3 and 3.4 are orthonormal bases for the Hilbert
space L2 [−π, π].

Theorem 3.3.6. Let {ei }i≥1 be an orthonormal system in a


Hilbert space H. Then {ei }i≥1 is a basis in H if and only if for
all x ∈ H,
X
kxk2 = |hx, ei i|2 .
i≥1

This identity is known as Parseval’s identity

Proof. Suppose the orthonormal system {ei }i≥1 is a basis in H.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.3. Orthogonal Bases 66

P
Then for any x ∈ H, we have x = i≥1 hx, ei iei , so that

kxk2 = hx, xi
X X
= h hx, ei iei , hx, ej iej i
i≥1 j≥1
X
2
= |hx, ei i| .
i≥1

To prove the converse, note that


kx − n1 hx, ei iei k2 = kxk2 − n1 |hx, ei ik2 → 0 as n → ∞, since
P P

by assumption, we have kxk2 = i≥1 |hx, ei i|2 .


P
P
Therefore, x = i≥1 hx, ei iei for any x ∈ H, showing that
{ei }i≥1 is a basis in H.

Remark 3.3.7. If the Parseval’s identity holds for a particular


P
x in H, then we have x = i≥1 hx, ei iei . Thus, if the Parseval’s
identity holds for every member of a dense subset of H, we see
that {ei }i≥1 is a complete system and hence by Theorem 3.3.4,
it is a basis in H.

Next we show that any two separable infinite dimensional


Hilbert spaces are the same in the following sense.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.3. Orthogonal Bases 67

Theorem 3.3.8. Any two separable infinite dimensional Hilbert


spaces H1 and H2 are isometrically equivalent; i.e., there exists
a linear isomorphism onto T : H1 → H2 such that kT xk = kxk
for every x ∈ H1 and, moreover, hT x, T yiH2 = hx, yiH1 for every
x and y in H1 .

Proof. Let H be a separable infinite dimensional Hilbert space.


To prove the theorem, it is enough to show that H is isometri-
cally equivalent to `2 , the space of all square summable scalar
sequences, which is also a separable infinite dimensional Hilbert
space.
Let {fi }∞
1 be an orthonormal basis for H. Then for every
x ∈ H, x = i≥1 hx, fi ifi and kxk2 = i≥1 |hx, fi i|2 .
P P

Let {ei }∞
1 be the natural basis of `2 . Define T : H → `2
P∞
by T x = 1 hx, fi iei = (hx, f1 i, hx, f2 i, hx, f3 i, ...) ∈ `2 , using

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.3. Orthogonal Bases 68

Bessel’s inequality. Then T is clearly linear and

kT xk2 = hT x, T xi
X∞ ∞
X
= h hx, fi iei , hx, fj iej i
1 1

X
= |hx, fi i|2
1
= kxk2 ,

showing that kT xk = kxk for every x ∈ H.


Now let (a1 , a2 , a3 , ...) ∈ `2 . Then ∞
P
1 ai fi converges in H
(Verify that its partial sums forms a Cauchy sequence in H and
hence converges.) and let x = ∞
P
ai fi ∈ H. Then hx, fi i = ai
P∞ 1
for all i and hence T x = 1 hx, fi iei = (a1 , a2 , a3 , ...). This
shows that T is onto. Also,

X ∞
X
hT x, T yi`2 = h hx, fi iei , hy, fj iej i
1 1

X
= hx, fi ihy, fi i
1
= hx, yiH

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.4. Orthogonal Decomposition 69

This completes the proof.

Exercise 3.6. Let {ϕi (t)}∞


1 be a sequence of continuous func-
tions that forms an orthnormal basis for L2 [a, b]. Then show
that the system

{ϕi (t)ϕj (τ ) = ψij (t, τ )}∞


i,j=1

is an orthonormal basis of L2 ([a, b] × [a, b]), which is the com-


pletion of the space of square integrableq(continuous) functions
RbRb
of two variables with norm kf (t, τ )k = a a
|f (t, τ )|2 dtdτ.

3.4 Orthogonal Decomposition

We begin with the definition of projections. Let L be a closed


subspace of a Hilbert space H. We define the distance of any
x ∈ H to L as ρ(x, L) = inf y∈L kx − yk. If this infimum is
achieved at some y ∈ L, we define y = PL x, and call this y as
the projection of x onto L.
We now consider a particular case that the subspace L con-
sidered is separable. Let {ei }i≥1 be an orthonormal basis in L
P
and let x ∈ H. Set y = i≥1 hx, ei iei . Then y ∈ L and clearly

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.4. Orthogonal Decomposition 70

x − y ⊥ ei for all i, which implies x − y ⊥ L.


Therefore, for any z ∈ L, we have x − z = (x − y) + (y − z)
and kx − zk2 = kx − yk2 + ky − zk2 , since x − y ⊥ y − z. It
follows
inf{kx − zk | z ∈ L} = kx − yk

and y = PL x. Also, such a y is unique.

Exercise 3.7. Show that a closed subspace of a separable Hilbert


space is separable.

Before proving the existence of a unique projection in the


non-separable case, we prove the following more general fact.

Theorem 3.4.1. Let M be a convex closed set in a Hilbert space


H. Then there exists a unique y ∈ M such that

ρ(x, M ) = kx − yk,

where ρ(x, M ) = inf w∈M kx − wk, the distance of x to the set


M.

Proof. Let d = ρ(x, M ) = inf w∈M kx − wk. As d is an infimum,


we can choose a sequence yn ∈ M and kx − yn k → d. We claim

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.4. Orthogonal Decomposition 71

that this sequence {yn } is a Cauchy sequence. From parallelo-


gram law, we find that

2(kx − yn k2 + kx − ym k2 ) = k2x − (yn + ym )k2 + kyn − ym k2 .

By the convexity of M, (yn + ym )/2 ∈ M and hence

kx − (yn + ym )/2k ≥ d,

we see that the right side of the above equality tends to 4d2 as
n, m → ∞.
But the right side is not smaller than 4d2 + limkyn − ym k2 .
From this, we conclude that kyn − ym k → 0 as n, m → ∞ and
hence {yn } is a Cauchy sequence in M so that there exists y ∈ H
such that y = limn→∞ yn . As M is a closed subspace of H, we
see that y ∈ M and ρ(x, M ) = kx − yk.
Suppose ρ(x, M ) = kx − y1 k = kx − y2 k for y1 , y2 ∈ M, then
again from parallelogram law and the convexity of M, it follows
that y1 = y2 . This proves the uniqueness of y.

Instead of closed convex set M, if we take a closed subspace


L in the above theorem, we see that there is a unique projection
y = PL x. Also, note that ρ(x, L) = kx − yk for some y ∈ L if

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.4. Orthogonal Decomposition 72

and only if x − y ⊥ L. To see this, first let y ∈ L be such that


x − y ⊥ L. Then, for any z ∈ L

kx − zk2 = kx − yk2 + ky − zk2 ≥ kx − yk2 ,

shows that y gives the distance and it is unique.


To see the opposite direction, assume that y ∈ L is the
projection PL x. Then for any z ∈ L,

kx−yk2 ≤ kx−(y +λz)k2 = kx−yk2 −2Reλhz, x−yi+|λ|2 kzk2 .

This implies,
2Reλhz, x − yi ≤ |λ|2 kzk2 .

Take λ = thz, x − yi for t ∈ R. Then

2t|hz, x − yi|2 ≤ t2 |hz, x − yi|2 kzk2

for all t ∈ R. Letting t → 0, we get hz, x − yi = 0 and hence


every z ∈ L is orthogonal to x − y. Thus, we have proved the
following theorem.

Theorem 3.4.2. For all x ∈ H, there exists a unique y ∈ L


such that x − y ⊥ L and y = PL x. Then, x = (x − y) + y and

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.4. Orthogonal Decomposition 73

kxk2 = kx − yk2 + kyk2 .

For any subspace L of a Hilbert space H, we define the or-


thogonal complement, denoted by L⊥ of L, as :

L⊥ = {x ∈ H : x ⊥ L} = {x ∈ H : x ⊥ y, for every y ∈ L}.

Exercise 3.8. For any subspace L of H, prove that L⊥ is a


closed subspace of H. Also, prove that L ∩ L⊥ = {0}.

The following theorem is known as the Projection Theorem.

Theorem 3.4.3. For every closed subspace L of a Hilbert space


H,
L ⊕ L⊥ = H.

Moreover,(L⊥ )⊥ = L.

Proof. Let x ∈ H. For any closed subspace L ,→ H, by Theorem


3.4.2, there exists a unique y ∈ L such that x = (x − y) + y with
x − y ∈ L⊥ . ( i.e., y = PL x.).
Also, from Exercise 3.8, L ∩ L⊥ = {0}. If x = z1 + y1 with
z1 ∈ L⊥ and y1 ∈ L, then z1 + y1 = (x − y) + y, so that
z1 − (x − y) = y − y1 .

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.4. Orthogonal Decomposition 74

But L and L⊥ are subspaces and L ∩ L⊥ = {0}, so that


z1 = (x − y) and y1 = y. This shows that L ⊕ L⊥ = H.
Now we show that (L⊥ )⊥ = L.
Let y ∈ L. If x ∈ L⊥ , then hy, xi = 0 ⇒ y ∈ (L⊥ )⊥ . This
implies that L ⊂ (L⊥ )⊥ .
Next, let y ∈ (L⊥ )⊥ . Since L ⊕ L⊥ = H, we have y = z1 +
z2 with z1 ∈ L and z2 ∈ L⊥ . Since L ⊂ (L⊥ )⊥ , z1 ∈ (L⊥ )⊥
and since (L⊥ )⊥ is a subspace (See Exercise 3.8), we see that
y − z1 ∈ (L⊥ )⊥ =⇒ z2 ∈ (L⊥ )⊥ But z2 ∈ L⊥ and hence
z2 ∈ L⊥ (L⊥ )⊥ = {0}, implies z2 = 0.
T

Hence y = z1 ∈ L ⇒ (L⊥ )⊥ ⊂ L.
Thus, we have (L⊥ )⊥ = L.

Exercise 3.9. Let L1 and L2 be closed subspaces of a Hilbert


space H such that L1 ,→ L2 ,→ H. Let x2 = PL2 x for any x ∈ H.
Prove that PL1 x = PL1 x2 .

Lemma 3.4.4. If E is a closed subspace of a Hilbert space H,


and codimE = 1, then the subspace E ⊥ is 1-dimensional.

Proof. By Theorem 3.4.3, we have E ⊕ E ⊥ = H. Suppose there


exists two linearly independent vectors x1 , x2 in E ⊥ . Then by

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.5. Linear Functionals 75

Gram-Schmidt orthogonalization process, we get two orthogonal


vectors y1 , y2 in E ⊥ . Now, if αy1 + βy2 = z ∈ E, since y1 , y2 ∈
E ⊥ , it follows that α = β = 0 and hence z = 0. This shows
that y1 + E, y2 + E are linearly independent vectors in H/E,
which is a contradiction to the assumption that codimE = 1.
This shows that the subspace E ⊥ is 1-dimensional.

3.5 Linear Functionals

Linear functionals on a general linear space E over a field K = R


or C, are just linear transformations from E to K. More pre-
cisely, we have the following definition.

Definition 3.5.1. Let E be a linear space over R (or C). Linear


functionals are functions f : E 7→ R (or correspondingly C) such
that
f (λx + µy) = λf (x) + µf (y),

for all x, y ∈ E and for all λ, µ ∈ R (or C).

For any linear functional f on E, the kernel of f,

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.5. Linear Functionals 76

kerf = {x ∈ E : f (x) = 0} is a linear subspace of E, which


we often denote by Hf .

Example 3.5.2. Consider the space c0 of all sequences that


P
converges to zero. For any scalar sequence (bi ) with |bi | <
∞, i.e., (bi ) ∈ `1 , define the linear functional on c0 as f (x) =
P∞
1 ai bi . (Verify it!) We will identify such a functional f with
the element (bi ) of `1 .
In a similar way, in the linear space `p , 1 < p < ∞, with any
(bi ) ∈ `q , one can define a linear functional f by f (x) = ∞
P
1 ai b i .
By using Hölder’s inequality for sequences (See Theorem 2.2.5),

X
|f (x)| ≤ |ai | · |bi | ≤ kxk`p kf k`q < ∞,
1

we see that f is defined for any x = (ai ) ∈ `p . We identify this


functional with the element (bi ) of `q .

Example 3.5.3. Now we will give two linear functionals in the


space C[0, 1].
R1
Define F (x) = 0 x(t)f (t)dt, where f is some integrable func-
R1
tion. This is well defined since |F (x)| ≤ maxt |x(t)| 0 |f (t)|dt.
Evaluation at some fixed point α ∈ [0, 1], defined by δα (x) =

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.6. Bounded Linear Functionals 77

x(α) is a linear functional on C[0, 1], since |δα (x)| ≤ kxkC[0,1] .

For any linear space E, the set of all linear functionals on


E is a linear space and is denoted by E # . The proof of the
following theorem is left as an exercise.

Theorem 3.5.4. Let f be a non-zero linear functional on a


linear space E. Then,

• codim ker f = 1

• If f, g ∈ E # \{0} and ker f = ker g, then there exists


λ 6= 0 such that λf = g.

• Let L ,→ E and codim L = 1. Then there exists f ∈ E #


such that ker f = L.

3.6 Bounded Linear Functionals

Let X = (E, k · k) be a normed space. A linear functional on X


is said to be a bounded functional if there exists C > 0 such that
|f (x)| ≤ Ckxk for all x ∈ X. i.e., if f is bounded on bounded
sets.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.6. Bounded Linear Functionals 78

We say that a linear functional f on X is a continuous func-


tional if xn → x, implies that f (xn ) → f (x).
Let X ∗ denotes the set of all bounded linear functionals on
the normed space X. Then X ∗ is clearly a linear space with
respect to point-wise addition and scalar multiplication. Define
a norm in X ∗ as follows: For f ∈ X ∗ , let

kf k∗ = sup |f (x)|/kxk = sup |f (x)|


x6=0 kxk=1

(Verify that this indeed defines a norm on X ∗ .) In particular,


we see that
|f (x)| ≤ kf k∗ · kxk.

We often write kf k instead of kf k∗ .


If f is a bounded linear functional on X, then for any se-
quence (xn ) with xn → x0 , we have

|f (xn ) − f (x0 )| = |f (xn − x0 )| ≤ kf k · kxn − x0 k → 0,

as n → ∞. This shows that f is a continuous linear functional.


Now assume that f is a continuous linear functional. Sup-
pose f is not bounded. Then for any n ∈ N, there exists xn

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.6. Bounded Linear Functionals 79

in X, with kxn k = 1 and |f (xn )| > n. Consider the sequence


yn = xn /n. Then, yn → 0, so that we must have f (yn ) → 0.
But |f (yn )| = |f (xn )|/n ≥ 1, which is a contradiction. Thus,
we have observed that:
A linear functional f on a normed space is a bounded
functional if and only if f is a continuous functional.
The linearity of f is very crucial and we can observe that if
a linear functional f on X, is continuous at 0, then it
is continuous at any x ∈ X. To see this, let xn → x, then
xn − x → 0, and by the continuity at 0, we have f (xn ) − f (x) =
f (xn − x) → f (0) = 0, implies that f (xn ) → f (x).

Exercise 3.10. A linear functional f on a normed space X is


continuous if and only if kerf is a closed subspace of X.

Exercise 3.11. For any bounded linear functional f on X,


prove that

kf k = sup{|f (x)| : kxk ≤ 1}


= sup{1/kxk : f (x) = 1}
1
=
inf{kxk : f (x) = 1}

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.6. Bounded Linear Functionals 80

Now, we turn our attention to bounded linear functionals on


a Hilbert space. For any fixed element y in a Hilbert space H,
define f (x) = hx, yi. Clearly, f is a linear functional in H and
by using Cauchy-Schwartz inequality (Theorem 3.1.3), we see
that |f (x)| = |hx, yi| ≤ kxk · kyk.
Now suppose that f is any linear functional on H that is
continuous too. Does there exist a y ∈ H such that f (x) =
hx, yi ?. The answer is yes, and this is the content of the following
result, known as Riesz representation theorem.

Theorem 3.6.1. For every ϕ ∈ H ∗ , there exists a unique y ∈ H


such that
ϕ(x) = hx, yi, x ∈ H;

(i.e., any continuous linear functional on a Hilbert space H is


represented by a unique element of the same space H.) Moreover,
this correspondence is an isometry: kϕk∗ = kyk. Conversely, for
any y ∈ H, the expression ϕ(x) = hx, yi defines a continuous
linear functional on H, with the norm kϕk∗ = kyk.

Proof. For any fixed y ∈ H, define ϕy (x) = hx, yi.


From above discussions, we see that ϕy is linear and satisfies
|ϕy (x)| ≤ kxk · kyk. This implies kϕy k∗ ≤ kyk.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.6. Bounded Linear Functionals 81

Setting x = y, we get ϕy (y) = hy, yi = kyk2 ≤ kϕy k∗ kyk, implies


kyk ≤ kϕy k∗ .
Thus, kϕy k∗ = kyk.
Now let us show that any continuous linear functional ϕ on
H is of the form ϕ(x) = hx, yi, for some y ∈ H.
If ϕ = 0, we can simply take y = 0.
If ϕ 6= 0, then L = kerϕ is closed in H and codimL = 1.
Therefore, L⊥ is 1-dimensional, so that L⊥ = span{ŷ} for some
ŷ ∈ H\{0}.
Using this vector ŷ, define the linear functional ϕŷ = hx, yi.
Then kerϕŷ = {ŷ}⊥ = (L⊥ )⊥ = L. Thus, kerϕŷ =kerϕ and
hence by the second part of Theorem 3.5.4, we get ϕ = λϕŷ , for
some scalar λ.
Put y = λŷ. Then, ϕ(x) = λϕŷ (x) = λhx, ŷi = hx, yi. Also, we
have noted in this case, kϕk = kyk.
Now to prove the uniqueness of y, let ϕ(x) = hx, yi = hx, y1 i
for all x ∈ H. Then, let x = y − y1 , we get hy − y1 , yi =
ϕ(y − y1 ) = hy − y1 , y1 i implies hy − y1 , y − y1 i = 0, which shows
that y = y1 . This completes the proof.

Most of the results about Hilbert spaces that we discussed

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.6. Bounded Linear Functionals 82

deal with separable Hilbert spaces, but the Projection theorem


(Theorem 3.4.3) and the Riesz representation theorem (Theo-
rem 3.6.1) are valid without such assumption. We now give an
example for a non-separable Hilbert space.
Consider the linear space L of continuous functions on R
such that Z T
1
p(f ) = ( lim |f (x)|2 dx)1/2
T →∞ 2T −T

is defined for f ∈ L. This p is only a semi norm on L, since


it may happen that p(f ) = 0 for some non-zero functions too.
(For instance, p(f ) is zero for any function with finite support.)
Let N = kerp and consider the quotient space L/N. Then it
is a normed space and p is a norm on it. Let H be the completion
of this space with respect to the norm p and the inner product
Z T
1
hf, gi = lim f (x)g(x)dx.
T →∞ 2T −T

Note that the set {eiλt }λ∈R is an uncountable orthonormal sys-


tem in H, and hence H is not separable. (If a Hilbert space H
has an uncountable orthonormal system, say {et }t∈R , then the
open balls {B(et , 1/2)} are pair wise disjoint and hence H can-

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

3.6. Bounded Linear Functionals 83

not have a countable dense subset, as any such set should put
at least one element in each of the above uncountable, disjoint
balls. )

Exercise 3.12. Prove that for any two subspaces L1 and L2 of


a Hilbert space, (L1 + L2 )⊥ = (L⊥
T ⊥
1 L2 ).

Exercise 3.13. Let L1 and L2 be closed subspaces of a Hilbert


space H. Then prove (L1 L2 )⊥ equals the closure of L⊥ ⊥
T
1 + L2 .

Exercise 3.14. If H1 and H2 are two Hilbert spaces and if


T : H1 → H2 is an isometry, then prove that hT x, T yi = hx, yi
for every x, y ∈ H1 .

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

Chapter 4
The Dual Space of a Normed
Space

This chapter deals with the duality between a normed space


X and the space X ∗ of all bounded linear functionals on it.
The linear space X ∗ of all bounded linear functionals on X is
equipped with the norm,

|f (x)|
kf k = sup ,
x6=0 kxk

84
MTH3C13 : FUNCTIONAL ANALYSIS
4.1. The Hahn-Banach Theorem and its
Consequences. 85

we call this norm as the dual norm (i.e., dual to the original
norm of X), and the normed space X ∗ as the dual space (i.e.,
dual to the space X). From the definition of dual norm, it is
clear that
|f (x)| ≤ kf k · kxk.

It can be proved that the dual space X ∗ is always complete


and hence is a Banach space for any normed space X.

4.1 The Hahn-Banach Theorem and


its Consequences.

We now state Hahn-Banach extension theorem, which is


one of the fundamental theorem in the theory of normed spaces.

Theorem 4.1.1. (Hahn-Banach)


Let X be a normed space over R, E be a subspace of X, and
f0 ∈ E ∗ . Then, there exists an extension f ∈ X ∗ such that
f |E = f0 (i.e., f (x) = f0 (x) for all x ∈ E ) and kf kX ∗ = kf0 kE ∗ .
i.e.,
|f (x)| |f0 (x)|
sup = sup .
x∈X\{0} kxk x∈E\{0} kxk

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS
4.1. The Hahn-Banach Theorem and its
Consequences. 86

We now derive some corollaries to this theorem.

Corollary 4.1.2. For all x0 ∈ S(X) = {x ∈ X : kxk = 1},


the unit sphere of X, there exists f0 ∈ X ∗ such that kf0 kX ∗ = 1,
and f0 (x0 ) = 1. So the functional f0 attains its supremum on
the unit ball on the vector x0 .

Proof. Let E0 = {λx0 : λ ∈ R. Then, E0 is a 1-dimensional


subspace of X. Define the functional ϕ on E0 as ϕ(λx0 ) = λ.
Then kϕkE0∗ = 1. By the Hahn-Banach theorem, there exists an
extension f0 with the desired properties.

Corollary 4.1.3. For all x0 ∈ X, there exists f0 ∈ X ∗ \{0} such


that f0 (x0 ) = kf0 k · kx0 k.

Proof. If x0 = 0, then the desired property holds for any f0 ∈


X ∗ \{0}. If x0 6= 0, set y = x0 /kx0 k. Then y ∈ S(X) and by
above corollary 4.1.2, there exists f0 ∈ X ∗ \{0} with the desired
properties.

Corollary 4.1.4. For all x1 ∈ X and for all x2 ∈ X such that


x1 6= x2 , there exists f ∈ X ∗ satisfying f (x1 ) 6= f (x2 ).

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS
4.1. The Hahn-Banach Theorem and its
Consequences. 87

Proof. Apply Corollary 4.1.3 for the element x0 = x1 − x2 , we


get f0 ∈ X ∗ \{0} such that f0 (x0 ) = f0 (x1 − x2 ) = f0 (x1 ) −
f0 (x2 ) = kf0 kkx1 − x2 k =
6 0, implies f0 (x1 ) 6= f0 (x2 ).

Corollary 4.1.5. X ∗ is a total set. i.e., if f (x) = 0 for all


f ∈ X ∗ , then x = 0.

Proof. If x 6= 0, then by Corollary 4.1.4, there exists f ∈ X ∗


satisfying f (x) 6= 0 = f (0), which is a contradiction.

Corollary 4.1.6. Let X be a normed space over R and let L ,→


X be a subspace. If x ∈ X is such that dist(x, L) = d > 0, then
there exists f ∈ X ∗ such that kf k = 1, f (L) = 0 and f (x) = d.

Proof. Let L1 = span{x, L}, i.e.,

L1 = {λx + y : λ ∈ R, y ∈ L}.

Then L1 is a subspace of X, and define


f0 : L1 → R by f0 (λx + y) = λ · d.
Then f0 is well-defined since any element z ∈ L1 has a unique
expression of the form z = λx+y with λ ∈ R and y ∈ L. Clearly
f0 is linear, hence f0 is a linear functional on L1 , f0 (L) = 0 and
f0 (x) = d.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS
4.1. The Hahn-Banach Theorem and its
Consequences. 88

Now, kzk = kλx + yk = |λ| · kx + y/λk ≥ |λ| · d = |f0 (z)|,


since (−y/λ) ∈ L and the distance from x to (−y/λ) is at least d.
This shows that kf0 k ≤ 1. Moreover, there exists yn ∈ L so that
kx+yn k → d. Hence d = |f0 (x+yn )| ≤ kf0 k·kx+yn k → d.kf0 k,
implies kf0 k ≥ 1. Thus kf0 k = 1.
By Hahn-Banach theorem 4.1.1, there exists an extension f
of f0 to all of X, with kf0 k = kf k. Therefore, we have f ∈ X ∗
such that kf k = 1, f |L1 = f0 , which means f (L) = 0 and
f (x) = d.

Corresponding to any subspace L of a normed space X, we


associate a subspace of X ∗ , namely, the annihilator L⊥ of L,
which is defined by

L⊥ = {f ∈ X ∗ : f (L) = 0}.

It is clear that L⊥ is a closed subspace of X ⊥ .


Now, if we start with a subspace F of X ∗ , then the annihi-
lator F ⊥ is a closed subspace of (X ∗ )∗ := X ∗∗ . Corresponding

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS
4.1. The Hahn-Banach Theorem and its
Consequences. 89

to this F, we consider the subspace F⊥ of X, defined by


\
F⊥ = {x ∈ X : f (x) = 0, for all f ∈ F } = kerf.
f ∈F

For any closed subspace L of a normed space X, we have the


following result.

Corollary 4.1.7. Let L be a closed subspace of X. Consider


the subspaces L⊥ ,→ X ∗ and (L⊥ )⊥ = {x ∈ X : f (x) =
0, for all f ∈ L⊥ }. Then (L⊥ )⊥ = L.

Proof. If x ∈ L, then f (x) = 0 for all f ∈ L⊥ , so that x ∈


kerf. Thus, L ⊂ (L⊥ )⊥ .
T
f ∈L⊥

Now, if x ∈/ L, we have dist(x, L) = d > 0, as L is closed. By


Corollary 4.1.6, there exists f such that f (L) = 0, i.e., f ∈ L⊥
/ (L⊥ )⊥ . This implies (L⊥ )⊥ ⊂ L.
and f (x) 6= 0. Therefore, x ∈
Thus, (L⊥ )⊥ = L.

Remark 4.1.8. Note that for any subspace L, not necessarily


closed, (L⊥ )⊥ = L̄.

Corollary 4.1.9. Let X be a normed space and let {x1 , x2 , ..., xn }


be a linearly independent subset of X. Then, there exists

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

4.2. Reflexive Spaces 90

f1 , f2 , ..., fn ∈ X ∗ such that fi (xj ) = δij , where δij is the Kro-


necker’s delta function,
(
1 if i = j
δij =
0 6 j
if i =

n
and every x ∈span{xi }n1 can be expressed as x =
P
fi (x)xi .
i=1

Proof. For any fixed i0 , let Li0 = span{xi }i6=i0 . As every finite
dimensional subspace of a normed space is closed (Try to prove
this!), Li0 is a closed subspace of X and xi0 ∈/ Li0 .
By Corollary 4.1.6, there exists fi0 ∈ X ∗ such that fi0 (Li0 ) =
P
0, fi0 (xi0 ) = 1. Thus, if x = ai xi ,then ai = fi (x). This com-
pletes the proof.

4.2 Reflexive Spaces

Let X be a normed space and let X ∗ be its dual space. We


have noted that X ∗ is always a Banach space. We now consider
the dual space (X ∗ )∗ of X ∗ , which we denote by X ∗∗ . We may
identify X with a subspace of X ∗∗ in the following way: For

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

4.2. Reflexive Spaces 91

x ∈ X, consider the functional i(x) ∈ X ∗∗ such that for every


f ∈ X ∗ , i(x)(f ) = f (x.) Then clearly, i : X → X ∗∗ defines a
one-to-one linear map. Also,

|i(x)(f )| ≤ kf k · kxk.

This implies that,

|i(x)(f )| |f (x)|
ki(x)k∗∗ = sup = sup ≤ kxk.
f 6=0 kf k f 6=0 kf k

But, Corollaries 4.1.2, and 4.1.3, shows that the last supremum
is attained and for any x ∈ X, there exists f 6= 0 such that
|f (x)| = kxk · kf k. Therefore, ki(x)k∗∗ = kxk. Thus, the map
i : X → X ∗∗ is an isometry into.
We say that a normed space X is a reflexive space if the
map i : X → X ∗∗ is onto. Thus, in case of reflexive spaces, the
spaces X and X ∗∗ are isometric, and we can identify them in a
natural way. For instance, in case of reflexive spaces, we do not
need to distinguish between F ⊥ and F⊥ , and in this case, the
conclusion of Corollary 4.1.7 becomes (L⊥ )⊥ = L.
We now consider the dual spaces of some normed spaces.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

4.2. Reflexive Spaces 92

Example 4.2.1. Consider the space c0 of null sequences (i.e.,


the set of all scalar sequences that converges to 0) equipped
with the norm kxk = sup |ai |, for x = (ai ). Fix a sequence
i
b = (bi ) ∈ `1 and using it, define a linear functional on c0 as
P
f (x) = i ai bi for x = (ai ) ∈ c0 . Then
X X
|f (x)| ≤ |ai bi | ≤ sup |ai | · |bi | = kxkc0 · kbk`1 .
i
i i

Therefore, kf kc∗0 ≤ kbk`1 .

Thus, starting with b = (bi ) ∈ `1 , we get f ∈ c∗0 , and


for this f, we have f (en ) = bn , for n = 1, 2, ..., where en =
(0, 0, ..., 0, 1, 0, ...) ∈ c0 , 1 occurs only at the nth place.
Similarly, starting with a f ∈ c∗0 , setting bn = f (en ) for
n = 1, 2, ..., we get b = (bn )n∈N ∈ `1 . To see this, for each
n = 1, 2, ..., consider the sequence

yn = (e−i argb1 , e−i argb2 , ..., e−i argbn


, 0, 0, ...)
Xn
= e−i argbk ek
k=1

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS
4.2. Reflexive Spaces 93

Then yn ∈ c0 and kyn kc0 = 1 for all n = 1, 2, ... By the definition


of norm of a functional,

kf kc∗0 ≥ |f (yn )|
≥ f (yn )
Xn
= f( e−i argbk
ek )
k=1
n
X
= e−i arg bk
bk
k=1
n
X
= |bk |.
k=1

(Here, we used the idea, if z = reiθ , then e−iθ z = |z|.) This


is true for any n = 1, 2, 3, ... This implies kf kc∗0 ≥ kbk`1 . This
shows that b ∈ `1 .
Thus, any f ∈ c∗0 is identified with an element b ∈ `1 and
kf kc∗0 ≥ kbk`1 .
For any such b ∈ `1 , we proved above that kf kc∗0 ≤ kbk`1 .
Combining these inequalities we obtain, kf kc∗0 = kbk`1 . Hence
we can identify the dual space of c0 as `1 , and this identification
is an isometry.(i.e., preserving the norms.)

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS
4.2. Reflexive Spaces 94

In a similar manner, one can prove that `∗1 = `∞. Thus c0 is


not a reflexive space.

Example 4.2.2. Consider the space of p−summable scalar se-


quences `p , with 1 < p < ∞. We claim that (`p )∗ = `q , where
1
p
+ 1q = 1. First let b = (bi ) ∈ `q . Then b defines a linear func-
tional on `p by the formula

X
f (x) = ai bi ,
i=1

for any x = (ai ) ∈ `p . Also, by Hölder’s inequality,


X X X
|f (x)| ≤ |ai bi | ≤ ( |ai |p )1/p · ( |bi |q )1/q .

i.e., |f (x)| ≤ kbk`q · kxk`p .

This implies, kf k`∗p ≤ kbk`q , and hence f ∈ `∗p .


Thus, corresponding to b = (bi ) ∈ `q , we identified a bounded
linear functional f ∈ `∗p , such that f (en ) = bn for n = 1, 2, ...,
where en = (0, 0, ..., 0, 1, 0, ...) ∈ `p , 1 occurs only at the nth
place.
Now, start with f ∈ `∗p , and set bn = f (en ).

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

4.2. Reflexive Spaces 95

n
e−i argbk
|bk |q−1 ek .
P
Then b = (bn ) ∈ `q . To see this, set yn =
k=1
Then yn ∈ `p and

Xn n
X
(q−1)p 1/p
kyn k`p = ( |bk | ) =( |bk |q )1/p ,
1 1

since (q − 1)p = q. (To see this recall that 1/p + 1/q = 1, so that
p
q = p−1 .)
Also,

kf k`∗p kyn k`p ≥ |f (yn )|


X n
= ( |bk |q )
k=1
Xn Xn
q 1/p
= ( |bk | ) · ( |bk |q )1/q
k=1 k=1
Xn
= kyn k`p · ( |bk |q )1/q
k=1

Letting n → ∞, we get kf k`∗p kyn k`p ≥ kyn k`p · kbk`q , which


implies, kf k`∗p ≥ kbk`q , showing that b = (bi ) ∈ `q .
Thus, any f ∈ `∗p is identified with an element b ∈ `q and

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

4.2. Reflexive Spaces 96

kf k`∗p ≥ kbk`q .
For any such b ∈ `q , we proved above that kf k`∗p ≤ kbk`q .
Combining these inequalities we get kf k`∗p = kbk`q . Therefore,
we can isometrically identify (`p )∗ with the space `q . This shows
that all the `p -spaces, 1 < p < ∞ are reflexive spaces.
A similar result holds also in case of function spaces Lp ,
namely, L∗p = Lq , for 1 < p < ∞.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

Chapter 5
Bounded Linear
Operators

In this chapter, we concentrate on bounded linear maps between


normed spaces. Also we introduce a norm on the linear space
of all bounded linear maps from a normed space to another.
We note that a linear map is a bounded operator if and only
if it is continuous.

97
MTH3C13 : FUNCTIONAL ANALYSIS

5.1. The Space of Bounded Linear Operators 98

5.1 The Space of Bounded Linear Op-


erators

Let X and Y be Banach spaces and let T : X → Y be a linear


map (usually called an operator) defined on X. Then, T is said
to be bounded if there exists C > 0 such that kT xkY ≤ CkxkX
for all x ∈ X. Thus, a linear map T : X → Y is bounded if T
maps bounded sets of X into bounded sets of Y.
If T is bounded, define

kT xk
kT k = sup
x6=0 kxk

Clearly, if kT xk ≤ C for every x ∈ X with kxk = 1, then


kT xk x  x
kT k ≤ C. (Since, kT k = sup = supkT k, here is
x6=0 kxk x6=0 kxk kxk
x
of norm 1, therefore we have kT ( )k ≤ C )
kxk
Moreover, the family of linear bounded operators is a linear
space with respect to the addition and the multiplication of
operators by scalars (Indeed, if T is such that kT xk ≤ C1 kxk and
U is such that kU xk ≤ C2 kxk, then k(αT + U )xk ≤ |α|kT xk +
kU xk ≤ (|α|C1 + C2 )kxk ). We now check that kT k defines a

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.1. The Space of Bounded Linear Operators 99

norm.

1. Clearly, kT k ≥ 0

2. T = 0 ⇒ kT k = 0, if kT k = 0 then T = 0. Suppose
if possible T 6= 0, then there exists an x 6= 0 such that
x c
T x = c > 0, then T ( ) = > 0 contradicts that
kxk kxk
kT k = 0

3.

k(T + U )xk
kT + U k = sup
x6=0 kxk
kT xk + kU xk
≤ sup
x6=0 kxk
≤ kT k + kU k

k(αT )xk |α|kT xk


4. kαT k = sup = sup = |α|kT k
x6=0 kxk x6=0 kxk

kT xk
Thus, kT k = sup defines a norm, and is called the oper-
x6=0 kxk
ator norm.
We write L(X → Y ) or L(X, Y ) to denote the linear space
of bounded operators with the above norm. From the definition

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.1. The Space of Bounded Linear Operators 100

of the norm kT k, it follows that for every x ∈ X we have

kT xk ≤ kT k · kxk.

If H is a Hilbert space, then for every bounded linear oper-


ator A : H → H we have

kAk = sup{|hAx, yi| : kxk ≤ 1, kyk ≤ 1}

To see this, let t = sup{|hAx, yi| : kxk ≤ 1, kyk ≤ 1}. Now by


applying the Cauchy-Schwartz inequality, we get
|hAx, yi| ≤ kAxkkyk ≤ kAkkxkkyk ⇒ t ≤ kAk.
kAxk
Now we need to show t ≥ kAk = sup .
x6=0 kxk

It is sufficient to show that kAxk ≤ tkxk whenever x 6= 0.


Now, use the fact that
x Ax
kAxk2 = kxkkAxkhA( ), i ≤ tkxkkAxk, we get
kxk kAxk
kAxk ≤ tkxk.

Theorem 5.1.1. Let X be a normed space and let Y be a com-


plete normed space. Then L(X, Y ) is a Banach space (i.e., it is
a complete normed space).

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.1. The Space of Bounded Linear Operators 101

Proof. Let {An }∞ n=1 be a Cauchy sequence in L(X, Y ), so that


for all  > 0 there exists N ∈ N such that if m, n ≥ N we have

kAn − Am k ≤ 

This implies that for every x ∈ X and m, n ≥ N

kAn (x) − Am (x)kY = k(An − Am )(x)kY


≤ kAn − Am k · kxk
≤ kxk.

Therefore, for all x ∈ X, the sequence {An (x)}∞


n=1 is a Cauchy
sequence in Y .
Since Y is a Banach space, this sequence has a limit, say A(x)
in Y , and thus, we define for all x ∈ X, A(x) = limn→∞ An (x).
Then A is a linear operator and it is also bounded, since

kA(x)kY ≤ supkAn (x)k ≤ kxkX · supkAn k,


n∈N n∈N

hence
kAk ≤ supkAn k.
n∈N

Here, we use the fact that a Cauchy sequence is bounded. Here

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.1. The Space of Bounded Linear Operators 102

both An and An x are Cauchy and therefore both supremums


are finite. Also kAxk = k lim An xk = lim kAn (x)k.
Thus, A ∈ L(X, Y ).
Now it remains to show that An → A.
Since {An } is a Cauchy sequence, for any  > 0 there exists N
such that for any m, n > N we have kAn − Am k < .
This means that for such m, n and for any x such that kxk ≤
1 we have kAm x − An xk < . Letting m tends to infinity, we
obtain kAx − An xk ≤  for any x with kxk ≤ 1.
Hence kA − An k ≤  for any n > N, which implies

lim kA − An k = 0.
n→∞

Next, we note that a linear map A is a bounded opera-


tor if and only if A is a continuous operator(i.e.,Axn −
Ax → 0 for any sequence xn → x).
First, assume that A is a bounded operator. We have to
show that xn → x ⇒ Axn → Ax.
kAxn − Axk = kA(xn − x)k ≤ kAkkxn − xk → 0 since xn → x.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.1. The Space of Bounded Linear Operators 103

Therefore, A is continuous.
Conversely suppose that A is continuous. Then A is con-
tinuous at 0 also, which implies ∃δ > 0 such that kxk < δ ⇒
kAxk < 1.
Therefore, whenever y 6= 0, kA(δy/kyk)k < 1 ⇒ kAyk ≤ kyk/δ,
when y = 0 trivial. Therefore A is a bounded operator. Thus,
a linear map A is a bounded operator if and only if A is a con-
tinuous operator.
Also, note that, for any bounded linear operator A,
the set ker A = {x : Ax = 0} is a closed subspace.
Indeed, if xn be a sequence in ker A, then Axn = 0 for all
xn . If xn → x, then 0 = Axn → Ax = 0. Therefore, x ∈ ker A.
In Chapter 4, we have observed that for any normed space
X, the dual space X ∗ is complete. Indeed, take Y to be the field
R or C (depending over which field, the space X is considered).
Then L(X, Y ) = X ∗ .
Next, let X, Y and Z be Banach spaces and let
A ∈ L(X, Y ), B ∈ L(Y, Z). Then BA ∈ L(X, Z) and

kBAk ≤ kBkkAk.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.1. The Space of Bounded Linear Operators 104

To see this, suppose xn → x ⇒ Axn → Ax ⇒ B(Axn ) →


B(Ax). Therefore, BA ∈ L(X, Z).
Also, kB(Ax)k ≤ kBkkAxk ≤ kBkkAkkxk.
Therefore, kBAk ≤ kBkkAk.

Now we illustrate some examples of linear operators.

Example 5.1.2. Consider the space C[0, 1]. For f ∈ C[0, 1],
define Z 1
Af = K(t, τ )f (τ )dτ,
0

where K is a continuous function of two variables. The operator


A is linear and
Z 1
kAf kC[0,1] ≤ max|f |.max |K(t, τ )|dτ.
t 0

R1
So, kAk ≤ max 0
|K(t, τ )|dτ .
t

It can be shown that


Z 1
kAk = max |K(t, τ )|dτ.
t 0

Example 5.1.3. In L2 [0, 1] and for an element K(t, τ ) of L2 (I 2 ),

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.1. The Space of Bounded Linear Operators 105

where I 2 = [0, 1] × [0, 1], define the operator

K : L2 [0, 1] 7→ L2 [0, 1]

by Z 1
(Kf )(t) = K(t, τ )f (τ )dτ.
0

The function of two variables K(t, τ ) is called the kernal(or the


kernal f unction) of the operator K. One may easily check using
the Cauchy-Schwartz inequality that the norm of the operator
K(denoted here by kKkop to avoid confusion) satisfies

kKkop ≤ kK(t, τ )kL2 (I 2 )

Indeed, we have
Z 1 Z 1 1/2 Z 1 1/2
2 2
|K(t, τ )f (τ )|dτ ≤ |K(t, τ )| dτ |f (τ )| dτ .
0 0 0

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.1. The Space of Bounded Linear Operators 106

Hence,
Z 1 Z 1 2
2
kKf k ≤ |K(t, τ )f (τ )|dτ dt
0 0
Z 1 Z 1
2
≤ kf k |K(t, τ )|2 dτ dt
0 0

and it follows that,

kKkop ≤ kK(t, τ )kL2 (I 2 ) .

Example 5.1.4. Let k(t) be a continuous function on [a, b]. In


L2 [a, b] define the operator A by Af = k(t) · f (t). Then A is a
bounded linear operator and

kAk := M := max |k(t)|.


a≤t≤b

It is immediate that kAk ≤ M . To show that, kAk = M ,


suppose M = |k(t0 )|. Define a sequence {xn } by

 n/2, t ∈ [t0 − 1 , t0 + 1 ],
p

xn (t) = n n
0, otherwise.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.1. The Space of Bounded Linear Operators 107

Now Z t0 +1/n
2 n
kxn k = = 1,
t0 −1/n 2
and from the continuity of k(t) at t0 , it follows that
Z t0 +1/n
2 n 2
kAk ≥ kAxn k = |k(t)|2 dt → |k(t0 )|2 .
2 t0 −1/n

Hence
kAk ≥ kk(t0 )| = M.

Example 5.1.5. The shift operator in `2 defined by T x =


(0, a1 , . . . , an , . . . ) for (an ) ∈ `2 satisfies kT xk = kxk for every x
and therefore kT k = 1

Example 5.1.6. Let (αij )∞ i,j=1 be an infinite matrix and let



K 2 = i,j=1 |αij |2 < ∞. Then the operator A defined in `2 by
P

A ((α)∞
i=1 ) = (βi )

where ∞
X
βi = αij αj
j=1

for i ∈ N is a bounded linear operator .

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.1. The Space of Bounded Linear Operators 108

To see this one has to check that kA(αi )k`2 ≤ K.k(αi )k`2 (again,
one should use the Cauchy-Schwartz inequality).

Example 5.1.7. Let H be a separable Hilbert space and let


{ei }∞
i=1 be an orthonormal basis. Let A : H → H be a bounded
linear operator.
P P
We have Ax = i≥1 hAx, ei iei and x = j≥1 xj ej , where
xj = hx, ej i.
From the second relation, since A is a bounded operator, we
P
get Ax = j≥1 hx, ej iAej . Substituting in the first relation, we
obtain
XDX E XX
Ax = hx, ej iAej , ei ei = hAej , ei ihx, ej iei .
i≥1 j≥1 i≥1 j≥1

P P
Set aij = hAej , ei i. Thus if x = i≥1 xi ei , y = i≥1 yi ei and
P
y = Ax, we get that yi = j≥1 aij xj .
 ∞
Hence, the sequence (αi ) is mapped by A to βi = rj hAej , ei iαi
P
.
i=1
Consequently, the operator A is defined by the matrix (hAej , ei i)i,j .
However, the example 5.1.6 may not be applicable, since the
|αij |2 < ∞ usually is not satisfied, for instance,
P
condition
consider the example of the identity operator I : H → H.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.2. Compact Operators 109

5.2 Compact Operators

In this section, we introduce the compact linear maps, which


are very important in applications. For instance, they play a
central role in the theory of integral equations and in various
problems of mathematical physics. Their theory served as a
model for the early work in functional analysis. Their properties
closely resembles those of operators on finite dimensional spaces.
Compact linear maps are defined as follows:

Definition 5.2.1. A linear operator A : X → Y is called a


compact operator if and only if for every bounded sequence
xn ∈ X the sequence {Axn } has a Cauchy subsequence.

Clearly, a compact operator must be bounded (and hence is


continuous) because the image of the unit ball of X must be a
bounded set in Y (otherwise we can easily build a sequence xn
inside the ball of X and such that kAxn k → ∞, thus unbounded
and hence without a Cauchy subsequence).
In order to work with compact operators, we should first
understand well, the notion of a compact set.
A set K ⊆ X is called a compact set if and only if for every

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.2. Compact Operators 110

sequence xn ∈ K there exists an x ∈ K and a subsequence xni


of xn so that xni → x.
K is said to be relatively compact (or precompact) if
every sequence xn ∈ K has a Cauchy subsequence xni . For
example, if X is complete and K is relatively compact, then the
closure K is compact.
From the Heine- Borel theorem of Real Analysis, it follows
that any bounded set in Rn is relatively compact.
We now state an important result due to Arzela .

Theorem 5.2.2. (Arzela) Let M ⊆ C[a, b]. M is relatively


compact (in C[a, b]) if and only if M is
(i) uniformly bounded( i.e., it is a bounded set in C[a, b]) and
(ii) equicontinuous: for any  > 0 there exists δ > 0 such that
|t1 − t2 | < δ implies |f (t1 ) − f (t2 )| <  for any f ∈ M .

Notice that if F is any metric precompact space, then the


same theorem is true for M ⊆ C(F ).
The proofs of the facts given in the following remark are left
as exercises.

Remark 5.2.3. (i) A subset of a precompact set is precompact.


(ii) For a bounded linear operator A if the image of the unit ball

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.2. Compact Operators 111

is precompact, then the image of any ball is precompact.


(iii)An operator A : X → Y is compact if and only if the image
of the unit ball of X is a precompact set in Y .

Example 5.2.4. Let C1 and C2 be positive constants and let

M = {x(t) ∈ C[a, b] : |x(t)| < C1 and |x0 (t)| < C2 }.

Then M is relatively compact. (To see this, use Arzela’s the-


orem: |x0 (t)| < C2 implies the uniform continuity: (x(t1 ) −
x(t2 ))/(t1 − t2 ) = x0 (θ) for t1 < θ < t2 , and therefore, |x(t1 ) −
x(t2 )| ≤ C2 · |t1 − t2 |)
Rt
Example 5.2.5. The operator Ax = 0 x(τ )dτ on C[0, 1] is a
compact operator (use the previous example).

Example 5.2.6. Let C1 [a, b] be a space of continuous functions


on [a, b] with continuous derivative, and let

kx(t)kC1 = max|x(t)| + max|x0 (t)|.


t t

Consider the embedding operatorA : C1 [a, b] 7→ C[a, b] with


Ax = x. Then, A is a compact operator. The unit ball of C1 [a, b]
is a precompact set in C[a, b] by Arzela’s theorem.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.2. Compact Operators 112

Example 5.2.7. Let K(t, τ ) be a continuous function of two


variables on [0, 1] × [0, 1].
R1
Let us prove that the operator Kx = 0 K(t, τ )x(τ )dτ : C[0, 1] →
C[0, 1] is a compact operator. Let M = {x(t)} be a bounded set
of functions from C[0, 1] such that kxk ≤ r. It is clear that the
R1
functions y(t) = 0 K(t, τ )x(τ )dτ , where x(t) is a function from
M , are uniformly bounded. Indeed, if R = maxK(t, τ ), then
t,τ
|y(t)| ≤ Rr. It means that the operator K is bounded, which
was already proved before.
Next, the functions y(t) are uniformly continuous. Indeed,
given  > 0 by the uniform continuity of the kernel K(t, τ ) there
exists δ > 0 such that |K(t1 , τ ) − K(t2 , τ )| < /r for |t1 − t2 | < δ
and any τ ∈ [0, 1]. It follows that |t1 − t2 | < δ implies
Z 1
|y(t1 ) − y(t2 )| ≤ |K(t1 , τ ) − K(t2 , τ )||x(τ )|dτ < 
0

for any function y(t) from the set under consideration, which
means that the set is equicontinuous.
Thus, by the Arzela theorem we conclude that the set {y(t)}
is relatively compact in C[0, 1]. Hence the operator K is com-
pact.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.2. Compact Operators 113

To build more examples of compact operators, we need some


properties of compact sets and operators.

Definition 5.2.8. Let X be a metric space and let A ⊆ X. A


set A ⊂ X is called an  − net of A if and only if for all x ∈ A
there exists y ∈ A such that the distance from x to y is less
than .

Lemma 5.2.9. M is relatively compact if and only if for every


 > 0 there exists a finite  − net in M

Proof. Assuming that there exists 0 > 0 such that there is no


finite 0 − net of M , take any x1 ∈ M and choose x2 ∈ M such
that kx1 −x2 k ≥ 0 . If {xi }n−1
1 are already chosen, we can choose
xn ∈ M such that kxn − x1 k, kxn − x2 k, . . . , kxn − xn−1 k ≥ 0 .
Such an xn exists(for every n), otherwise {x1 , x2 , . . . xn−1 } is a
finite 0 − net of M . Therefore no subsequence of {xn }∞ n=1 is
a Cauchy sequence, which means that M is not precompact, a
contradiction.
For the converse, for all k ∈ N take an k ≤ 1/2k. M has a
finite k − net for every k(by the assumption). Let {xn }∞
n=1 ∈ M
and consider a finite 1 − net. It divides the sequence between a
finite number of balls around the net points. So, there is atleast

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.2. Compact Operators 114

one ball which contains an infinite subsequence from the original


sequence. Let us denote the subsequence contained within this
(1)
ball by {xn }∞ n=1 . In similar way, an k −net divides the sequence
(k−1) ∞
{xn }n=1 between a finite number of balls and there is one ball
which contains an infinite subsequence of the previously chosen
(k−1) (k)
sequence {xn }. Let us call it {xn }∞ n=1 . We know that on
this step the ball’s radius is less than k ; hence for all m, n ∈ N
we have
1
kx(k) (k)
m − xn k ≤ 2k ≤ .
k
(n)
The subsequence {xn }∞ ∞
n=1 of {xn }n=1 is a Cauchy sequence.
Indeed, for all η ≥ 0 we can choose ν ∈ N such that 1/ν ≤ η
and then for all m, n ≥ ν we have

1
kx(n) (m)
n − xm k ≤ ≤η
ν
(ν)
since both vectors belong to {xi }. This implies that M is
relatively compact.

We now list some properties of the space of compact opera-


tors as a theorem.

Theorem 5.2.10. The set K(X → Y ) of compact operators

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.2. Compact Operators 115

from X to Y satisfy
(i) K(X → Y ) is a linear subspace of L(X → Y ).
(ii) If A ∈ K(X → Y ), B ∈ L(Z → X), C ∈ L(Y → Z), then
AB ∈ K(Z → Y ), CA ∈ K(X → Z). In the case X = Y it
means that K(X) ≡ K(X → X) is a two-sided ideal of L(X →
X) ≡ L(X).
(iii) K(X → Y ) is a closed subspace of L(X → Y ).

Proof. (i) One can easily verify that K(X → Y ) is a linear


subspace of L(X → Y ).
(ii) For the operator AB, let {zn }∞
n=1 be a bounded sequence
in Z. B is bounded; hence {B(zn )}∞ n=1 is also bounded. A is

compact; hence {A(B(zn ))}n=1 has a converging subsequence
and thus AB is compact.
For the operator BA, let {xn }∞n=1 be a bounded sequence in
X. A is compact; hence the sequence {A(xn )}∞ n=1 has a con-

verging subsequence {A(xnk )}k=1 . B is bounded, so it follows
that the sequence B(A(xnk ))∞k=1 is also converging and thus BA
is compact.
(iii) Let An → A( i.e., kAn − Ak → 0) and let An ∈ K(X →
Y ). In order to show that A is compact, it is enough to prove

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.3. Dual Operators 116

that AD(X) (the image of the unit ball) is precompact. Thus,


we have to find for all  > 0 a finite  − net for AD(X).
For all  > 0 there exists n ∈ N such that kAn − Ak < /2
and An D(X) is precompact (An is a compact operator).
Take an /2−net {yi }N N
1 of An D(X). Then {yi }1 is an −net
for AD(X).
Indeed, for any y ∈ AD(X) there is x ∈ D(X) such that
y = Ax. Since {yi }N
1 is an /2 − net of An D(X), take yi0 from
the net such that kyi0 − An xk ≤ /2.
Then ky − yi0 k ≤ kAx − An xk + kAn (x) − yi0 k ≤ .

5.3 Dual Operators

Let A : X → Y be a bounded operator and let φ ∈ Y ∗ . Then


f (x) = φ(Ax) for x ∈ X defines a linear function on X.
Moreover,
|f (x)| ≤ kφkY ∗ · kAk · kxk.

Thus, f ∈ X ∗ . Hence, we have a map φ 7→ f := A∗ φ. It is clear


that this map A∗ : Y ∗ 7→ X ∗ is linear. So, for every x ∈ X and

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.3. Dual Operators 117

φ ∈ Y ∗ we have
φ(Ax) ≡ (A∗ φ)(x).

We check now that A∗ is also a bounded operator:

kAk = sup kA(x)kY


kxk=1

= sup sup |φ(A(x))|


kxk=1kφk=1

= sup sup |φ(A(x))|


kφk=1kxk=1

= sup sup |(A∗ φ(x))|


kφk=1kxk=1

= sup k(A∗ (φ))kX ∗


kφk=1

= kA∗ k.

We thus see that we always have kAk = kAk∗ . The operator


A∗ is called the dual operator to the operator A. Note that
A∗ : Y ∗ → X ∗ .
Let us introduce now more convenient notation to deal with
duality and dual operators. Let X be a (normed) linear space
and X ∗ be its dual. For f ∈ X ∗ and x ∈ X we write the action
f on x in a symmetric way as hx, f i instead of f (x)(i.e., by

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.3. Dual Operators 118

definition hx, f i = f (x).)


For a general linear space E and a space of functionals F on
E ( F may not coincide with E ∗ ) we call hx, f i the “pairing”
between E and F . We “copy” in this notation the inner product
notation but now the first and the second arguments belong to
different spaces. Applying this notation to the pairing between
X and X ∗ , we can rewrite the definition of the dual operator
A∗ of an operator A in the following(symmetric) way:
For all x ∈ X and for all f ∈ Y ∗ ,

hAx, f i = hx, A∗ f i

Observe that if A ∈ L(X, Y ) and B ∈ L(Y, Z), then (BA)∗ =


A∗ B ∗ : Z ∗ 7→ X ∗
(To see this, note: ((A∗ B ∗ )f )(x) = (A∗ (B ∗ f ))(x) = (B ∗ f )(Ax) =
f (B(Ax)) = ((BA)∗ f )(x))
In the case of Hilbert spaces, for an operator A ∈ L(H1 →
H2 ), the adjoint operator A ∈ L(H2 → H1 ) is defined via the
relation

hAx, yi = hx, A∗ yi for all x ∈ H1 , and for all y ∈ H2 .

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.3. Dual Operators 119

Example 5.3.1. Consider K(t, τ ) ∈ L2 (I 2 )(I = [0, 1]) and de-


R1
fine the operator K on L2 [0, 1] by (Kx)(t) = 0 K(t, τ )x(τ )dτ .
Let us show that the adjoint operator K ∗ is given by (K ∗ y)(t) =
R1
0
K(τ, t)y(τ )dτ .
Indeed, we have
Z 1 Z 1 
hKx, yi = K(t, τ )x(τ )dτ y(t)dt
0 0
Z 1 Z 1 
= x(τ ) K(t, τ )y(t)dt dτ
0 0

= hx, y i,

where Z 1

y (τ ) = K(t, τ )y(t)dt
0

Theorem 5.3.2. A bounded operator A : X → Y is compact


implies A∗ : Y ∗ → X ∗ is compact.

Proof. We show that A∗ D(Y ∗ ) = K ⊂ X ∗ is precompact. Let


f ∈ A∗ D(Y ∗ ).
Then f (x) = (A∗ φ)(x) for x ∈ D(X) and some φ ∈ D(Y ∗ );
that is, f (x) = (A∗ φ)(x) = φ(Ax).

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.3. Dual Operators 120

We have

kf kX ∗ = sup |f (x)| = sup |φ(Ax)| = sup|φ(y)|,


kxk≤1 kxk≤1 y∈T

where T = AD(X) is precompact in Y since A is compact.


Consider the collection of continuous functions on T, M = {φ ∈
D(Y ∗ )}.
For any φ ∈ D(Y ∗ ) we have |φ(y)| ≤ kAk, y ∈ T and |φ(y1 ) −
φ2 (y2 )| ≤ ky1 − y2 k <  for ky1 − y2 k < .
Thus by Arzela’s theorem, M is precompact.
Let {fn }∞ ∗ ∗
n=1 be a bounded sequence in A D(Y ). From the
above expression for kf kX ∗ , it follows that the corresponding
sequence of functions {φn }∞
n=1 is also bounded.

Hence there exists a subsequence φn0 such that


kφn0 − φn00 kC(T ) → 0 as n0 , n00 → ∞ which implies that
kfn0 − fn00 kX ∗ → 0 as n0 , n00 → ∞.
So, A ∈ K(X → Y ), where A is a compact operator, implies
that A∗ ∈ K(Y ∗ → X ∗ )(i.e, A∗ is also a compact operator).

Remark 5.3.3. In the case where X and Y are Hilbert spaces,


the above theorem may be proved without use of Arzela’s the-
orem.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.4. Finite Rank Operators 121

5.4 Finite Rank Operators

We call the number dim(Im A) the rank of the operator A and


we write rank A=dim(ImA).
We say that the operator A has finite rank if rank A < ∞.
For an example of an operator of finite rank, consider a finite
number of elements yi ∈ Y and fi ∈ X ∗ where X and Y are
Banach spaces and i = 1, 2, . . . , n for some n ∈ N. Define the
operator A by setting Ax = n1 fi (x)yi . This operator has rank
P

no greater than n and it is a bounded operator since kAk ≤


Pn
i kfi k.kyi k.

Indeed, we have kAxk = k n1 fi (x)yi k ≤


P Pn
1 kfi (x)yi k ≤
Pn Pn Pn
1 kfi (x)kkyi k ≤ 1 kfi kkxkkyi k = kxk 1 kfi kkyi k.

Let us compute the dual operator A∗ :


n
X n
X 
φ(Ax) = fi (x)φ(yi ) = φ(yi )fi (x)
1 1

So, A∗ φ = n1 φ(yi )fi . Thus the operator A may be written as


P
Pn
A(.) = 1 fi (.)yi ( where we put the “.” in the place where
the variable “x” should be applied) and its dual is A∗ (.) =
Pn ∗
1 (.)(yi )fi ( now “.” is in the position of the variable φ ∈ Y ).

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.4. Finite Rank Operators 122

Every bounded operator of finite rank is a compact opera-


tor because it maps the unit ball to a bounded set in a finite
dimensional subspace which is always precompact.

Example 5.4.1.
Compactness of the integral operator in L2

Consider K(t, τ ) ∈ L2 (I 2 ), where I = [0, 1], and let


s
Z 1 Z 1
C = kK(t, τ )kL2 (I2 ) = kK(t, τ )k2 dtdτ .
0 0

Define the operator K on L2 [0, 1] by


Z 1
Kx = K(t, τ )x(τ )dτ.
0

Then K is a compact operator in L2 [0, 1] and kKkop ≤ C =


kKkL2 (I 2 ) .
In order to prove this, we first recall that kKkop ≤ kKkL2 (I 2 )
(See Example 5.1.3.)
Let {φi (t)}∞
1 be an orthonormal basis in L2 ([0, 1]). From
Exercise 3.6, we see that {φi (t)φj (τ )} is an orthonormal basis

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.4. Finite Rank Operators 123

in L2 [I 2 ]; so
X
K(t, τ ) = aij φi (t)φj (τ ),
ij

where the series converges in the sense of L2 (I 2 ).


Let n X
n
X
Kn = aij φi (t)φj (τ ).
i=1 j=1

Then kKn − KkL2 (I 2 ) → 0 as n → ∞; therefore, considering the


operator Z 1
Kn x = Kn (t, τ )x(τ )dτ,
0

we also know that kKn − Kkop → 0 as n → ∞. But, Kn is an


operator of finite rank (rank Kn ≤ n) and this means that K is
approximable by finite rank operators, which are compact. This
implies that the limit operator K is also compact.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.5. Different Notions of Convergences in L(X) 124

5.5 Different Notions of Convergences


in L(X)

In this section, we introduce different types of convergence in the


space of operators L(X) := L(X → X), namely norm conver-
gence, also called uniform convergence, strong convergence and
weak convergence. The concept of weak convergence illustrates
a basic principle of functional analysis, namely, the fact that the
investigation of normed spaces is often related to that of their
dual spaces.
Norm convergence, or Uniform convergence, is de-
fined by saying that the sequence of operators An converges in
norm to the operator A (and we write An → A) if kAn −Ak → 0
as n → ∞.
If X is a Banach space, then the space L(X) is complete and
so if {An } is a Cauchy sequence with respect to the norm, then
it always converges to a bounded operator.
Another important notion is that of strong convergence:
We say An → A strongly if for all x ∈ X we have An x → Ax.
s
For strong convergence we write An → A.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.5. Different Notions of Convergences in L(X) 125

We note here that if the sequence {An } is Cauchy in the


strong sense, that is, for all x ∈ X the sequence An x is Cauchy
in X, then there exists A ∈ L(X) such that An → A strongly.
This fact is an immediate consequence of the Banach-Steinhaus
theorem, which is beyond the scope of our syllabus.
It is easy to see that strong convergence is weaker than
norm convergence. That is, if An → A ( converges in norm),
then for any x ∈ X, An x → Ax since

kAn x − Axk ≤ kxk.kAn − Ak → 0.

However the opposite implication is not true. Let us give an


example.
Consider the projections in L2 [0, 1] defined for n > 0 by

f (t), for t < 
n
Pn f =
0, for t ≥ n .

It is easy to see that Pn f → 0 as n → 0 for all f ∈ L2 [0, 1];


that is, Pn converges in the strong sense to zero. But kPn k = 1,
and consequently it does not converge in norm to zero.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.5. Different Notions of Convergences in L(X) 126

(To see kPn k = 1,, note that kPn f k ≤ kf k by its definition.


If we take g = Pn f , then kPn gk = kPn Pn f k = kPn f k = kgk,
for f ∈ L2 [0, 1], Pn f ∈ L2 [0, 1].)
Our third notion of convergence is that of weak conver-
gence. We say An converges weakly to A, and we write
w
An → A, if for all x ∈ X and for all f ∈ X ∗ we have f (An x) →
f (Ax).
Again, in order to distinguish this notion from that of strong
convergence, consider the example of the shift operator A in `2
(The shift operator in `2 defined by Ax = (0, a1 , . . . , an , . . . ) for
n times
(an ) ∈ `2 , also we have An (x) = (0, . . . , 0, a1 , a2 , . . . ) ).
For this operator one can easily see that f (An x) → 0 for
every f = ∞
P P∞
1 bi ei and x = 1 ai ei since


X
f (An x) = bi ai−n
i=n+1

and therefore

X ∞
1/2  X 1/2
n 2 2
|f (A (x))| ≤ |ai | |bi | → 0,
1 n+1

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.6. Invertible Operators 127

as n → ∞.
w
Thus, we have An → 0, but kAxk = kxk for every x and so
kAn xk = kxk. Hence, An does not converge strongly to zero.
(For any x 6= 0, kxk 6= 0. Hence kAn xk does not converge to
zero⇒ An x does not converge to zero. )

5.6 Invertible Operators

Let A ∈ L(X). We call B(:= A−1 ) the inverse operator of A


if and only if BA = Id and AB = Id, where Id is the identity
operator.
In the finite dimensional case, the notion of the determinant
implies that BA = Id is enough to deduce that A is invertible
and A−1 = B. The reason is that detA 6= 0 leads to a formula
for A−1 .
In the infinite dimensional case, though, this is not the case.
One may consider for example the case of the right shift operator
in `2 : for x = (xi )∞
1 let T x = (0, x1 , x2 , . . . ). Define the left shift
operator U : `2 7→ `2 by Ux = (x2 , x3 , . . . ). Then UT = Id but
TU = 6 Id and kerT U = 6 0.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.6. Invertible Operators 128

We now list some properties of invertible operators.

1. If A and B are invertible operators, then AB is also in-


vertible and its inverse satisfies (AB)−1 = B −1 A−1 .
((AB)(B −1 A−1 ) = Id = (B −1 A−1 )AB)

2. If kAk = q < 1, then (I − A) is invertible and (I − A)−1 =


P∞ k −1
0 A . Moreover, k(I − A) k ≤ 1/(1 − kAk). Indeed,
first note that kAkk ≤ kAkk ≤ q k → 0 as k → ∞. The
sequence Sn = n0 Ak is Cauchy.
P

kSn − Sm k = kAm+1 + . . . An k
= kAm+1 (I + A + . . . An−m−1 )k
≤ kAm+1 kk(I + A + . . . An−m−1 )k
≤ kAkm+1 k(I + A + . . . An−m−1 )k
kAkm+1
≤ → 0,
1 − kAk

since kAk < 1 ⇒ kAkm+1 → 0, and therefore Sn con-


verges.
Moreover, (I − A)Sn = (I − A)(I + A + · · · + An ) =
I − An+1 → I as n → ∞ and Sn (I − A) → I as n → ∞.
Thus, lim Sn is the inverse operator of (I − A)

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.6. Invertible Operators 129

3. Let A be an invertible operator and let B be such that


kA − Bk < 1/kA−1 k. Then B is also invertible. Indeed,
write B = A(I − A−1 (A − B)). The operator A is invert-
ible and I − A−1 (A − B) is invertible (kA−1 (A − B)k ≤
1
kA−1 kk(A − B)k < kA−1 k −1 = 1 and apply (2)).
kA k
Hence, their product is invertible by (1).

We conclude this Chapter by stating one of the main theo-


rems of functional analysis.
Theorem 5.6.1. (The Banach open map theorem) Let
X,Y be Banach spaces and let A : X → Y be a bounded linear
operator which is one-to-one ( i.e, ker A=0) and onto ( i.e., Im
A=Y). Then the formally defined inverse operator A−1 : Y → X
is bounded.

Recall that a map f from a metric space X to a metric space


Y is said to be open if for every open set E in X, its image f (E)
is open in Y.
Note that a map f : X → Y is continuous if and only if for
every open set E in Y, its inverse image f −1 (E) is open in X.
Thus, a bijective map f : X → Y is open if and only if
f −1 : Y → X is continuous.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

5.6. Invertible Operators 130

Therefore, the Banach open map theorem states that if X


and Y are Banach spaces, and if A : X → Y is a bijective,
bounded linear map, then A is an open map.

School of Distance Education,University of Calicut


MTH3C13 : FUNCTIONAL ANALYSIS

References

Text Books (As per Syllabus)

1. Yuli Eidelman, Vitali Milman & Antonis Tsolomitis: Func-


tional Analysis: An Introduction, Graduate Studies in Math-
ematics, Volume 66, American Mathematical Society, Prov-
idence, Rhode Island, 2004

Books For Further Reading

1. B.V. Limaye: Functional Analysis, New Age International


Ltd., New Delhi, 1996.

2. E. Kreyzig: Introductory Functional Analysis with Appli-


cations , John Wiley and Sons, 1978.

3. M. Thamban Nair: Functional Analysis: A First Course,


2nd Edition, PHI Learning Pvt. Ltd., 2021.
MTH3C13 : FUNCTIONAL ANALYSIS

SYLLABUS FOR MSc MATHEMATICS (CBCSS) PG PROGRAMME

EFFECTIVE FROM 2019 ADMISSION ONWARDS

SEMESTER 3

MTH3C13: FUNCTIONAL ANALYSIS


No. of Credits: 4
No. of hours of Lectures/week: 5

TEXT : YULI EIDELMAN, VITALI MILMAN & ANTONIS TSOLOMITIS; FUNC-


TIONAL ANALYSIS AN INTRODUCTION; AMS, Providence, Rhode Island, 2004

Module 1
Linear Spaces; normed spaces;first examples: Linear spaces, Normed spaces;first exaples,
Holder’s inequality, Minkowski’s ineaquality, Topological and geometric notions, Quotient
normed space, Completeness;completion. [Chapter 1 Sections 1.1- 1.5]

Module 2
Hilbert spaces: Basic notions; first examples, Cauchy- Schwartz ineaquality and Hilber-
tian norm,Bessels ineaquality,Completesystems, Gram-Schmidt orthogonalization proce-
dure, orthogonal bases, Parseval’ identity; Projection; orthgonal decompositions; Sepa-
rable case, The distance from a point to a convex set, Orthogonal decomposition; lin-
ear functionals; Linear functionals in a general linear space,Bounded linear functionals,
Bounded linear functionals in a Hilbert space, An example of a non separable Hilbert
space. [Chapter 2; Sections 2.1-2.3(omit Proposition 2.1. 15)]
Module 3
The dual space; The Hahn Banach Theorem and its first consequences, corollories of
the Hahn Banach theorem, Examples of dual spaces. Bounded linear Operators; Com-
pleteness of the space of bounded linear operators, Examples of linear operators, Compact
operators, Compact sets, The space of compact boperators, Dual operators ,Operators
of finite rank, Compactness of the integral operators in L2, Convergence in the space of
bounded operators, Invertible operators[ Chapter3; Sections 3.1 , 3.2; Chapter4; Sections
4.1- 4.7]

You might also like