IGNOU MSC MATHS FUNCTIONAL ANALYSIS BLOCK 3

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 73

MMT-006

FUNCTIONAL ANALYSIS
Indira Gandhi National Open
University School of Sciences

Block

3
Hilbert Spaces
Block Introduction 3
Notations and Symbols 4
UNIT 8
Geometry of Hilbert Spaces 5
UNIT 9
Orthonormal Sets 29
UNIT 10
Operators on Hilbert Spaces 55
Course Design Committee
Dr. A.K Vijayrajan, Faculty Members
Kerala School of Mathematics, Calicut School of Sciences, IGNOU
Prof. Parvin Sinclair
Prof. K. Parthasarathy, (Retd.) Prof. Poornima Mital
Ramanujan Institute, University of Madras, Chennai Prof. Sujatha Varma
Prof. Deepika
Prof. M.S. Balasubhramani , (Retd.) Dr. S. Venkataraman
University of Calicut, Kerala

Dr. Sachi Srivastava,


Delhi University, Delhi

Dr. Shravan Kumar,


IIT Delhi

Prof. V. Muruganandhan
NISER, Orissa

Block Preparation Team


Prof. K. Parthasarathy (Editor) Prof. Sujatha Varma
Ramanujan Institute, University of Madras, Chennai School of Sciences, IGNOU

Dr. N. Shravan Kumar,


Indian Institute of Technology, Delhi

Course Coordinator: Prof. Sujatha Varma

Acknowledgements: Mr. Santosh Kumar Pal for word processing the Manuscript and Mr. Surender
Singh Chauhan for the CRC.

Disclaimer – Any materials adapted from web-based resources in this course are being used for educational purposes
only and not for commercial purpose.

, 2023

© Indira Gandhi National Open University

ISBN-
All right reserved. No part of this work may be reproduced in any form, by mimeograph or any other means, without
permission in writing from the Indira Gandhi National Open University.
Further information on the Indira Gandhi National Open University courses may be obtained from the University’s
Office at Maidan Garhi, New Delhi-110 068.
Printed and published on behalf of the Indira Gandhi National Open University, New Delhi by the Director, School of
2 Sciences.
BLOCK 3 HILBERT SPACES
In Block 1 you were introduced to inner product spaces. You have also learnt how an
innerproduct induces a norm. An innerproduct space which is complete, under the norm
induced by the innerproduct, is called a Hilbert space.

Hilbert spaces are important in the mathematical formulation of quantum mechanics,


especially infinite dimensional Hilbert spaces. These spaces enhances the study how
particles are represented by vectors..

Hilbert spaces were named after David Hilbert, who studied them in the context of integral
equations. Van Neumann, was the mathematician, who most clearly recognized the
importance of these spaces as a result of his work on the foundation of quantum mechanics
begun with Hilbert and Lothar Nordheium. The name “Hilbert space” was soon adopted
by others. Among all the infinite dimensional vector spaces, the Hilbert space are the most
well-behaved. The innerproduct allows one to adopt “geometrical” view and use
geometrical language familiar from finite dimensional spaces. Another importance of
Hilbert spaces is that every Hilbert space has an orthonormal basis and every element of
the Hilbert space can be written in a unique way as a sum of multiples of these basis
elements.

This Block is divided into 3 units. In Unit 1 we discuss Hilbert spaces which are
innerproduct spaces that are complete. These spaces are similar to two and three
dimensional Euclidean spaces. Here we discuss two important theorems – Riesz
Representation Theorem and Projection Theorem.

Unit 2 deals with the notion of orthonormal basis in Hilbert spaces. They are also known as
complete orthonormal basis. The importance is that the basic elements in the basis allow
the spaces to be decomposed into linear combination of these basic elements. This makes
computations in these simpler. Using the orthonormality property, many important
properties are established in Hilbert spaces. One such property is Bessel’s inequality which
has the geometric meaning that the orthogonal projection of an element f on the linear
span of its basis elements has a norm which does not exceed the norm of f i ".e. f ". This is
similar to saying that in a right angled triangle hypotenuse is not shorter than any of the
otherside. Another important result is Parseval’s formula which has lot of applications in
physical sciences such as transformation energy of a signal from time domain to frequency
domain. Another important theorem discussed in this unit is Riesz-Fischer theorem.

Unit 3 deals with bounded linear maps from a Hilbert space to itself. We discuss self-
adjoint, normal and unitary operators. These operators have essential role in quantum
mechanics, particularly to quantum states.

3
NOTATIONS AND SYMBOLS (used in Block 3)

, Inner product
H Hilbert space
BL( X ) Space of bounded linear maps from X to itself
BL( X , Y ) Space of bounded linear maps from X to Y
C( X ) Space of bounded continuous scalar-valued functions on X
c Space of convergent sequences of scalars
c0 Subspace of c consisting of sequence converging to 0.
c00 Subspace of c consisting of sequences having at most a finite number of
non-zero terms
d ( x, y ) Distance between x and y
T Norm of linear map T
Lp Lebesgue space of p -integrable scalar-valued functions
L∞ Lebesgue space of essentially bounded measurable function
lp Space of p -summable scalar sequences
l∞ Space of bounded scalar sequences
Ea Annihilartor of a set E
E⊥ Orthogonal complement of a set E
B X (0,1) Open unit ball
B X ( a, r ) Open ball with centre a and radius r
X′ Dual space of X
X ′′ Dual of X ′
~
X Completion of X
x p p -norm of x
x∞ Sup norm of x
X ×Y Cartesian product of X and Y
X /Y Quotient space
x +Y Quotient norm of x + Y

4
UNIT 8

GEOMETRY OF HILBERT SPACES

Structure Page No
8.1 Introduction
Objectives
8.2 Complete Inner Product Spaces
8.3 Orthogonality
8.4 Existence of Best Approximation
8.5 Projection Theorem
8.6 Riesz Representation Theorem
8.7 Summary
8.8 Hints/Solutions

8.1 INTRODUCTION
In this unit, we provide a geometric structure to a linear space. This leads us to
Hilbert spaces. Hilbert spaces are a special kind of Banach spaces, those
which possess an additional structure in which we can tell when two elements
are orthogonal (i.e. perpendicular).

In Sec. 8.2 we familiarize you with the topological properties of Hilbert space.
Sec. 8.3 deals with geometrical properties of Hilbert spaces viz orthogonality.
You will see that the geometry is very much like the familiar two-and three
dimensional Euclidean geometry. We shall explain the orthogonality
properties. In Sec. 8.4, 8.5 and 8.6 we state and prove three important
theorems known as existence of best approximation, Projection Theorem and
Riesz representation theorem.

Objectives
The objectives of this unit are to
• Define innerproduct spaces and identify inner product spaces which are
complete and are called Hilbert spaces,
• Establish the fact that the norm on a normed linear space is induced by an
inner product if and only if the norm satisfies the parallelogram identity,
• Explain the existence of best approximation in a Hilbert space,
• Familiarize you with orthogonality property for inner product space,
• Explain the orthogonality of elements in a Hilbert space which helps in
extending the property of right angles of 2D and 3D Euclidean spaces to
infinite dimensional Hilbert spaces,
Block 3 Hilbert Spaces
• State and prove the projection theorem,
• State and prove the Riesz representation theorem which characterizes the
dual space of a Hilbert space.

8.2 COMPLETE INNER PRODUCT SPACES


In this section we study an inner product which is a new structure on a linear
space. You may recall that we have introduced you to inner product spaces in
Block 1. There we have shown that this new structure defines a norm and this
helps to extend the concept of length to infinite dimensional spaces.

Let us start with some basic notions of inner product spaces. We denote these
spaces by H .

Definition 1: Let H be a vector space over K. An inner product on H is a


map , : H × H → K satisfying the following conditions: for all x, y, z ∈ H and
α, β ∈ K ,
a) x, x ≥ 0 and is 0 iff x = 0.
b) x, y = y, x , the complex conjugate if K = C. If K = R, then
x, y = y, x .
c) αx + β y, z = α x, z + β y , z .

The pair (H , , ) is called an inner product space.

Here are some examples.


n
Example 1: H = C n , then < z , w >=  z w is an inner product. If H = R
i =1
i i
n
, then
n
the dot product x. y = x y
i =1
i i is an inner product.

Example 2: Let H = l 2 . Define an inner product on H , given by, for z, w ∈ l 2 ,

< z , w >=  z k wk … (1)


k =N

You may note that the infinite series in (1) defining the inner product < z , w > is
absolutely convergent and hence convergent. This is because using the
Cauchy-Schwaz inequality, we can show that the sequence of partial sums of
 z k wk is convergent. Now it is not difficult to verify the three conditions
k ∈N

in Definition 1. Hence l 2 is an inner product space.

Example 3: Let M (n, C) be the complex vector space of all n × n matrices with
entries in C. For X ∈ M (n, C ), let X * denote the complex conjugate of the
transpose of X . That is, if X = ( xij ), then X * = ( x j i ), where x j i stands for the
n
complex conjugate of x ji . Define X , Y := tr ( XY *) where tr ( A) = a
i =1
ii for any

A = ( aij ) ∈ M ( n, C). Then (M ( n, C), , ) is an inner product space.


6
Unit 8 Geometry of Hilbert Spaces
2 2
Example 4: H = L (R) The inner product f , g ∈ L is defined for the inner
product ( f , g ) is given by

( f , g ) =  f g dx … (2)
X

It can be shown that the integral in (2) exists by applying Holder’s inequality.
We have already studied that an inner product defines a norm given by

2
x = x, x

This norm is said to be induced by the inner product. Also we have the
Cauchy-Schwarz inequality

x, y ≤ x y for all x, y ∈ H .

All topological notions on an inner product space are with respect to the
induced norm.

You can easily verify the following propositions.

Proposition 1: Let {xn } be a sequence in an inner product space X . Let us


show that if xn , x → x, x and xn → x , then xn → x.

Proof: Since

2
xn − x = xn − x, xn − x
= xn , x n − x n x − x , xn + x , x ,
2 2
xn , xn = xn → x , x, xn = xn , x → x, x ,

We see that

2 2
xn − x → x − x, x − x, x + x, x = 0.

That is, xn → x.

Proposition 2: Let X be an inner product space and let x, y ∈ X . Suppose


(α n ) is a sequence of scalars converging to a scalar α prove that
α n x, y → αx, y in X .

( )
Proposition 3: If xn → x and y n → y in H , , , then xn , yn → x, y .

The proofs of the above propositions 2 and 3 are left as exercises to do.

It is possible to construct more Hilbert spaces using the known Hilbert spaces.
For instance, it is possible to construct a product Hilbert space from a
sequence of Hilbert spaces in the same way as l 2 is constructed from the
sequence K, K,... . The quotient of a Hilbert space by one of its closed
subspaces is again a Hilbert space. The proof is left it as a exercise for you to
do. 7
Block 3 Hilbert Spaces
You may also note that a subspace of a Hilbert Space is Hilbert space under
the induced inner product if and only if the subspace is closed. (It is useful to
recall a similar result for complete metric spaces)

Here is another interesting result.

Example 5: Show that the spaces Lp [0,1] 1 ≤ p ≤ ∞ is an inner product space


if and only p = 2.

Solution: To observe that let us take

f (t ) = t , g (t ) = 1 − t ∀ t ∈ [0,1],

We have f + g p
= 1 and

1 /(1 + p )1/ p if 1 ≤ p < ∞


f p
= g p
= f −g p
=
1 if p = ∞

From these relations it follows that

f +g
2
p
+ f −g
2
p
=2 f( 2
p
+ g
2
p
) ⇔ p = 2.
Thus, ⋅ p
is not induced by an inner product if p ≠ 2.
***

You can now try the following exercises.

E1) Prove Proposition 2.

E2) Prove Proposition 3.

E3) Prove that the quotient of a Hilbert space by one of its closed
subspaces is again a Hilbert space.

Next we shall consider a notion special to inner product spaces.

8.3 ORTHOGONALITY
In this section we discuss orthongonality property of inner product spaces.

We shall begin with a definition.

Definition 2: Let X be an inner product space, and x, y ∈ X . Then x and y in


X are said to be orthogonal if ( x, y ) = 0.

The following identity, called Pythagoras Theorem, due to its geometric


implication, can be easily deduced using the definition of an inner product.

Theorem 1 (Pythagoras theorem): Let X be an inner product space, and x


and y be orthogonal in X . Then
8
Unit 8 Geometry of Hilbert Spaces
2 2 2
x+ y = x + y .

It is easily seen that, if the scalar field is R, then the converse of the
Pythagoras theorem also holds. However, the converse need not be true if the
scalar field is C. A simple example shows this: Let X = C with standard inner
product, and for nonzero real numbers α, β ∈ R, let x = α, y = iβ. Then we
have

2 2 2 2 2 2
x + y = α + iβ = α + β = x + y ,

but ( x, y ) = −i αβ ≠ 0.

We shall give another definition.

Definition 3: A subset S of X is said to be an orthogonal set if every pair of


distinct elements in S are orthogonal.

An orthogonal set S is said to be an orthonormal set if x = 1 for every x ∈ S .


Thus, if E is an orthogonal set which does not contain the zero element, then
the set
~
E = {x / x : x ∈ E }

is an orthonormal set.

For a subset S of an inner product space X , we write

S ⊥ = {x ∈ X : x, u = 0 ∀ u ∈ S}.

The set S ⊥ is called called the orthogonal complement of S and is denoted


by S ⊥ .

We may note that

X ⊥ = {0}, {0}⊥ = X .

The following proposition follows from the definition.

Proposition 4: Let X be an inner product space and S is a subset. Then

i) S ∩ S ⊥ ⊂ {0} . Also S ∩ S ⊥ = {0} if S is a subspace.


ii) S ⊂ S ⊥⊥
iii) For any inner product space H and any subsets S1 and S 2 such that
S1 ⊂ S 2 , we have S1⊥ ⊃ S 2⊥ .

The proof of the following proposition is very easy and is left as an exercise for
you to do.

Proposition 5: Let X be an inner product space and S ⊆ X . Then

i) S ⊥ is closed subspace of X ,
9
Block 3 Hilbert Spaces

ii) S⊥ =S ,

iii) If S is dense in X , then S ⊥ = {0},


iv) If S is an orthogonal set and 0 ∉ S , then S linearly independent.

We conclude by leaving some exercises for you to try.

E4) Prove (i), (ii), (iii) and (iv) of Proposition 5.

Next we shall discuss an important geometric property of an inner product


space.

8.4 EXISTENCE OF BEST APPROXIMATION


In this section we shall discuss a geometric property known as the “Best
approximation property”. We consider the following question. What best can
be done if we want to come close to a given element of an inner product space
H and we remain in a given subset of H ? To study this, we introduce the
following notion.

Definition 4: Let H be an inner product space and E be a subset of H . Given


an element x of H , an element y of E is said to be a best approximation from
E to x if x − y ≤ x − z for all z ∈ E , that is, x − y = dist ( x, E ).

We address three basic questions in this regard.


1. Does a best approximation from E to x exist for any x ?
2. Can there be more than one best approximations from E to x ?
3. How can we find a best approximation from E to x ?

To begin with we consider the following result.

Theorem 2: Let H be an inner product space.

a) Let E ⊂ H and x ∈ E. Then there exists a best approximation from E to x


if and only if x ∈ E.
b) If E ⊂ H is convex, then there exists at most one best approximation from
E to any x ∈ H .
c) Let F be a subspace of H and x ∈ H . Then y ∈ F is a best approximation
from F to x if and only if x − y ⊥ F and in that case

1/ 2
dist ( x, F ) = x, x − y

Proof: a) If x ∈ E , then clearly x is a best approximation to itself from E.


Conversely, let y ∈ E be a best approximation from E to x. Then
x − y = dist ( x, E ) = 0, that is, x = y. Hence x ∈ E.

b) Let y1 and y2 be best approximations from a convex subset E to H to


10 x ∈ H . By the parallelogram law (21.2(b)) for x − y1 and x − y2 , we have
Unit 8 Geometry of Hilbert Spaces
2 2 2 2
2 x − y1 + 2 x − y2 = 2 x − y1 − y2 + y2 − y1 .

Let d = dist ( x, E ). Then x − y1 = d = x − y 2 . Also, since


( y1 + y2 ) / 2 ∈ E , we have x − ( y1 + y2 ) / 2 ≥ d . Hence

2
0 ≤ y 2 − y1 ≤ 2d 2 + 2d 2 − 4d 2 = 0,

that is, y2 = y1 , showing that there is at most one best approximation from
E to x.

c) Let y ∈ F such that x − y ⊥ F . Consider x ∈ F . Since F is a subspace,


y − z ∈ F , so that x − y ⊥ y − z. By the Pythagoras theorem (22.1(a)),

2 2 2 2
x − z = ( x − y) + ( y − z ) = x − y + y − z .

Hence x − y ≤ x − z for all z ∈ F , that is, y is a best approximation from


F to x.

Conversely, let y ∈ F be a best approximation from F to x. To show that


x − y ⊥ F , we may consider z ∈ F with z = 1. Then w = y + x − y, z z
belongs to F , since F is a subspace. Hence

2 2
x − y ≤ x − w = x − w, x − w
2
= x − y, x − y − x − y , z z, z .
2 2
= x − y − x − y, z .

Thus x − y , z = 0, that is, x − y ⊥ z. Further, since y ⊥ x − y, we have

2
[dist ( x, F )]2 = x − y = x − y, x − y = x, x − y .

The result Theorem 2 (a) shows that if a subset E of H is not closed, then
there is some x in H such that there is no best approximation from E to x,
while 2 (b) shows that if E is convex, then there cannot be more than one
best approximations from E to any x ∈ H . We shall show later that if E is a
nonempty closed convex subset of a complete inner product space H , then
for each x ∈ H , there is a unique best approximation from E to x.

x–y

F
y
Fig. 1 11
Block 3 Hilbert Spaces
While the result in Theorem 2 (c) does not guarantee the existence of a best
approximation from a subspace F of an inner product space H to x ∈ H , it
does give a clue for finding it. It tells us to look for some y ∈ F such that
x − y ⊥ F . The elements x, y, x − y and the subspace F are schematically
shown in Fig. 1. In fact y is obtained by dropping a perpendicular from x to
F.

We shall now show that if F is a finite dimensional subspace of H , then the


best approximations from F always exist and they can be written down
explicitly.

Theorem 3: Let H be an inner product space, {x1 ,..., xm } be a linearly


independent subset of H and x ∈ H . Then the following holds.

a) Let F = span{x1 ,..., xm }. Then the unique best approximation from F to x


is given by

y = k1 x1 + ... + k m xm ,

where k1 ,..., k m form the unique solution of the normal equations:

x1 , x1 k1 + ... + xm , x1 k m = x, x1
M M M
x1 , xm k1 + ... + xm , xm k m = x, xm

Also,

1/ 2
dist ( x, F ) = x, x − k1 x1 − ... − k m xm

b) Let c1 ,...., cm be scalars and

E = { y ∈ X : y , x j = c j for j = 1,..., m}.

Then the unique best approximation from E to x is given by

y = x + k1 x1 + ... + k m xm ,

where k1 ,..., k m form the unique solution of the equations

x1 , x1 k1 + ... + xm , x1 k m = c1 − x, x1
M M M
x1 , xm k1 + ... + xm , xm k m = cm − x, xm

Also,

dist ( x, E ) = k1 x1 + ... + k m xm

The proof of the theorem is omitted.


12
Unit 8 Geometry of Hilbert Spaces
We now prove the existence and the uniqueness of a best approximation from
a nonempty closed convex subset of a Hilbert space.

Theorem 4: Let E be a nonempty closed convex subset of a Hilbert space


H . Then for each x ∈ H , there exists a unique best approximation from E to
x. In particular, there is a unique element in E of minimal norm.

Proof: Let x ∈ H and d = dist ( x, E ). Then there is a sequence ( yn ) in E


such that x − yn → d . Applying the parallelogram law to x − yn and x − ym ,
we obtain

2 2 2 2
2 x − yn + 2 x − ym = 2 x − yn − ym + y m − y n

for all m, n = 1, 2,... . Since E is convex, ( yn + ym ) / 2 belongs to E and


x − ( yn + ym ) / 2 ≥ d . Hence

2 2 2
ym − yn ≤ 2 x − yn + 2 x − ym − 4d 2 → 2d 2 + 2d 2 − 4d 2 = 0.

This shows that ( yn ) is a Cauchy sequence in E. Since H is complete and


E is closed in H , we see that y n → y in E. It is clear that

x − y = lim x − yn = d .
n→∞

Thus y is a best approximation from E to x. The uniqueness of y follows


from Theorem 3 (b), the proof being very similar to the argument given above.

The unique best approximation from E to 0 is the unique element in E of


minimal norm.

Remark 1: If E is, in fact, a closed subspace of H , then Theorem 2(c) shows


that the best approximation from E to x is given by the unique y ∈ E for
which x − y ⊥ E .

Why don’t you try some exercise.

E5) Let F be a subspace of X and G be a subspace of F . Let x ∈ X . If y is


a best approximation from F to x and z is a best approximation from G
to y, then z is a best approximation from G to x.

Next we shall discuss a theorem which gives the best approximation of


elements of a Hilbert from a closed subspace.

8.5 PROJECTION THEOREM


Now we discuss Projection Theorem which is a generalization of Pythagoras
theorem. You might have learnt the notion of direct sum (internal/external) of
vector spaces in Linear Algebra. We briefly review the notion here.
13
Block 3 Hilbert Spaces
Let X be a vector space, and V , W be linear subspaces of X . We say that
X is a direct sum of V and W and denote it by X = V ⊕ W , if any vector
x ∈ X can be written uniquely as x = v + w where v ∈ V and w ∈ W .

We then have a linear map P : X → X defined by Px = v if x = v + w.

Note that P 2 = P and range of P , R ( P ) = V and zero space, Z ( P ) = W . Note


that R ( P) = Z ( I − P ). Conversely, if P ∈ L( X ) is such that P 2 = P, then
X = R( P) ⊕ Z ( P). All these are easy to verify.

We now deal with analogous concepts in the context of inner product spaces.

Definition 5: Let V be a closed subspace of an inner product space H . We


say that V has a complement in H if there exists a closed subspace W such
that V ∩ W = (0) and H = V + W . In such a case, we also say that V is a
direct summand of H or V admits a topological complement in H . We also
say H has complemented subspace property if every closed subspace has
a topological complement.

Example 6: Let V be a finite dimensional subspace of H then show that V


admits a complement.

Solution: Let {v1 ,...., vn } be a basis of V . Let f j ,1 ≤ j ≤ n be continuous linear


functional on V , defined by f i (vi ) = δ ji . Note that each f j is continuous on V .
By Hahn-Banach theorem there exist g j ∈ X which extend f j . Given x ∈ X ,
let v := g
j
j ( x) v j and w := x − v. Note that w ∈ ∩ j Z ( g j ). If we let

W := ∩ j Z ( g j ), then X = V ⊕ W .

Another way is to see that if P ( x) := gj


j ( x ) v j , then P is continuous linear

and P 2 = P. A linear map P such that P 2 = P is called a projection.


***

What about infinite dimensional spaces? We will show that any closed vector
subspace of a Hilbert space has a complement.

Theorem 5 (Projection Theorem): If F is a closed subspace of the Hilbert



space H , we have the direct sum decomposition of H , H = F ⊕ F .
Equivalently there is an orthogonal projection onto F . Moreover we have
F ⊥⊥ = F .

Proof: Since it is clear that F ∩ F ⊥ = {0}, we need only check that


H = F + F ⊥ . To this end, consider

N = F + F⊥

It is clear that N is a closed subspace of H ; further, we have


14
Unit 8 Geometry of Hilbert Spaces

F ⊂ N and F ⊂ N ,

which implies that

N ⊥ ⊂ F ⊥ and N ⊥ ⊂ F ⊥⊥
or

N ⊥ ⊂ F ⊥ ∩ F ⊥⊥ = {0}

Therefore,

N ⊥ = {0}
and

N = N ⊥⊥ = {0}⊥ = H .

Hence the result.

Finally, we prove that F ⊥⊥ = F . Let x ∈ F . For every z ∈ F ⊥ , we have


( x, z ) = 0. Hence that x ∈ ( F ⊥ ) ⊥ = F ⊥⊥ . Conversely, let x ∈ F ⊥⊥ . Then
x = y + z for some y ∈ F and z ∈ F ⊥ . By what we have just seen, y ∈ F ⊥⊥ .
⊥⊥
Thus z = x − y ∈ F , so that z ∈ F ⊥ ∩ F ⊥⊥ . Hence z = 0, that is, x = y ∈ F .
⊥⊥
Therefore F = F .

The projection theorem shows that every Hilbert space H has the
complemented subspace property, that is, for every closed subspace F of
H , there is a closed subspace G of H such that F + G = H and
F ∩ G = {0}. In fact, we can let G = F ⊥ . The closed subspace F ⊥ is called
the orthogonal complement of the closed subspace F . The complemented
subspace property characterizes Hilbert spaces among all complete normed
linear spaces.

Remark 2: Note that in Example 6 we have shown that if F is a finite-


dimensional subspace of any inner product space, such a decomposition of
the space is possible.

We shall now prove some theorems which are consequences of Projection


Theorem.

Theorem 6: Let F be a closed subspace of H and a ∈ H , a ∉ F . Then there


is a unique f ∈ H ′ such that f (a) = d (a, F ), f ( x ) = 0 for all x ∈ F and f = 1.

Proof: By the Projection theorem, there exist b, c such that

a = b + c, b ∈ F, c ∈ F ⊥.
15
Block 3 Hilbert Spaces

Since a − b = c ∈ F , we see that b is the best approximation to a from F .
Further, c ≠ 0 since a ∉ F . Let

c
y= and f ( x) = x, y , x∈H
c

Then, f ∈ H ′ and f = y = 1. If x ∈ F , then x ∈ H , and hence

c
f ( x ) = x, =0
c

Thus, f ( x ) = 0 for all x ∈ H , further,

c 1 1
f ( x) = a, = b + c, c = c, c = c .
c c c

Since d ( a , F ) = a − b = c , we see that f (a) = d (a, F ). The completes the


proof.

Theorem 7: Let Y be a closed subspace of a Hilbert space X . Then X / Y is


linearly isometric to Y ⊥ .

Proof: Let Y ⊥ = Z . Define T : Z → X / Y by T ( z ) = z + Y , z ∈ Z.

The relation

k ( z1 + Y ) + ( z 2 + Y ) = kz1 + z 2 + Y

states that T is linear. If z ∈ Z and T ( z ) is the zero element of X / Y , then


z ∈ Y , and hence z = 0 since Y ∩ Z = Y ∩ Y ⊥ = {0}. This shows that T is
one-one. Let x + Y be any element of X / Y . Since

x = y + z, y ∈Y , z ∈Y ⊥ = Z,

we have

x + Y = y + z + Y = z + Y = T ( z ).

Thus T is onto.

Let z ∈ Z . For all y ∈ Y , we have y, z = 0, and hence

2 2 2 2
y+z = y + z ≥ z .

There

inf { z + y : y ∈ Y } = z ,
16
Unit 8 Geometry of Hilbert Spaces
that is, z + Y = z . This means that T is an isometry. Thus we get a linear
isometry T between Y ⊥ and X / Y . This proves the result.

You can try some exercises now.

E6) Let X be an inner product space. Then show that the projection
theorem may not hold if X is not complete.

E7) Suppose that F1 and F2 are closed subspace of a Hilbert space H .


Show that

i) ( F1 ∩ F2 ) ⊥ equals the closure of F1⊥ + F2⊥ ;

ii) If, F1 ⊥ F2 , then F1 + F2 is closed subspace. Show also that


F1 + F2 , need not be closed even though F1 and F2 are closed
subspaces without the condition F1 ⊥ F2 .

Next we shall give a characterization of continuous linear functionals of inner


product spaces.

8.6 RIESZ REPRESENTATION THEOREM

You were introduced to bounded (continuous) linear functionals on normed


spaces in Unit 7, Sec. 7.2. There we have shown that any linear functional
over a finite-dimensional space was represented by some element in the
space itself. This allows us to think of the same result for infinite dimensional
inner product spaces. Now in this section we discuss the linear functionals on
inner product spaces. Mainly we wish to discuss the relationship between
inner product spaces and continuous linear functionals on them. It will be seen F. Riesz
that every continuous linear functional on a Hilbert space (completeness of the
space is important) has a simple representation, so simple that you can
literally lay your hands on them. This result was established by the Hungarian
mathematician F. Riesz in 1907 and therefore it is called Riesz Representation
Theorem.

We recall the definition of continuous linear functional and make note of some
of the properties of continuous linear functionals defined on inner product
space.

Let X be an inner product space over K and f be a linear functional on X .


Then we observe the following.

1) Let f be continuous on X . Then f is continuous at 0 and f (0) = 0.


Hence there is some δ > 0 such that f ( x ) ≤ 1 whenever x ∈ X and
x ≤ δ. The linearity of f shows that f ( x) ≤ x / δ for all x ∈ X .

17
Block 3 Hilbert Spaces
Conversely, if f ( x) ≤ α x for some α > 0 and all x ∈ X , then for all x
and y in X , we have

f ( x) − f ( y) = f ( x − y) ≤ α x − y ,

showing that f is (uniformly) continuous on X . That means continuity


and uniform continuity are the same for linear functionals.

2) The set X ′ of all continuous linear functionals on X ′ is called the dual of


X . As you have observed for a normed linear space X ′ is a linear space
under pointwise addition and scalar multiplication. For f ∈ X ′, let

f = sup { f ( x) : x ∈ X , x ≤ 1}.

Then is a norm on X ′. We shall later show that there is an inner


′ ′ 2
product , on X ′ such that f, f = f for all f ∈ X ′.

We have the basic inequality

f ( x) ≤ f x, f ∈ X ′, x ∈ X .

It implies that if X ≠ {0}, then

f = sup { f ( x) : x ∈ X , x = 1}.

Let us now consider some examples of continuous linear functionals on X .


Fix y ∈ X and define

f ( x ) = x, y , x ∈ X .

The linearity and the continuity of the inner product in the first variable imply
that f ∈ X ′ . Let us calculate f . By the Schwarz inequality, we have

f ( x) ≤ x y , x∈ X.

Hence f ≤ y . If y = 0, then f = 0 and f = 0 = y . If y ≠ 0 , let x = y / y .


Then x = 1 and

y
f ( x) = ,y = y ,
y

showing that f = y .
18
Unit 8 Geometry of Hilbert Spaces
We shall now show that if X is complete, then this procedure gives all the
continuous linear functional on X . We give two proofs. The first is based on
the projection theorem 8, while the second uses the existence of an
orthonormal basis for X along with the Fourier expansion.

Theorem 8 (Riesz representation theorem (F. Riesz and M. Frechet,


1907)): Let H be a Hilbert space and f ∈ H ′. Then there is a unique y ∈ H
such that

f ( x ) = x, y , x ∈ H.

In fact, if z is a nonzero element of H such that z ⊥ Z ( f ), then


M. Frechet
f ( z) z (1878 – 1973)
y= .
z, z

Proof: If f = 0, then let y = 0, so that for all x ∈ H , we have

f ( x) = 0 = x,0 .

Let f ≠ 0. Since Z ( f ) is a closed subspace of H , the projection Theorem 8



shows that H = Z ( f ) + Z ( f ) . As Z ( f ) ≠ H , consider a nonzero element z
in Z ( f ) ⊥ . Let x ∈ H , Z ( f ) is a hyperspace in H . That means

x = w + kz

for some w ∈ Z ( f ) and k ∈ K . Then

x, z = w, z + k z , z = k z , z ,

so that k = x, z / z , z . Hence

x, z f ( z)
f ( x) = f (w) + k f ( z ) = k f ( z ) = f ( z ) = x, z .
z, z z, z

Thus we let y = f ( z ) z / z , z . Hence the result.

There is an alternate proof of the theorem using another theorem, the Riesz-
Fischer Theorem, which we will be discussing in the next unit.

Remark 3: For f ∈ H ′, the unique element y ∈ H such that f ( x) = x, y for


all x ∈ X is known as the representer of f . It satisfies y = f , as we have
already seen.

Let us see some examples. 19


Block 3 Hilbert Spaces
n
Example 7: Let us consider H = K .

The Riesz representation theorem for K n simply states the familiar fact that
every (continuous) linear functional f on K n is given by

f ( x) = k1 x (1) + ... + k n x(n), x ∈ Kn,

for fixed k1 ,..., k n in K. For if y is the representer of f , then we can let

k j = y ( j ), j = 1,..., n.

Example 8: Let us consider H = l 2 .

Every continuous linear functional f on l 2 is given by

f ( x) = k1 x(1) + k 2 x(2) + ..., x ∈ l2 ,

for a fixed square summable sequence (k n ) in K . For again if y is the


representer of f , then we can let k j = y ( j ), j = 1, 2,...

2
Example 9: Let us consider H = L ( E ) where E be a measurable subset of
R. Every continuous linear functional f on L2 ( E ) is given by

f ( x) =  k (t ) x(t ) dm(t ), x ∈ L2 ( E ),
E

for fixed square-integrable function k on E. For again if y is the representer of


f , then we can let k (t ) = y (t ) for t ∈ E.

The Riesz representation theorem does not hold for incomplete inner product
spaces as the following example shows.

Example 10: Show that the Riesz representation theorem does not hold for
X = c00 , the linear space of all scalar sequences having only a finite number
of nonzero entries with the inner product induced by l 2 .

Solution: Define f : X → K by


xn
f ( x) =  , x∈ X.
n =1 n

Then f is linear. By Holder’s inequality

 ∞ 1  ∞ 2 π2
f ( x ) ≤   2    x (n)  =
2 2
x
 n =1 n   n =1  6
20
Unit 8 Geometry of Hilbert Spaces
Hence f is continuous and f ≤ π / 6 .

Since f ≠ 0, F = Z ( f ) is a proper closed subspace of X . Let z ∈ F ⊥ . Since


z ∈ c00 , we see that z = ( z (1),..., z (m),0,0,...) for some positive integer m. For
1 ≤ n ≤ m, define xn ∈ X by letting xn (n) = 1, xn (m + 1) = −(m + 1) / n and
xn ( j ) = 0 for all j ≠ n or m + 1. Then f ( xn ) = 1 / n − (m + 1) / n(m + 1) = 0, so

that xn ∈ F and z (n) = z , xn = 0. Thus F = {0}. This shows that
F + F ⊥ ≠ X , that is, the projection theorem does not hold for X .

Also, f ∈ X ′ has no representer in X . To see this, let y ∈ X . Were


f ( x) = x, y for all x ∈ X , then by letting en = (0,...,0,1,0,0,...), where 1 occurs
only in the nth entry, we see that y (n) = en , y = f (en ) = 1 / n ≠ 0 for all
n = 1, 2,... . But this is impossible since y ∈ c00 . This shows that the Riesz
representation theorem does not hold for X .
***

As a consequence of the Riesz representation theorem we have the following


result.

Proposition 6: If H is a Hilbert space, then H is reflexive.

Proof: Let f ∈ H ′′. Since H ′ is a Hilbert space with inner product


x′, y′ = Ty′, Tx′ , where Tx′ ∈ H is the representer of x′. By the Riesz
representation theorem applied to H ′ there is a unique y ′ ∈ H ′ such that
y ′ = f and

f ( x′) = x′, y ′ for all x′ ∈ H ′.

Let T ( y′) = y be the representer of y ′. Then for all x′ ∈ H ′,

f ( x′) = x′, y ′ = T ( y′), T ′( x′) = y, T ( x′) = x′( y ) = ( Jy) ( x′),

where J : H → H ′′ is the canonical mapping. Thus, f = J ( y ). This means


that H is reflexive. This completes the proof.

Here are some exercises for you.

E8) Let H = R 3 . Define f ∈ H ′ by

f ( x) = x1 + x2 − x3 , x = ( x1 , x2 , x3 )

Find the representer of f . 21


Block 3 Hilbert Spaces
E9) Let E be a non-empty subset of a Hilbert space H . Show that
E a = {x′ ∈ H ′ : T ( x′) ∈ E ⊥ }, where T ( x′) is the representer of x′. Using
the same notation for annihilator as defined for normed linear spaces
and E ⊥ , orthogonal complement of E.

With this we come to an end of this unit.

8.7 SUMMARY

In this unit we have covered the following points.

1. Explained inner product spaces and Hilbert spaces.

2. Familiarized with the orthogonality property of inner product spaces

3. Stated and proved the following theorems

i) Existence of best approximation.


ii) Projection theorem
iii) Riesz-Representation Theorem.

8.8 SOLUTIONS AND ANSWERS

E1) We have to prove that (α n x, y ) → (αx, y ) where α n → α. That is we


have to show that

(α n x, y ) − (αx, y ) → 0 as n → ∞.

By Cauchy-Schwarz inequality

((α n − α) x, y ≤ (α n − α) x y → 0 as n → ∞ since
α n → α as n → ∞ .

E2) ( xn , yn ) − ( x, y ) = ( xn , yn ) − ( xn , y ) + ( xn , y ) − ( x, y )
= ( xn , yn − y + ( xn − x) y
≤ xn yn − y + ( xn − x ) y (by Cauchy Schwarz inequality)

Since {xn } is convergent { xn } is bounded and their exists on M


such that xn ≤ M ∀ n ∈ N and xn − x → 0 as n → ∞. Similarly
since { yn } is convergent yn − y → 0 as n → ∞. Hence

( xn , yn ) − ( x, y ) → 0

22 Therefore ( xn , yn ) → ( x, y ) .
Unit 8 Geometry of Hilbert Spaces
E3) Let x1 + Y and x2 + Y be any two elements of X / Y . There exist
sequences { yn } and {z n } in Y such that

x1 + yn → x1 + Y , x2 + z n → x2 + Y

as n → ∞. Since yn ± z n ∈ Y , we have

x1 + x2 + Y ≤ x1 + x2 + y n + z n ,
x1 − x2 + Y ≤ x1 − x2 + yn − z n .

Hence,
2 2 2
x1 + x2 + Y + x1 − x2 + Y ≤ ( x1 + y n ) + ( x2 + z n )
2
+ ( x1 + yn ) − ( x2 + z n )
2 2
= 2 x1 + yn + 2 x2 + z n

by the parallelogram law in X . Letting n → ∞, we get

2 2 2 2
x1 + x2 + Y + x1 − x2 + Y ≤ 2 x1 + Y + x2 + Y (3)

Next let {un }, {vn } be sequences in Y such that

x1 + x2 + u n → x1 + x2 ,+Y ,

x1 − x2 + vn → x1 − x2 + Y

1
as n → ∞. Since (un ± vn ) ∈ Y , we have
2
2 2
2 2 un + vn un − vn
2 x1 + Y + 2 x2 + Y ≤ 2 x1 + + 2 x2 +
2 2
2 2
= x1 + x2 + u n + x1 − x2 + vn

by the parallelogram law in X . Letting n → ∞, we obtain

2 2 2 2
2 x1 + Y + 2 x2 + Y ≤ x1 + x2 + Y + x1 − x2 + Y (4)

Combining (3) and (4), we see that

2 2 2 2
( x1 + Y ) + ( x2 + Y ) + ( x1 + Y ) − ( x2 + Y ) = 2 x1 + Y + 2 x2 + Y

Thus the quotient norm on X / Y satisfies the parallelogram law.


Hence, by this norm is induced by an inner product.
23
Block 3 Hilbert Spaces

E4) i) Let x, y ∈ S and k ∈ K .

Let z = kx + y

Then for all u ∈ S , we have

kx + y , u = kx, u + y , u
= 0 + 0 , since x, y ∈ S ⊥
=0
Thus S ⊥ is a subspace.

To show that S ⊥ is closed, let {xn } be a sequence in S ⊥ such that


xn → x in X . Then

x, u = lim xn , u , ∀ u ∈ S .
=0
Thus show that x ∈ S ⊥ . Hence S ⊥ is closed.

ii) We first show that S ⊥ ⊂ S ⊥

Let x ∈ S ⊥ and v ∈ S . Then v = lim un , u n ∈ S .


n

x, v = x, lim u n
n

= lim x, un = 0.
n

This show that x ∈ S ⊥ . Hence S ⊥ ⊂ S ⊥ .

On the other hand, let x ∈ S ⊥ . Then


x, u = 0 ∀ u ∈ S .

In particular x, u = 0 ∀ u ∈ S . This shows that x ∈ S ⊥ . Therefore


S ⊥ ⊂ S⊥ .

Hence S ⊥ = S ⊥ .

iii) Suppose that S ⊥ ≠ {0}. Let x ∈ S ⊥ , x ≠ 0. Since S is dense in X .


Then for any y ∈ X , there exists sequence { yn } in S , such that
yn → y. Therefore

x, y = x, lim yn , yn ∈ S
n

= lim x, yn = 0, since x ∈ S ⊥
n

24
= 0.
Unit 8 Geometry of Hilbert Spaces
This is true for all y ∈ X . This implies that x = 0 which is a
contradiction.

Hence the claim.

iv) To show that S is linearly independent. Let us assume that


α 0 ≠ 0, such that
α 0u0 + α1u1 + ... + α nun = 0
n

α u
i =0
i i =0

 n

Then 

 α u , y  = 0
i=0
i i

E5) ' y ' is a best approximation from F to x implies that

x − y ≤ x − t ∀t ∈ F.

' z ' is a best approximation from E to y implies that

y − z ≤ x − r ∀ r ∈G

x− z = x− y+ y−z

≤ x− y + y−z

≤ x − t + x − t ∀ t ∈ G since G ⊂ F .

Hence z is a best approximation from G to x.

E6) For example, let X = c00 , the linear space of all scalar sequences
having only a finite number of nonzero entries with the inner product.
Then X is not complete.

Define f : X → K by


x ( n)
f ( x) =  , x∈ X.
n=1 n

Then f is linear. By Holder’s inequality,

 ∞ 1  ∞ 2 π2 2
f ( x ) ≤   2    x (n)  =
2
x .
 n=1 n   n=1  6

Hence f is continuous and f ≤ π / 6 .


25
Block 3 Hilbert Spaces
Since f ≠ 0, F = Z ( f ) is a proper closed subspace of X . Let z ∈ F ⊥ .
Since z ∈ c00 , we see that z = ( z (1),..., z (m), 0, 0,...) for some positive
integer m. For 1 ≤ n ≤ m, define xn ∈ X by letting
xn (n) = 1, xn (m + 1) = −(m + 1) / n and xn ( j ) = 0 for all j ≠ n or m + 1.
Then f ( xn ) = 1 / n − (m + 1) / n(m + 1) = 0, so that xn ∈ F and
z (n) = z , xn = 0. Thus F ⊥ = {0}. This shows that F + F ⊥ ≠ X , that is,
the projection theorem does not hold for X .

E7) i) Since F1 I F2 ⊂ F1 and F1 I F2 ⊂ F2 , we have by Problem 5.3.2,

F1⊥ ⊂ ( F1 I F2 ) ⊥ , F2⊥ ⊂ ( F1 I F2 ) ⊥

Thus, ( F1 I F2 ) ⊥ is a closed subspace containing F1⊥ and F2⊥ ,


and hence F1⊥ + F2⊥ . So if F is the closure of F1⊥ + F2⊥ , then

( F1 I F2 ) ⊥ ⊃ F . (5)

On the other hand,

F1⊥ ⊂ F , F2⊥ ⊂ F ;

F1⊥⊥ ⊃ F ⊥ , F2⊥⊥ ⊃ F ⊥ .

But F1 , F2 , F are closed subspaces of a Hilbert space H . So, by


problem 5.3.4(b),

F1⊥⊥ = F1 , F2⊥⊥ = F2 , F ⊥⊥ = F .

Hence,

F1 ⊃ F ⊥ , F2 ⊃ F ⊥ , F1 I F2 ⊃ F ⊥ ,

( F1 I F2 ) ⊥ ⊂ F ⊥⊥ = F .

This together with Eqn. (5) proves (i).

ii) Let x be any element in the closure of F1 + F2 . Then there is a


sequence {xn } in F1 + F2 such that xn → x. Thus,

xn = yn + zn , yn ∈ F1 , zn ∈ F2 .

Since F1 ⊥ F2 , we get

2 2
xn − xm = ( yn − ym ) + ( z n − z m )

2 2
= yn − y m + z n − z m
26
Unit 8 Geometry of Hilbert Spaces
for all n, m . This shows that { yn }, {zn } are Cauchy sequences in
F1 , F2 respectively. But F1 , F2 are complete as they are closed in
H . Hence there exist y ∈ F1 , z ∈ F2 such that yn → y, zn → z.
Hence, xn → y + z. So x = y + z ∈ F1 + F2 . This shows that F1 + F2
is closed.

For the last part consider the Hilbert space H = l 2 . Let

F1 = {x ∈ l 2 : x(2 j ) = 0, j ≥ 1},

F2 = {x ∈ l 2 : x(2 j − 1) = jx(2 j ), j ≥ 1}.

Then F1 and F2 are easily seen to be subspaces of l 2 . If xn → x


in l 2 , then xn ( j ) → x( j ) for each j. Using this we also see that
F1 and F2 are closed. Let en be the n -th unit vector,
n
 1 1
y n =   −  e2 j −1 ,
j =1  2 j − 1 2

n
1 1 
z n =   e2 j −1 + e2 j 
j =1  2 2j 

Then, yn ∈ F1 ,

1 1
z n (2 j − 1) = , z n (2 j ) = for 1 ≤ j ≤ n,
2 2j

zn (2 j − 1) = 0 = zn (2 j ) for j > n.

Hence z n ∈ F2 . Also,

2n
1
yn + z n =  ei
i =1 i

1
So, if x(m) = for all m, then x ∈ l 2 and yn + z n → x. Hence,
m
x ∈ F1 + F2 . Suppose that x = y + z , where y ∈ F1 and z ∈ F2 . Then,

1
= x (2 j ) = y (2 j ) + z (2 j ) = z (2 j );
2j

1
z (2 j ) − 1) = jz (2 j ) = .
2

This is not possible since z ∈ l 2 . Thus F1 + F2 is not closed. This


completes the claim.
27
Block 3 Hilbert Spaces
E8) By the Riesz representation theorem for Hilbert spaces a representer is
an element (α1 , α 2 , α 3 ) of R 3 such that

f ( x ) = α1 x1 + α 2 x2 + α 3 x3

= ( x, α)

for some α ∈ R 3 . Take α 0 = (1,1,−1). Then

f ( x) = x1 + x2 − x3 = t

= ( x, α 0 ).

Therefore (1, 1, − 1) is a representer of f .

E9) By definition

E a = {x′ ∈ H ′ : x′( x) = 0 ∀ x ∈ E}

Since H ′ = H (by the Riesz representation theorem), x′ represents an


element x0 ∈ H , known as representor. Then we have
x0 , x = 0 ∀ x ∈ E. This shows that x0 ∈ E ⊥ .

Suppose y0 ∈ E ⊥  ( y0 , x ) = 0 ∀ x ∈ E. Since H = H ′, y0 is
representer of y. Therefore, corresponding to y0 , there exists a linear
functional x0′ = y0 such that
x′0 ( x ) = y0 , x = 0 ∀ x ∈ E  x0′ ( x) = 0  x′0 ∈ E a . Hence the result.

28
UNIT 9

ORTHONORMAL SETS
SETS

Structure Page No
9.1 Introduction
Objectives
9.2 Orthornormal Basis
9.3 Bessel’s Inequality
9.4 Parseval’s Identity
9.5 Riesz-Fischer Theorem
9.6 Summary
9.5 Solutions/Answers

9.1 INTRODUCTION
In the last unit, Unit-8, Sec. 8.3, we have already discussed orthogonality
property of inner product spaces. We have explained orthonormal sets and
proved that every finite dimensional inner product space have a basis which is
an orthonormal set. In this unit we make a deeper study of orthonormal sets.
We study orthonormal basis in infinite dimensional spaces.

We shall consider some important results on inner product spaces and Hilbert
spaces which are exclusively based on the inner product structure of the
space. Certain results may not be applicable to all normed linear spaces or
Banach space. We prove Bessel’s inequality, Parseval’s identity and the
Riesz-Fischer theorem. These theorems infact characterizes inner product
spaces or Hilbert spaces and have got lot of applications.

Objectives
The objectives of this unit are to
• explain orthonormal basis and their significance,
• state and prove the following theorems
i) Bessel’s inequality
ii) Parseval’s identity
iii) Riesz-Fischer theorem.

9.2 ORTHORNORMAL BASIS


Let X be an inner product space with inner product ⋅, ⋅ , and let ⋅ be the
corresponding norm, i.e.
Block 3 Hilbert Spaces
1/ 2
x = x, x ∀ x∈ X.

Unless otherwise stated explicitly, the inner product spaces that we consider
are nontrivial, i.e., of dimension at least one.

Recall from Unit 8 that a subset S of X is an orthogonal set in X if x, y = 0


for every distinct x, y ∈ S , and an orthogonal set S is called an orthonormal
set if x = 1 for every x ∈ S . Recall also that every orthonormal set is linearly
independent.

By an orthonormal sequence we mean a sequence (u n ) in X whose terms


form an orthonormal set.

Infact, Gramm-Schmidt orthogonalization is a procedure which takes a non


orthogonal set of linearly independent sets and constructs an orthogonal set.
We explain this in the following theorem.

Theorem 1: Let X be an inner product space and {x1 ,..., xn } be linearly


independent in X . Let u1 = x1 , and for j = 2,..., n, let

j −1 x j , ui
uj = xj −  ui .
i =1 ui , u i

Then {u1 ,..., u j } is an orthogonal set, and

span {x1 ,...., x j } = span {u1 ,...., u j }, j = 1,...., n.

x2 , u1
Proof: Define u1 = x1 and u 2 = x2 − u1.
u1 , u1

Clearly, u2 ∈ span{x1 , x2 } and u2 , u1 = 0. Since {x1 , x2 } is linearly


independent, u2 ≠ 0. Therefore, by Proposition 4 (iv), {u1 , u 2 } is linearly
independent, and hence, it follows that

span{u1 , u2 } = span{x1 , x2 }.

Having defined an orthogonal set {u1 ,...., u j −1} such that

span {u1 ,..., u j −1} = span {x1 ,..., x j −1},

let

j −1 x j , ui
uj = xj −  ui .
i =1 u i , ui

Then we have u j ∈ span {x1 ,..., x j } and u j , u k = 0, k < j. Again, since


{x1 ,..., x j } is linearly independent, u j ≠ 0. Thus, the orthogonal set {u1 ,..., u n }
has the required property.
30
Unit 9 Orthonormal Sets
Remark 1: a) From the above theorem, we can deduce that if {x1 , x2 ,...} is a
denumerable linearly independent set in X , then {u1 , u 2 ,...} defined by

j −1 x j , ui
u1 = x1 , uj = xj −  ui , j = 2, 3,...,
i =1 ui , u i

is an orthogonal set, and

span {x1 ,..., x j } = span {u1 ,..., u j } ∀ j ∈ N.

b) Theorem 4 also shows that every finite dimensional inner product space
has a basis which is an orthonormal set. Let X be an inner product space
of dimension n, and let E = {u1 ,..., un } be a basis of X which is also an
orthonormal set. Then, for every x ∈ X , we have

n n 2
x =  x, u j u j , x =  x, u j
2
.
j =1 j =1

Indeed, if x = α1u1 + ... + α n u n ∈ X , then

x, ui = α1 u1 , ui + ... + α n un , ui = αi ∀i = 1,..., n.

From this, the required representations of x and x follow.

Do we have analogous representations in infinite dimensional spaces? We


shall discuss this question later.

Using the concept of orthonormality, we define a new concept in an inner


product space, namely, an orthonormal basis.

Definition 1: Let E be an orthonormal set in an inner product space X . Then


E is said to be an orthonormal basis of X or a complete orthonormal
system for X if it is a maximal orthonormal set in X , that is, E is an
~ ~
orthonormal set, and for every orthonormal set E satisfying E ⊆ E , we have
~
E = E.

Then we have the following result.

Proposition 1: If E is an orthonormal set in an inner product space X , then


X has an orthonormal basis containing E.

Proof: Let X be an inner product space and let E be an orthonormal set in


X . Consider the family E of all orthonormal sets containing E. E is
nonempty, as E ∈ E . It is seen that E is a partially ordered set with respect to
set inclusion. If T is a totally ordered subset of E , then the union of all
elements of T is an upper bound of T . Therefore, by Zorn’s lemma, E has a
maximal element which is an orthonormal basis of X .

Since every orthonormal set is linearly independent, it is clear that every basis
which is an orthonormal set is an orthonormal basis as well. But, an
orthonormal basis need not be a basis. 31
Block 3 Hilbert Spaces
Here is an example to this effect.

Example 1: Let X = l 2 , and E = {e1 , e2 ,...}, with e j (i ) = δij ; i, j ∈ N. Clearly,


E is an orthonormal set. Since x, en = x(n) for all x ∈ l 2 and for all n ∈ N,

x ∈l2, x, en = 0 ∀n ∈ N  x = 0.
~
This shows that, there cannot be an orthonormal set E which properly
contains E. Thus, E is an orthonormal basis. But, since l 2 is a complete
space, by Theorem 2.30, E is not a basis of l 2 . Another way to see this is
that any finite linear combination of the en has only finitely many zero terms.
1
For example   ∈ l 2 , but is not in the span of {en : n = 1, 2...}.
n

In fact, the property of the set E that we used in the above example is a
characterization property of any orthonormal basis, as we shall see in the
following theorem. We shall show, in Sec. 9.5, that an orthonormal basis of a
Hilbert space is a basis if and only if the space is finite dimensional.

Recall that for a set E ⊆ X , we have the orthogonal complement

E ⊥ = {u ∈ X : u, x = 0 ∀x ∈ S}.

Here is an interesting criterion for an orthonormal set to be an orthonormal


basis.

Theorem 2: Let E be an orthonormal set in an inner product space X . Then


E is an orthonormal basis if and only if E ⊥ = {0}, i.e.,

x∈ X, x, u = 0 ∀ u ∈ E  x = 0.

Proof: Suppose E is an orthonormal set in X . We show that E ⊥ ≠ {0} if and


only if E is not an orthonormal basis.

If E ⊥ ≠ {0}, and x ∈ E ⊥ is a nonzero element, then taking u = x / x , we see


~
that E := E ∪ {u} is an orthonormal set which properly contains E so that E
is not an orthonormal basis.

Conversely, suppose that E is not an orthonormal basis. Then there exists an


~ ~ ~
orthonormal set E ≠ E such that E ⊂ E . Let v = E \ E. Then we have v ≠ 0
and v ∈ E ⊥ . Thus, E ⊥ ≠ {0}.

We have already mentioned that every orthonormal set which is also a basis is
an orthonormal basis. From Theorem 1, we can derive something more.

Corollary 1: Let E be an orthonormal set in an inner product space X . Then


E is an orthonormal basis of span E and its closure.

Proof: Since E is a basis of span E , it is an orthonormal basis of span E as


well. Now, to show that E is an orthonormal basis of span E , by Theorem 1, it
is enough to prove that
32
Unit 9 Orthonormal Sets
x ∈ span E , x, u = 0 ∀ u ∈ E  x = 0.

For this, suppose x ∈ span E such that x, u = 0 for every u ∈ E. Then it


follows that x, y = 0 for every y ∈ span E. Now, let ( xn ) be a sequence in
span E such that x = lim n →∞ xn . Then, using the continuity of inner product
and the above observed fact, we get

x, x = lim x, xn = 0,
n→∞

proving x = 0.

Let us see some examples of orthonormal basis.

Example 2: (i) Consider X = K n with the standard inner product

n
x, y =  α j β j , x, y ∈ K n .
j =1

Clearly, {e1 ,...., en } with e j (i ) = δij for i, j = 1,..., n, is an orthonormal set which
is also a basis of X . Hence, it is an orthonormal basis of X .

ii) Let X = c00 with the l 2 − inner product, i.e.,


x, y =  x( j ) y ( j ), x, y ∈ c00 .
j =1

Then E = {e1 , e2 ,....} with e j (i ) = δij for i, j = 1, 2,..., is an orthonormal set,


and it is also a basis of X . Therefore, E is an orthonormal basis of X .

We have already observed in Example 1 that E is an orthonormal basis


of l 2 . This is a consequence of Corollary 1 as well, since c00 is dense in
l 2 and E is an orthonormal basis of c00 .

iii) Let X 0 = P[a, b] with the L2 -inner product

b
x, y =  x (t ) y (t ) dt , x, y ∈ P[a, b] .
a

Let x j (t ) = t j −1 for t ∈ [a, b], j = 1, 2,..., and let E = {u1 , u2 ,...} be the
orthonormal set obtained from {x1 , x2 ,...} by Gramm-Schmidt
orthogonalization procedure. Since E is a basis of P[a, b], by Corollary
1, it is an orthonormal basis of X 0 .

Recall that if [a, b] = [−1,1], then the polynomials u1 , u 2 ,... are known as
Legendre polynomials.

iv) Let X = C[a, b] with L2 -inner product, i.e.,


33
Block 3 Hilbert Spaces
b
x, y =  x(t ) y (t ) dt , x, y ∈ C[a, b],
a

and let X 0 and E be as in (iii) above. Since E is an orthonormal basis of


X 0 and since P[a, b] is dense in X , by Corollary 1, E is an orthonormal
basis of X .

v) Let X and E be as in (iv). Since E is an orthonormal basis of X and


since P[a, b] is dense in L2 [a, b] , by Corollary 1, E is an orthonormal
basis of L2 [a, b].

We give another orthonormal basis of L2 [a, b] using the definition of


Fourier coefficients. We take K = C and for convenience of
representation we consider [a, b] = [0, 2π]. Let

e int
u n (t ) = , t ∈ [0, 2π], n ∈ Z.

Then it can be seen that

un , u m = δ nm ∀ n, m ∈ Z,

so that E = {u1 , u2 ,..} is an orthonormal set. It is known by Weierstrass


approximation theorem, that span E , the set of all trigonometric
polynomials, is dense in C[0, 2π] with respect to the norm ⋅ ∞
. Since

x 2 ≤ 2π x ∞
∀ x ∈ C[0, 2π],

it follows that span (E ) is dense in C[0, 2π] with respect to ⋅ 2


as well.
2
Now since C[0, 2π] is dense in L [0,2π] , by Corollary 1, E is an
orthonormal basis of L2 [0, 2π].

vi) Let u1 (t ) = 1 / 2π , and for n ∈ N,

sin( nt ) cos(nt )
u2 n (t ) = , u2 n+1 (t ) = .
π π

Then it can be seen that E = {u1 , u2 ,...} is an orthonormal set in L2 [0, 2π].
By using the same arguments as in the last example, it also follows that
E is an orthonormal basis of L2 [0, 2π].

We note that the orthonormal bases for all the spaces K n , c00 , l 2 , C[ a, b],
L2 [a, b] considered in Example 1 are countable. Recall that these spaces are
separable. Now we see that this is, in fact, the case for every separable inner
product space.

Proposition 1: Every orthonormal set in a separable inner product space is


34 countable.
Unit 9 Orthonormal Sets
Proof: Let X be an inner product space and E be an orthonormal set in X .
We observe that for every u , v ∈ E with u ≠ v,

2 2 2
u − v = u + v = 2.

Therefore, if E is uncountable, then there is an uncountable family of disjoint


open sets, namely, {B (u, 2 ) : u ∈ E} in X . Hence, the space X is not
separable.

We know that the spaces C[a, b], l 2 and L2 [a, b] cannot have denumerable
basis. Hence, the orthonormal basis considered in Example 1 for these spaces
are not basis.

But if the space is finite dimensional, then every orthonormal basis is a basis.

n
Suppose {u1 ,..., u n } is an orthonormal set in a Hilbert space H . If x = α u ,
1
j j

then, on taking innerproduct with ui , we see that α i = x, ui and so


n
x =  x, ui ui . Further
1

n n

α u , α v =  α j =  x, ui
2 2 2
x = j j i i
1 1

When H is finite dimensional and {u1 ,...., u n } is a basis for H , every x ∈ H


2
x, u j u j and x =  x, u j
2
has an expansion x = . What happens when
H is infinite dimensional? If {u j } is an inner products we get α j = x, u j and
2
x =  x, u j
2
. If there is an x ∈ H , x ≠ 0, such that x ⊥ u for all j, then
2
clearly both the equalities fail. However we will see that the series  x, u j


2

2
and x, u j u j always converge for all x and x, u j ≤ x . Equality


2

2
may not hold. In fact x, u j = x for all x if and only if x, u j u j = x
for all x. These both hold precisely when {u j } is a maximal orthonormal set or
an orthonormal basis. We will illustrate all these below.

Let us look at the example of X = l 2 and E = {e1 , e2 ,...}. In this case, we know
that x, e j = x( j ) for all j ∈ N, so that

∞ ∞ 2
x =  x( j ) =  x, e j
2 2
∀ x ∈l 2.
j =1 j =1

n
For x ∈ l 2 , if we take s n =  x( j )e , then it follows that
j =1
j

 x( j)
2 2
x − sn = →0 as n → ∞.
j = n+1
35
Block 3 Hilbert Spaces
Thus, we also have


x = lim sn =  x, e j e j .
n →∞
j =1

Before going further, we observe a few facts concerning a series of the form


j =1
x, u j u j , where {u1 , u 2 ,...} is a denumerable orthonormal set. Suppose a

series j =1
x, u j u j converges. Can we rearrange the series and still have its

convergence? The answer is in the affirmative, as the following proposition


shows.

Proposition 2: Let E = {u1 , u2 ,...} be a denumerable orthonormal set in an


inner product space X , and let ψ : E → K be a function. Suppose (vn ) is a

sequence obtained by rearranging the terms of (un ). If the series  ψ(u ) u
j =1
j j


converges, say to s ∈ X , then the series  ψ(u )u
j =1
j j also converges to s.

∞ ∞
Proof: Suppose that the series  ψ(u j )u j converges, say s =  ψ (u j ) u j .
j =1 j =1

For n ∈ N, let
n n
s n =  ψ (u j ) u j , ξ n =  ψ (v j ) v j ∀ n ∈ N.
j =1 j =1

We show that (ξ n ) converges to s. For this, first we observe that


2 2 2
s − ξn = s + ξ n − s, ξ n − ξ n , s ,
where

n
= ξ n , ξ n =  ψ (v j ) ,
2 2
ξn s, ξn = lim sm , ξ n .
m→∞
j =1

For a fixed n, taking m large enough such that {v1 ,..., vn } ⊆ {u1 ,...., u m }, it
follows that
n
s m , ξ n =  ψ (v j ) = ξ n , s m .
2

j =1

Thus,
2 2 2
s − ξn = s − ξn ∀ n ∈ N.

Now, since
∞ ∞
=  ψ (v j ) =  ψ (u j ) = lim sn
2 2 2 2 2
lim ξ n = s ,
n→∞ n →∞
j =1 j =1

36 we have lim n→∞ ξ n = s.


Unit 9 Orthonormal Sets
Now we introduce some notations.

Notation 1: Suppose E is a countable orthonormal set in an inner product


space X and ψ : E → K be a function. If E is a finite orthonormal set, then
the meaning of  u ∈E ψ (u ) u is obvious. Now, suppose that E is a
denumerable orthonormal set and x ∈ X . Suppose u1 , u 2 ,... and v1 , v2 ,... are
two different enumerations of E. Then by Proposition 2, we know that
∞ ∞

 ψ(u )u
j =1
j j converges if and only if  ψ (v )v
j =1
j j converges with the same sum.

Thus, we say that the expression  u ∈E ψ(u )u converges or well defined



provided the series  ψ(u )u
j =1
j j converges for a particular enumeration, say

u1 , u 2 ,..., of elements E.

Next, suppose that E is an arbitrary, but not necessarily countable,


orthonormal set in X , and let ψ : E → K. If the set E0 = {u ∈ E : ψ(u ) ≠ 0} is
countable, and if  u∈E0 ψ (u )u is convergent, then we denote this sum by
 u∈E ψ (u ) u , and say that  u ∈E ψ (u )u is convergent or well defined.

Let us see some examples.

Example 3: Let H = L2 ( [− π, π]) and for n = 0, ± 1, ± 2,...,

e int
u n (t ) = , t ∈ [− π, π].

Then {u n : n = 0, ± 1, ± 2,...} is an orthonormal set in H . For f ∈ H and


n = 0, ± 1, ± 2,..., we have

π
1
 f (t ) e
−int
f , un = dm(t ) = 2π xf (n),
2π −π

Another useful orthonormal basis for H = L2 ( [− π, π] ) is obtained by noting


that eint = cos nt + i sin nt, cos nt = (eint + e − int ) / 2 and sin nt = (eint − e − int ) / 2i for
t ∈ [− π, π], n = 1, 2,... . Then

 1 cos nt sin nt 
 , , : n = 1, 2,...
 2π π π 

is an orthonormal basis for H . If we let

π π
1 1
a0 = 
2π −π
f (t ) dm(t ), an =  f (t ) cos nt dm(t ),
π −π

π
1
π −π
bn = f (t ) sin nt dm(t ), n = 1, 2, ....,
37
Block 3 Hilbert Spaces
then for x ∈ H the following expansion is known as the Fourier expansion
of f

f (t ) = a0 +  (an cos nt + bn sin nt ),
n =1

where the series converges in . 2 . Also, for x ∈ H we have the following


formula which is known as the Parseval formula.

π ∞
1

π −π
f
2
dm = 2 a0
2
+ 
n =1
2 2
an + bn . ( )
Example 4: Let H = L2 ( [− 1,1]) and for n = 0,1, 2..., let un denote the
Legendre polynomial of degree n, obtained by applying the Gram-Schmidt
orthogonalization to the linearly independent set { f 0 , f1 ,...}, where
f n (t ) = t n , t ∈ [− 1,1]. Then {u0 , u1 ,...} is an orthonormal set in H .

Let f ∈ H and ε > 0. As the set of all continuous functions on [− 1,1] is dense
in H , there is some g ∈ C ( [− 1,1] ) such that f − g 2
< ε. By the Weierstrass
theorem, there is a polynomial p such that g (t ) − p (t ) < ε for all t ∈ [− 1,1].
Then

 g (t ) − p (t )
2 2
g− p 2 = dt ≤ 2ε 2 .
−1

Thus f − p 2
≤ f − g 2 + g − p 2 < (1 + 2 )ε . This shows that the set of all
polynomials on [−1,1] is dense in H . Since

span{u0 ,..., un } = span{ f 0 ,..., f n }


= { p : p is a polynomial of degree ≤ n},

we see that span{u0 , u1 ,...} is the set of all polynomials on [−1,1]. Also,
{u0 , u1 ,...} is an orthonormal basis for H .

Example 5: Let X be the real space C[−1, 1] of all real-valued continuous


functions on [−1,1] with the inner product given by

1
( f , g) =  f (t ) g (t ) dt , x, y ∈ X .
−1

If Y is the set of all odd functions in X , then let us find its orthogonal
complement Y ⊥ .

To find this, let us note that a function f is odd if f (−t ) = − f (t ) for all t ; it is
even if f (−t ) = f (t ) for all t. The orthogonal complement of Y is

Y ⊥ = { f ∈ X : f , g = 0 for all y ∈ Y }.
38
Unit 9 Orthonormal Sets
Let f be any even function and g ∈ Y . Then f (t ) g (t ) is an odd function of
t , and hence

1
f ,g  f (t ) g (t ) dt = 0.
−1

Thus, f , g = 0 for all y ∈ Y . Therefore, f ∈Y ⊥ .

On the other hand, let f ∈Y ⊥ , and g (t ) = f (t ) − f (−t ). Then,


g (−t ) = f (−t ) − f (t ) = − g (t ), and so g ∈ Y . Hence, f , g = 0 . Since

( f (t )) 2 = ( f (t ) − f (−t ) g (t ) = f (t ) g (t ) + f (−t ) g (−t ),


we get
1 1 1

 ( g (t ))  f (t ) g (t ) dt +  f (t ) g (−t )dt
2
dt =
−1 −1 −1

= f , g + f , g = 0.

Since f is continuous this implies that g (t ) = 0 for all t , that is, f (−t ) = f (t ).
So f is an even function if f ∈Y ⊥ . Thus, Y ⊥ is the set of all even functions.

Example 6: Let f1 (t ) = t 2 f 2 (t ) = t and f 3 (t ) = 1. Let us orthonormalise


f1 , f 2 , f 3 in this order, on the interval [−1,1] with respect to the inner product

1
f , g =  f (t ) g (t ) dt.
−1

We find orthonormal functions u1 , u 2 , u3

1 1
2
=  ( f1 (t )) 2 dt =  t 4 dt =
2
Since f1 ,
−1 −1
5

We take

f1
u1 = = (5 / 2) f1.
f1

Let

f 2 = f 2 − f 2 , u1 u1.

Then g 2 , u1 = 0. Since

1
f 2 , u1 =  t 5 / 2t 2 dt = 0,
−1

We have g 2 = f 2 , 39
Block 3 Hilbert Spaces
1
2
=  ( f 2 (t )) 2 dt = .
2
g2
−1
3

So we take
g2
u2 = = (3 / 2) f 2 .
g2
Let

g 3 = f 3 − f 3 , u 2 u 2 − f 3 , u1 u1

Then, g 3 , u 2 = 0 = g 3 , u1 . Since

1
g 3 , u 2 =  (3 / 2) f 2 (t ) dt = 0,
−1
1
2
f 3 , u1 = 
−1
(5 / 2) f1 (t ) dt = (5 / 2) ,
3

We get

2 5
g3 = f3 − (5 / 2)u1 = f 3 − f1 ,
3 3
1 1 2
 5  8
=  ( g 3 (t )) dt =  1 − t 2  dt = .
2 2
g3
−1 
−1
3  9

Take

y3  5 
u3 = = (9 / 8)  f 3 − f1 .
y3  3 

Thus,

2 3 − 5t 2
u1 (t ) = (5 / 2)t , u 2 (t ) = (3 / 2)t u 3 (t ) = (1 / 2) .
2
2
Example 7: Let H be the Hilbert space l 2 or L [−π, π]. In either case, find an
infinite orthonormal set in l 2 , which is (i) an orthonormal basis, (ii) not an
orthonormal basis.

Solution: Consider the case H = l 2 . Let en be the n -th unit vector in l 2 .


Then, ei , e j = ei ( j ) = 0 or 1 according as i ≠ j or i = j. So {en } is an
orthonormal set in l 2 . Since x, en = x( n ), we see that x, en = 0 for all n
only if x(n) = 0 for all n, that is x = 0. This implies that {en } is a total
orthonormal set in l 2 .

Removing some vectors from {en }, we find an infinite orthonormal set which is
not total.
40
Unit 9 Orthonormal Sets
2
Next, let H = L [− π, π]. For integers n, m, we have

π
0 if n ≠ m
e
int −imt
e dt = 
−π 2π if n = m.

So let

u n (t ) = ( 2 π) −1/ 2 eint , n = 0, ± 1, ± 2,....

Then, {u n : n = 0, ± 1, ± 2,...} is orthonormal in L2 [−π, π]. To show that it is an


orthonormal basis, we use Fejer’s theorem which states that if x is a 2 π -
periodic continuous function and

π
1 1
xˆ (n) = 
2π −π
x(t ) e −int dt =

x, un ,

N
s N (t ) =  xˆ ( n ) e int , N = 0,1, 2,.....,
−N

then (1 / N ) ( s0 + s1 + ... + s N −1 ) → x uniformly.

Suppose that u ∈ L2 [−π, π] and u , u n = 0 for all integers n, that is, uˆ (n) = 0
for all n. Let

t
x(t ) =  u ( s) ds, − π ≤ t ≤ π.
−π

Then x is absolutely continuous, x′ = u a.e. and

π
x(π) =  u ( s) ds = 2πuˆ (0) = 0 = xˆ (−π).
−π

If n ≠ 0, then integration by parts gives

π
2π xˆ (n) =  x(t ) e −int dt
−π

1 2π
=
− in
[
x(π) e −in π − x(−π) e in π + ]
in
uˆ (n) = 0.

Thus, xˆ (n) = 0 for all n ≠ 0, and

1
s N (t ) = xˆ (0), ( s0 + ... + sN −1 ) = xˆ (0)
N

for all N . It follows from Fejer’s theorem that x(t ) = xˆ (0) for all t. Hence
u = x′ = 0 a.e.. Thus, u , u n = 0 for all n only if u is the zero element of
L2 [−π, π]. Therefore, {u n } is complete.
41
Block 3 Hilbert Spaces
Removing some vectors from {un }, we get an orthonormal set which is not
complete.

Here is an exercise for you.

E1) Show that every orthonormal set in an inner product space is a closed
set.

E2) Let u n be the sequence in l 2 with 1 in the nth place and zeros
elsewhere. Prove that the set {u n } is an orthonormal basis for l 2 .

E3) Let X = l 2 , for n = 1, 2, ... let xn = (1,...,1,0,0,...) where 1 occurs only in


the first n places. Prove that the Gram-Schmidt orthogonalization yields
the sequences {u n } where u n = (0,...,0,1,0,...0,...) where 1 occurs only in
the nth place.

E4) Let f n (t ) = t n for n = 0,1, 2 ... and − 1 ≤ t ≤ 1.

i) Let {g n } be the orthogonal sequence constructed out of { f n }


using Gram-Schimdt process. Show that {g n } is an orthonormal
basis.

ii) Use the Gram-Schmidt process on { f 0 , f1 ,...} to determine the first


three vectors in the orthonormal set.

Next we consider another important result.

9.3 BESSEL’S INEQUALITY


Bessel’s inequality is a statement about coefficients of an element x in a
Hilbert space H with respect to an orthonormal sequence. The inequality was
established by F.W. Bessel in 1828.
Friedrich Bessel
German mathematician In this section we shall study the inequality, which is generalization of Schwarz
(1784-1845) inequality. The inequality is very useful to establish certain important results.

Theorem 3(Bessel’s inequality): Let E be a countable orthonormal set in an


inner product space X . Then for every x ∈ X ,


2 2
x, u ≤ x .
u ∈E

Moreover, for x ∈ X ,

 = x ⇔  x, u u = x.
2 2
x, u
u∈E u∈E

Proof: Suppose {u1 ,..., u n } ⊆ E. For x ∈ X , let


n
s n =  x, u j u j .
42 j =1
Unit 9 Orthonormal Sets
Then, using the orthonormality of E , we have

n 2
=  x, u j
2
x, sn = sn , x = sn
j =1

so that

n 2
= x −  x, u j
2 2
0 ≤ x − sn .
j =1

From this the conclusions in the theorem follow.

Remark: An obvious consequence of Theorem 3 is that if {u1 , u 2 ,...} is a


denumerable orthonormal set in an inner product space X , then

x, u n → 0 as n→∞

for every x ∈ X . This implies, in view of Example 3, for every x ∈ L2 [0, 2π],

2π 2π

 x(t ) cos (nt ) dµ(t ) → 0,


0
 x(t ) sin (nt ) dµ (t ) → 0
0

as n → ∞. These results are known as the Riemann-Lebesgue Lemma.

Next we consider another important theorem as a consequence of Theorem 3.

Theorem 4: Let E be an orthonormal set in an inner product space X . Then


for every x ∈ X , the set

E x = {u ∈ E : x, u ≠ 0}

is a countable set.

Proof: For x ∈ X , we note that


E x = {u ∈ E : x, u ≠ 0} = U E x , n ,
n =1

where

 1
E x,n = u ∈ E : x, u > .
 n

Thus it is enough to show that E x ,n is finite for each n . For this, let, u1 ,..., uk in
E x ,n . Then, by Theorem 3, it follows that
k
k 2

2
≤ x, u j ≤ x
n 2 j =1

43
Block 3 Hilbert Spaces
2
so that k ≤ n 2 x . This proves (How?) that E x ,n is a finite set for each n, and
hence, E x is a countable set.
Here is an exercise for you.

E5) Let X be an inner product space and E be an orthonormal set.

i) If E is a finite set, then show that equality occurs in Bessel’s


inequality if and only if x ∈ span E.
ii) If E = {u1 , u2 ,...}, a denumerable orthonormal set, then show that
the function

T : x → ( x, u1 , x, u2 ,...), x∈ X,

is a bounded linear operator from X into l 2 , and it is injective if and


only if E is an orthonormal basis. Give an example to show that T
need not be surjective even if E is an orthonormal basis.

Next we shall see another important result.

9.4 PARSEVAL’S IDENTITY


Parsevel’s identity named after Marc-Antoine Parseval is a fundamental result
Marc-Antoine Parseval
which is a generalized Pythagorean theorem for innerproduct spaces. In
French Mathematician general the identity asserts that the square of the norm of an element is equal
(1755-1836) to the square of the coefficients in the series expansion known as ‘Fourier
Expansion’, as shown in Notation 1, of the Sec. 9.2 with respect to a countable
orthonormal set.

In this section, we prove the equalities in (2) of Sec. 9.2 whenever X is a


Hilbert space and {un : n ∈ N} is an orthonormal basis. In fact, in view of
Theorem 4, we prove generalized forms of inequalities with respect to an
arbitrary (not necessarily countable) orthonormal set. We first we prove a
general result.

Proposition 3: Let X be a Hilbert space and E be an orthonormal set in X .


Then for every x ∈ X , the series 
x, u u is well defined, and
u ∈E

x −  x, u u ∈ E ⊥ .
u ∈E

Proof: Let x ∈ X and E x = {u ∈ X : x, u ≠ 0}. By Theorem 4, we know that


E x is a countable set. Suppose E x is a finite set, say Ex = {u1 ,..., un }. Then
n
y =  x, u j u j =  x, u u, and for every v ∈ E,
j =1 u ∈E

n
y , v =  x, u j u j , v = x, v .
j =1

44
Unit 9 Orthonormal Sets

Hence, x − y ∈ E . Next, suppose that E x = {u1 , u2 ,...}, a denumerable set.
n
Let s n = 
j =1
x, u j u j for n ∈ N. Then for n > m, we have
n 2

2
sn − sm = x, u j
j = m +1

Now, by Bessel’s inequality, it follows that ( sn ) is a Cauchy sequence. Since


X is a Hilbert space, the sequence ( sn ) converges. Let y be its limit, i.e.,

y = lim sn =  x, u j u j . Then, using the continuity of the inner product and
n →∞
j =1

the orthonormality of {u1 , u 2 ,...}, it follows that


y , v =  x, u j u j , v = x, v ∀ v ∈ E,
j =1

so that x − y ∈ E ⊥ . This completes the proof.

As a corollary to the above proposition, we prove two results. The first one is
the Fourier expansion theorem and the next one is the Parseval’s identity.

Theorem 5: Let X be a Hilbert space and E be an orthonormal set in X .


Then the following are equivalent:

i) E is an orthonormal basis of X .

ii) (Fourier expansion) For every x ∈ X , x = 


u∈ E
x, u u.

=  x, u .
2 2
iii) (Parseval’s formula) For every x ∈ X , x
u∈ E

Proof: Let x ∈ X and E x = {u ∈ E : x, u ≠ 0}. By Bessel’s inequality


(Theorem 4), we know that E x is a countable set. Then by Theorem 3, we
have

 
2 2
x= x, u u ⇔ x = x, u .
u∈E x u∈E x

Thus, we have the equivalence of (ii) and (iii).


To see the equivalence of (i) and (ii), recall from Theorem 1 that E is an
orthonormal basis if and only if E ⊥ = {0}. But, by Proposition 3, E ⊥ = {0} if
and only if x = 
u∈E
x, u u. Thus, (i) and (ii) are equivalent.

Remark: If E is a countable orthonormal set, then in the proof of the above


theorem, we can take E itself in place of E x . Thus, if {u1 , u 2 ,...} is an
orthonormal set in a Hilbert space, then it is an orthonormal basis if and only if
every x ∈ X can be written as

45
Block 3 Hilbert Spaces
x =  x, un un .
n

In particular, we can conclude that every countable orthonormal basis of a


Hilbert space is a Schauder basis.
Let us see some examples.

Example 8: (i) We have observed earlier that if X = l 2 and the orthonormal


basis E = {e1 , e2 ,...}, then


x =  x( j ) e j ∀ x ∈l2
j =1

Parseval’s formula in this case is nothing but the definition of the norm in l 2 .

(ii) Consider the Hilbert space X = L2 [0, 2π], and orthonormal basis as in
Example 2 (v). By Theorem 5, we have

∞ ∞

 xˆ (n) un ,  xˆ (n)
2 2
x= x = ,
n = −∞ n − −∞

where xˆ (n) is the n -th Fourier coefficient, i.e.,


1
 x(t ) e
−int
xˆ (n) = dt.
2π 0

Here is an exercise for you.

E6) Let X be a Hilbert space and E be an orthonormal set.

 f (u )
2 2
i) If E is a countable set, then prove that ≤ f for every
u ∈E

f ∈ X ′.

ii) Show that, for every f ∈ X ′, E f = {u ∈ E : f (u ) ≠ 0} is countable,


and  f (u) u converges.
u∈E

In the next section we show that the convergence of  f (u) u for every
u∈E

f ∈ X ′ without resorting to the Riesz representation theorem. This can be


answered affirmatively, after proving Riesz-Fischer theorem in Section 9.5.

Before concluding this section, we give an interesting characterization of


Hilbert spaces having countable orthonormal basis.

9.5 RIESZ-FISCHER THEOREM

46
Frigyes Riesz
Hungarian Mathematician
(1880-1956)
Unit 9 Orthonormal Sets
In this section we discuss Riesz-Fischer theorem. The theorem was
independently proved by two mathematicians Frigyes Riesz and Ernst
Sigismund Fishcer in 1907. Initially the theorem was established for L2 -
spaces. But the theorem also applies to more general Hilbert spaces. We have
already discussed in Sec. 9.2 that every orthonormal set in a separable inner
product space is countable. Infact the next theorem asserts that if X is a
Hilbert space, then the existence of a countable orthonormal basis is one of
the characterizing properties of separable Hilbert spaces.

Theorem 6: Let X be a Hilbert space and E be an orthonormal basis of X .


Then E is countable if and only if X is separable.

Proof: Suppose X is a separable Hilbert space. Then we already know, by


Proposition 3, that every orthonormal basis of X is countable.

Conversely, suppose E is a countable orthonormal basis of X . By Theorem


5, we know that span E is dense in X . In fact, E is a Schauder basis of X .
Hence, X is a separable space.

You may recall that every finite dimensional inner product space X is linearly
isometric with l 2 (n) for some n ∈ N, with isometry given by

x → ( x, u1 ,..., x, un ), x∈ X,

where {u1 ,...., u n } is an orthonormal basis of X .

Ernst Sigismund Fischer


Also, if X is an infinite dimensional separable Hilbert space and {u1 , u 2 ,...} is (1875-1954)
an orthonormal basis of X , then by Theorem 5, we know that the map

x → ( x, u1 , x, u 2 ,...), x∈ X ,

is linearly isometry from X into l 2 . A question that naturally arises is whether


this map is onto. We answer this affirmatively, by making use of the following
theorem.

Theorem 7 (Riesz-Fischer theorem): Let {u1 , u 2 ,...} be an orthonormal set in


a Hilbert space X and let (α n ) be a sequence of scalars. Then
∞ ∞

 α n converges if and only if α u


2
n n converges, and, in that case,
n =1 n =1


α n = x, un ∀n∈ N, where x = α u ..
n =1
n n

n
Proof: Let sn = α u
j =1
j j for n ∈ N. Then for n > m, it follows that

α
2 2
s n − sm = n .
j = m +1

47
Block 3 Hilbert Spaces

α
2
From this it is clear that n converges if and only if ( sn ) is a Cauchy
n =1

sequence. Since X is complete, ( sn ) is a Cauchy sequence if and only if


α u
n =1
n n converges. Now the results follow by using the properties of an inner

product.
As a consequence of the above theorem, we show that every infinite
dimensional separable Hilbert space is linearly isometric with l 2 . More
generally, we have the following.

Theorem 8: Let X be an infinite dimensional Hilbert space and (u n ) be an


orthonormal sequence in X . Consider the linear operator T : X → l 2 defined
by

Tx = ( x, u1 , x, u 2 ,...), x∈ X.

Then we have the following:

i) T is surjective

ii) T is an isometry if and only if E is an orthonormal basis.

Proof: Clearly, by Bessel’s inequality, the map T is well defined. By Riesz-


Fischer Theorem 7, it is seen that T is surjective. By Theorem 5, T is linearly
isometry if and only if E is an orthonormal basis.

We have observed earlier that an orthonormal basis of an inner product space


need not be a (Hamel) basis. We have also given examples of infinite
dimensional inner product spaces with orthonormal bases which are bases as
well. The following theorem shows that, in an infinite dimensional Hilbert
space, an orthonormal basis can never be a basis.

Theorem 9: Let X be a Hilbert space and E be an orthonormal basis of X .


Then E is a basis of X if and only if X is finite dimensional.

Proof: Clearly, if X is finite dimensional, then every orthonormal basis is a


basis.

Conversely, suppose that X is an infinite dimensional Hilbert space and E is


an orthonormal basis of X . Let {u1 , u 2 ,...} be a denumerable subset of E. Let

α
2
(α n ) be a sequence of nonzero scalars such that j < ∞. Then, by
j =1

Riesz-Fischer Theorem 7, it follows that the series α u
j =1
j j converges.


Let x =  α u . If E is a basis of
j =1
j j X , then there exists {v1 ,...., vk } ⊆ E and
k
scalars β1 ,..., β k such that x =  β v . Suppose n
j =1
j j is large enough such that

un ∉{v1 ,..., vk }. Then, using the orthonormality of {u1 , u 2 ,...}, we have


48
Unit 9 Orthonormal Sets
∞ k
αn =  α j u j , un =
j =1
β v ,u
j =1
j j n =0

a contradiction to the assumption that α j ≠ 0 for all j ∈ N. Thus, if X is


infinite dimensional, then E cannot be a basis of X .

Next we discuss an alternate proof of the Riesz representation theorem using


Riesz Fischer theorem.

Alternate proof of the Reisz representation theorem (Theorem 10, Unit 8).

We proceed as follows. Let {u α } be an orthonormal basis for H and


{uα : f (uα ) ≠ 0} = {u1 , u 2 ,...}, as in Theorem 9 (b) of unit 8. Then

 f (u )
2 2
n ≤ f <∞
n

by 9 (b). Since H is a Hilbert space, the Riesz-Fischer theorem shows that


the series  f (u ) u
n
n n converges in H . Let

y =  f (u n ) u n .
n

We claim that f ( x) = x, y for all x ∈ H . Let x ∈ H and


{uα : x, uα ≠ 0} = {v1 , v2 ,...} . The Fourier expansion shows that

x =  x, vm vm
m

Hence

f ( x) =  x, vm f (vm ).
m

On the other hand,

x, y =  x, vm vm , y .
m

We show that vm , y = f (vm ) for each m = 1, 2,..., Fix m. Then

vm , y = vm ,  f (un ) un =  f (un ) vm , u n .
n n

Now, if vm = un0 for some n0 , then vm , y = f (un0 ) = f (vm ). Next, let vm ≠ un


for any n. Then vm , y = 0, and also f (vm ) = 0, since
vm ∉ {u1 , u 2 ,...} = {uα : f (uα ) ≠ 0}. It therefore, follows that f ( x ) = x, y .

49
Block 3 Hilbert Spaces
Finally, let us prove the uniqueness of y ∈ H . If f ( x) = x, y1 for all x ∈ H
and some y1 ∈ H as well, then letting x = y − y1 , we obtain
y − y1 , y = f ( y − y1 ) = y − y1 , y1 . Hence y − y1 , y − y1 = 0, that is y1 = y.

Hence the result.

Before concluding this section, we summarise the above theorem and give
some interesting characterization of a Hilbert space having a countable
orthonormal basis.

Theorem 10: Let H be a nonzero Hilbert space over K . Then the following
conditions are equivalent.

i) H has a countable orthonormal basis.

ii) H is linearly isometric to K n for some n, or to l 2 .

iii) H is separable.

We end with a remark followed by an exercise for you to do. The proof of the
theorem is omitted.

Remark: In fact, on every linear space with an uncountable basis, we can


define an inner product under which the basis is an orthonormal basis so that
the space and its completion are not separable.

E7) Let X = c00 , un = (0, ...0, 1,0,0,...0) where only the nth entry is 1 and let
1
kn = for n = 1, 2,... . Prove that the Riesz-Fischer theorem does not
n
hold good for this space.

With this we come to an end of this unit and end of this Block also.

9.6 SUMMARY
In this unit we have covered the following:

1. We have explained orthonormal bases and discussed the relationship of


orthonormal sets and orthonormal basis.

2. We have explained the expression for the series representation of


elements of Hilbert space with reference to an orthonormal set. The
expression is denoted an 
( x, y ) u where E is an orthonormal set.
u ∈E

3. We have stated and proved the following three important theorem

i) Bessel’s Inequality

ii) Parseval’s Identity

50 iii) Riesz Fischer Theorem


Unit 9 Orthonormal Sets

9.7 SOLUTIONS / ANSWERS


E1) Let E be an orthonormal set in an innerproduct space. Let x ∈ E .
Then there exists {xn } in E such that xn → x. To prove that x ∈ E i.e.
( x, y ) = 0 ∀ y ∈ E and x = 1.
Let y ∈ E , then ( x, y ) = lim ( xn , y )
n→∞

Since xn , y ∈ E , ( xn , y ) = 0.

∴ ( x, y ) = 0.

This is true ∀ y ∈ E. Hence x is orthogonal to every element of E.


Also x = lim xn = 1, since xn ∈ E and E is orthonormal. Hence
n

x ∈ E.

Thus E is closed.

E2) Hint: Verify that un , u m = 0 when n ≠ m and un = 1.

E3) Apply Theorem 2 and get

x1
u1 = = (1, 0, 0,...)
x1

u 2 = x2 − x2 , u1 u1

= x2 − u1
= (1,1, 0, 0,...) − (1, 0, 0,...)
= (0,1, 0, 0, 0....)
u3 = x3 − x3 , u1 u1 − x3 , u 2 u 2

= x3 − u1 − u 2
= (0, 0,1, 0, 0,...)

Similarly

un = xn − u1 − u 2 .... − un−1
= (0, 0, ...,1, 0, 0,....)
nth place

Hence the result.

E4) i) The sequence {g n } is an orthonormal sequence in L2 [−1,1]


obtained by Gram-Schimdt process.

yn = span{ f 0 , f1 ,..., f n } = span{g 0 , g1 ,...g n }.


51
Block 3 Hilbert Spaces

Since f n (t ) = t n , we see that Yn consists of polynomials of degree


not exceeding n. To prove that {g n } is an orthonormal basis for
L2 [−1,1], let f ∈ L2 [−1,1] and x, f n = 0 for all n ≥ 0. It follows
that f , h = 0 if h∈ Ym for some m.

Let h be any continuous function on [−1,1] . Then there is a


sequence { yn } of polynomials converging uniformly to y [−1,1] .
So, hn → h in the norm of L2 [−1,1] . Hence f , hn → f , h . But
each polynomial hn lies in some Ym , and so f , hn = 0 for all n.
Thus, x, x = 0 and x = 0. Therefore by Thoerem 2, {g n } is an
orthonormal basis.

ii) The first three terms f 0 , f1 , f 2 of the desired orthonormal


sequence are determined as follows: Since x0 (t ) = 1 for all t ,

1
=  1dt = 2
2
x0
−1

So we take

x0 1
u0 = = .
x0 2

Let

y1 = x1 − x1 , f 0 f 0 .

Then, y1 , f 0 = 0. Since

1
1
x1 , f 0 =  t dt = 0,
−1 2

We have y1 = x1 , and so

1
2
y1 =  t 2 dt = .
2

−1
3

Take

y1
f1 = = (3 / 2) x1.
y1

Let

y2 = x2 − x2 , f1 f1 − x2 , f 0 f 0 .

52
Unit 9 Orthonormal Sets
Then y2 , f1 = 0 y2 , f 0 . Since

1
3
x2 , f1 =  t 2 t dt = 0
−1
2

1
1 2
x2 , f 0 =  t 2 dt = ,
−1
2 3
we get

2 1
y2 = x2 − f 0 = x2 − ,
3 3
1 1
8
=  ( y2 (t )) 2 dt =  (t 2 − 1 / 3) 2 dt =
2
y2 .
−1 −1
45

Take

y2 45  1
f2 = =  x2 − .
y2 8 3

Thus,

1 3 5  3t 2 − 1 
f 0 (t ) = , f1 ( t ) = t , f 2 (t ) =  .
2 2 2  2 

E5) Let E = {x1 , x2 ,...., xm } be an orthonormal set.

m
For each m, let xm =  ( x, x ) x .
n=1
n n

Since {x1 , x2 ,...., xm } is an orthonormal set for each x ∈ X , we have

=  x, xn
2 2
( x, xm ) = ( xm , x ) = ( xm , xm ) = xm
2
0 = x − xm = ( x − xm , x − xm )
m
= x −  ( x, xn )
2 2

n=1

=0
i.e. x = xm

Therefore x ∈ span(E ).

Conversely suppose x ∈ span (E ).

m
Then x = c x ,
n =1
n n {x1 ,....xm } orthonormal

53
Block 3 Hilbert Spaces
( x, xn ) = ( cn xn , xn )

= cn ( xn , xn )

= cn

i.e. x =  ( x, x ) xn n

 m n

∴ x = ( x, x) =   ( x, xn ) xn ,  ( x, xn ) xn 
2

 n=1 n=1 

=  ( x, xn )
2

Hence the equality holds in Bessel’s inequality.

E6) i) Hint: Follows from Riesz representation theorem and Bessel’s


inequality.

ii) Use Riesz representation theorem and Proposition 3.

E7) Note that

k
2
n <∞

 1   1 
k u
n =1
n n = (1, 0,....,0) +  0, ,0,... +  0, 0, ...
 2   3 

 1 1 
= 1, , ,...
 2 3 

1
=   ∈ l2
n

1
=   ∉ c00
n

54
UNIT 10

OPERATORS ON HILBERT
SPACES
SPACES

Structure Page No
10.1 Introduction
Objectives
10.2 Adjoint of an Operator
10.3 Normal, Unitary and Self-Adjoint Operators
10.4 Summary
10.5 Solutions/Answers

10.1 INTRODUCTION
In this unit, we study bounded linear maps from a Hilbert Space to itself. You
are aware that such maps are known as linear operators on a Hilbert space.
Several types of operators, such as, self-adjoint, normal, unitary, and compact
self adjoints operators, will be studied. Banach spaces are little too general to
yield some rich results on operators. Therefore, we restrict our attention to
Hilbert Spaces. In Sec. 10.2, we define the adjoint of an operator on a Hilbert
space. The adjoint of an operator enables us to define many other important
classes of operators known as self-adjoint, normal and unitary operators which
are discussed in Sec. 10.3.

Objectives
The objectives of this unit are to
• introduce you to the following important classes of bounded linear
operators namely,
i) Self-adjoint
ii) Normal
iii) Unitary
• explain the interrelationship with them.

10.2 ADJOINT OF AN OPERATOR


In this section, we introduce you to the adjoint of an operator on a Hilbert
space.
Block 3 Hilbert Spaces
To begin unit with we recall certain definitions which we have already covered
in unit on bounded linear maps.

By an operator A on an inner product space X over K , we mean a linear


map A from X to X . It is said to be bounded if A( x ) ≤ α x for all x ∈ X
and some α > 0, where x = ( x, x) . A bounded operator A is uniformaly
continuous on X , since for all x and y in X , A( x) − A( y ) ≤ α x − y .
Conversely, if A is a linear map from X to X and A is continuous at 0, then
A is a bounded operator on X .

Here we shall restrict ourselves to bounded linear operators on Hilbert spaces.

The set of all bounded operators on H is denoted by BL(H ) . We say that


A ∈ BL(H ) is invertible if there is some B ∈ BL(H ) such that AB = I = BA,
where I is the identity operator on H .

For A ∈ BL(H ), let

A = sup{ A( x) : x ∈ H , x ≤ 1}.

Then it follows that is a norm on BL(H ) and we have

A( x ) ≤ A x

for all x ∈ X . For A, B in BL(H ) and x in H , we have

AB ( x) ≤ A B ( x ) ≤ A B x ,

so that AB ≤ A B .

If ( An ) and ( Bn ) are sequences in BL(H ) such that An → A and Bn → B in


BL(H ), then we see that An + Bn → A + B and An Bn → AB, the verification is
left as an exercise.

An + Bn − A − B ≤ An − A + Bn − B ,

An Bn − AB ≤ An Bn − An B + An B − AB

≤ An Bn − B + An − A B

To calculate the norm of a bounded operator A on H is, in general, not easy.


It requires to maximize A( x ) , subject to x ∈ H with x ≤ 1. We claim that for
x ∈ H , we have

A( x ) = sup { }
A( x ), y : y ∈ X , y ≤ 1 .

This follows by nothing A( x ), y ≤ A( x) y for all y ∈ H and


56
Unit 10 Operators on Hilbert Spaces

A( x), y = A( x) if we let y = A( x) / A( x) , provided A( x) ≠ 0. Hence we


have

A = sup { A( x), y : x, y ∈ X , x ≤ 1, y ≤ 1 .}
Now we are ready to consider the adjoint of an operator. We introduce the
notion of adjoint of an operator for finite dimensional spaces as a
generalization of the concept of conjugate transpose of a matrix. We shall
explain this

Let ( aij ) be an m × n matrix of scalars. Let A : K n → K m be the linear


operator associated with the above matrix. That is, A is defined by

n
( Ax)(i ) =  aij x( j ), i = 1, ..., m; x ∈ K n .
j =1

Let B : K m → K n be the linear operator associated with the conjugate


transpose of ( aij ), i.e.,

m
( By ) (i) =  a ji y ( j ), i = 1,..., n; y ∈ K m .
j =1

Then, it is easily seen that

Ax, y Km
= x , By Kn
∀ x ∈ Kn , y ∈ K m.

Here, ⋅ , ⋅ Km
, ⋅,⋅ Kn
denote the standard inner products on K m and K n ,
respectively.

Since every linear operator from K n to K m is induced by an m × n matrix, it


can be seen that, corresponding to every linear operator A : K n → K m , there
is a linear operator B : K m → K n such that

Ax, y Km
= x, By Kn
∀x ∈ K n , y ∈ K m .

Since K n is a prototype of any n -dimensional inner product space, it is also


not difficult to show (verify) that, if X and Y are finite dimensional inner
product spaces, and if A : X → Y is a linear operator, then there is a linear
operator B : Y → X such that

Ax, y = x, By ∀ x ∈ X , y ∈Y.

We shall now extend this to infinite dimensional spaces.

Definition 1: Let X and Y be any two inner product spaces (not necessarily
finite dimensional) and A : X → Y be a linear operator. Then A is said to
have an adjoint if there exists a linear operator B : Y → X such that

Ax, y = x, By ∀ x ∈ X , y ∈Y.
57
Block 3 Hilbert Spaces
Such an operator B is called an adjoint of A.

We may observe that, if an adjoint exists, then it is unique:


Suppose B1 , B2 are adjoints of a linear operator A : X → Y . Then we have

x, B1 y = Ax, y = x, B2 y ∀ x ∈ X , y ∈Y

so that x, ( B1 − B2 ) y = 0 for all x ∈ X , y ∈ Y . Consequently, B1 = B2 .

Notation: The adjoint of a linear operator A, if it exists, is denoted by A*.


Note that, if a linear operator A : X → Y has the adjoint A* , then

A* y, x = x, A* y = Ax, y = y , Ax ∀x ∈ X , y ∈ Y ,

and, consequently,

( A* )* = A.

We have already noticed that every linear operator between finite dimensional
inner product spaces has an adjoint. Does this hold for if X , Y are infinite
dimensional spaces? The following example shows that the adjoint may not
exist for infinite dimensional spaces.

Example 1: Let X = c00 with l 2 -inner product, i.e.,


x, y =  x( j ) y ( j ), x, y ∈ X .
j =1

Consider the linear operator A : X → X defined by

 ∞ x( j ) 
Ax =   e1 ,
 x∈ X.
 j = 1 j 

Note that A ∈ B ( X ), and

1
Aen , e1 = ∀n ∈ N.
n

Now if there exists an operator B : X → X such that

Ax, y = x, By ∀ x ∈ X , y ∈Y ,

then, as a particular case of the above relation, we will have

1
= Aen , e1 = en , Be1 = ( Be1 ) (n) ∀n ∈ N,
n

which is not possible since Be1 must belong to c00 . Thus, in this example,
there exists no linear operator B : X → X such that Ax, y = x, By for all
x, y ∈ X .
58
Unit 10 Operators on Hilbert Spaces
2
However, if we consider the same operator as an operator on l , then the
adjoint does exist. Indeed, for every x, y ∈ l 2 ,

∞ ∞
x( j ) x( j)
Ax, y =  e1 , y =  y (1) = x, By ,
j =1 j j =1 j

where B : l 2 → l 2 is defined by

y (1)
( A* y ) ( j ) = , j ∈ N; y ∈ l 2 .
j

We now give a few more examples of linear operators between infinite


dimensional inner product spaces having adjoints.

Example 2: Let X and Y be C[a, b] or L2 [a, b] with ⋅ 2


.

i) For u ∈ C[a, b], consider the linear operator A : X → X defined by

( Ax)(t ) = u (t ) x(t ), x ∈ X , t ∈ [a, b].

We note that for all x, y ∈ X ,

( A* y )(t ) = u (t )y (t ) ∀y ∈ X , t ∈ [a, b]

which is the adjoint of A.

ii) Suppose k (⋅,⋅) ∈ C ([a, b] × [a, b]). Then we know that for every x ∈ L2 [a, b]
the function Ax defined by

b
( Ax) ( s) =  k ( s, t ) x (t ) dµ(t ), x ∈ L2 [a, b]; s ∈ [a, b],
a

belongs to C[a, b] . A : x → y is bounded linear operator where [a, b] and


x = L2 [a, b] and y = c[a, b]. Note that, for all x, y ∈ L2 [a, b],
b
Ax, y =  ( Ax) ( s) y ( s) dµ( s)
a

b
b 
=    k ( s, t ) x (t ) dµ (t )  y ( s ) dµ( s )
aa 
b
b 
=  x (t )   k ( s, t ) y ( s) dµ( s)  dµ(t ).
a a 

Thus, A* exists and is given by

b
( A y ) ( s) =  k (t , s) y (t ) dµ(t )
*
∀y ∈ Y , s ∈ [a, b].
a
*** 59
Block 3 Hilbert Spaces
You can try this exercise now.

E1) Let X = L2 [a, b], and for φ ∈ L∞ [a, b], let Ax = φx, x ∈ X . Show that the
operator B : X → X defined by Bx = φx, x ∈ X , is the adjoint of A.

E2) Show that the adjoint operation is a bijective map of BL(H ) to itself.

E3) Prove that O* = O and I * = I where O and I denote the zero operator
and identify operator on H .

E4) Let { An } be a sequence of operators in BL(H ) and A ∈ BL(H ) . Prove


that if An − A → 0, then An* An − A* A → 0 and An An* − AA* → 0 as
n → ∞.

We shall now show that if X is a Hilbert space, then every bounded linear
operator A : X → Y has the adjoint. Before that we prove the following result.

Theorem 1: Let X , Y be inner product spaces, and A : X → Y be a linear


operator such that the adjoint A* exists. If A ∈ B ( X , Y ), then A* ∈ B (Y , X ),
and

2
A* = A , A* A = A .

Proof: Suppose A ∈ B ( X , Y ). For every y ∈ Y ,

A* y = sup { A y, u
*
:u∈ X, u ≤1 }
{
= sup }
y, Au : u ∈ X , u ≤ 1

Hence, by continuity of A, we have

A* ∈ B ( X , Y ), A* ≤ A .

Since ( A* )* = A, it follows from the above that

A = ( A* )* ≤ A* .

Thus, we have proved that A* = A .

Also, for every x ∈ X ,


2 2 2
Ax = Ax, Ax = A* Ax, x ≤ A* Ax x ≤ A x .

2
From this we obtain A = A* A .

Now we state the theorem which shows that existence of the adjoint for linear
60 operators.
Unit 10 Operators on Hilbert Spaces
Theorem 2: Let X be a Hilbert space and Y be an inner product space.
Then every A ∈ B ( X , Y ) has the adjoint.

Proof: Let A ∈ B ( X , Y ). We have to show that there exists a linear operator


B : Y → X such that

Ax, y = x, By ∀x ∈ X , y ∈ Y .

For this purpose, for each y ∈ Y , consider the map f y : X → K defined by


f y ( x) = Ax, y , x ∈ X . Clearly, f y ∈ X ′. Hence there exists by the Riesz
representation theorem, a unique u y ∈ X such that f y ( x) = x, u y for all
x ∈ X , i.e.,

Ax, y = x, u y ∀x ∈ X .

It is easily seen that the map B : Y → X defined by

By = u y , y ∈Y

is a linear operator, and it satisfies

Ax, y = x, By ∀x ∈ X , y ∈ Y .

Thus, the adjoint A* exists and is given by A* y = u y for all y ∈ Y .

Remark 1: In the above theorem, although we used completeness of the


space X for the existence of an adjoint of an operator, it is not a necessary
condition. Recall that in Example 2, the space C[a, b] with ⋅ 2 is not a
Hilbert space.

We have already observed that ( A* )* = A for every A ∈ B ( X , Y ). Next we


shall state some more properties of adjoints.

Theorem 3: Let X , Y , Z be Hilbert spaces.

i) For A, A1 , A2 ∈ B ( X , Y ) and α ∈ K,

( A1 + A2 )* = A1* + A2* , (αA)* = α A* .

ii) For A1 ∈ B ( X , Y ) and A2 ∈ B (Y , Z ) ,

( A2 A1 )* = A1* A2*.

The proof is left as an exercise for you to try.

Remark 2: By part (i) in the above theorem and the facts that ( A* )* = A and
A* = A for all A ∈ B ( X , Y ), it is clear that the map A a A* is a conjugate
linear isometry from B ( X , Y ) onto B (Y , X ).

Let us find the adjoint of a few more bounded linear operators. 61


Block 3 Hilbert Spaces
Example 3: i) Let X and Y be infinite dimensional separable Hilbert spaces,
U = {u1 , u 2 ,...} and V = {v1 , v2 ,...} be orthonormal bases of X and Y ,
respectively. Let (λ n ) be a bounded sequence of scalars. For x ∈ X , define

Ax =  λ n x, un vn , x∈ X.
n

Clearly (How?), A ∈ B ( X , Y ). It is easily seen that the adjoint of A is given by

A* y =  λ n y, vn un , y ∈Y.
n

The following example includes the previous one as a particular case.

ii) Let X , Y , U = {u1 , u 2 ,...}, V = {v1 , v2 ,...} be as in (i). Let ( aij ) be an infinite
matrix of scalars, and let

α =  aij , β = sup  aij , γ = sup  aij .


2

ij i j j i

Suppose that min {α, βγ} < ∞. Then, similar to what we have seen in the
beginning of this section,

 
Ax =    aij x, u j vi ,
 x∈ X, (1)
i  j 

belongs to Y for all x ∈ X , and the map A : X → Y is a bounded linear


operator (Verify). It can be easily seen that A* is given by

 
A* y =    a ji y , v j  ui ,
 y ∈Y. (2)
i  j 

Note that

Au j , vi = aij , A*v j , ui = v j , Aui = a ji ∀ i, j ∈ N.


***

In the above example, the operator A is defined by a given matrix ( aij ) which
satisfies certain conditions. The following example shows that every bounded
linear operator A : X → Y between separable Hilbert spaces X and Y can
be represented by a matrix in the sense of (1), and then its adjoint is given by
(2).

Example 4: Let us now look at the adjoints of the right shift and left shift
operators: Let X = l 2 and A be the right shift operators, i.e.,

A : (α1 , α 2 ,...) a (0, α1 , α 2 ,...).


In this case, we have Ae j , ei = δ j +1, i for all i, j ∈ N, and
Ax =  x(i) ei+1 , A* x =  x(i + 1) ei ∀x ∈ l 2 .
i i

Note that A* is the left shift operator (α1 , α 2 ,...) a (α 2 , α 3 ,...).


62
Unit 10 Operators on Hilbert Spaces
* *
Since ( A ) = A, it follows at once that the adjoint of the left shift operator is
the right shift operator.
***

In the following proposition, we list a few results concerning ranges and null
spaces of A and its adjoint which will be useful for deriving more results for
Hilbert space operators.

Proposition 1: Let X and Y be Hilbert spaces, and A ∈ B ( X , Y ) . Then we


have the following:

i) N ( A) = R( A* ) ⊥ .

ii) N ( A* ) = R( A) ⊥ .

iii) N ( A) ⊥ = R ( A* ).

iv) N ( A* ) ⊥ = R ( A).

v) N ( A* A) = N ( A).

Proof: We note that for x ∈ X ,

x ∈ R( A* ) ⊥ ⇔ x, A* y = 0 ∀ y ∈ Y

⇔ Ax, y = 0 ∀ y ∈ Y

⇔ Ax = 0
⇔ x ∈ N ( A).
Thus we have proved (i). The result in (ii) is obtained by replacing A in (i) by
A* . The results in (iii) an (iv) follow from (i) and (ii) by observing that for every
subset S of a Hilbert space, ( S ⊥ ) ⊥ = span S (Verify). For the proof of (v), first
we observe that N ( A) ⊆ N ( A* A). The reverse inclusion is a consequence of
2
the relation Ax = Ax, Ax = A* Ax, x for all x ∈ X .

Here is an exercise for you to try.

E5) Prove Theorem 3.

In the next section we shall discuss some bounded operators which are
related to their adjoints in special ways.

10.3 NORMAL, UNITARY AND SELF-ADJOINT


OPERATORS
In this section we shall introduce you to classes of some more operators.
Since commutativity of elements of BL(H ) usually does not happen, a
bounded operator on H which commutes with its own adjoint is specially
considered.
63
Block 3 Hilbert Spaces
* *
Let A ∈ BL(H ) . Then A is called normal if A A = A A , unitary if
A* A = I = AA* , that is, A* = A−1 , and self-adjoint if A* = A.

Clearly, if A is unitary or self-adjoint, then A is normal. However, a normal


2
operator need be neither unitary nor self-adjoint. For example, let H = K and
for x = ( x(1), x(2)) in H ,

A( x ) = ( x (1) − x(2), x(1) + x (2)).

Then it can be seen that for x ∈ H ,

A* ( x) = ( x(1) + x(2),− x(1) + x(2)),

A* A( x ) = 2( x(1), x(2)) = AA* ( x).

Hence A* A = A A* , but A* A ≠ I and A* ≠ A.

It is easy to see that if B is a normal operator and a C is a bounded operator


such that C *C = I , then the operator A = CBC * is normal. Note that a
bounded operator C may satisfy C *C = I without being a unitary operator.
For example, let H = l 2 and for x = ( x(1), x(2),...) in H , let

C ( x) = (0, x(1), x(2),...)

Then C * ( x) = ( x(2), x (3),...) for x ∈ H , as we have seen in 25.4(a). Hence

C *C ( x) = C * ((0, x(1), x(2),...)) = ( x (1), x(2)....),

CC * ( x) = C (( x(2), x (3),...)) = (0, x(2), x (3),...)

for all x ∈ H , so that C *C = I but CC * ≠ I .

Let A ∈ BL(H ). Since A( x), y = x, A* ( y ) for all x, y ∈ H , we immediately


see that A is normal if and only if A( x ), A( y ) = A* ( x ), A* ( y ) , self adjoint if
and only if A( x ), y = x, A( y ) for all x, y ∈ H . We have seen in earlier that
the inner product , on H characterizes the geometry of H . Hence an
operator A is normal if A and A* transform the geometry of H in the same
fashion, and A is unitary if neither A nor A* change the geometry of H . For
this reason, a unitary operator is known as a Hilbert space isomorphism.

To get a clear idea of these operators, we consider a number of examples.

Example 4: Consider H = L2 ( E ) where E is a measurable subset of R. Fix


z in L∞ (E ) and define

A( x) = zx , x ∈ H .

Then A is an operator on H . Since

A( x ) 2 =  zx dm ≤ z
2 2 2 2

x 2, x∈ H,
64 E
Unit 10 Operators on Hilbert Spaces
We see that A is bounded. (See problem 25.2). For x, y ∈ H ,

y, A* ( x ) = A( y ), x =  zyx− dm = y , −zx
E

Thus A* ( x ) = −
zx for all x ∈ H . Since
2
A* A( x) = z x = AA* ( x ), x ∈ H ,

We see that A is normal. Further, A is unitary if and only if z = 1 is almost


every where in the sense that the set of points for which z ≠ 1 can be
neglected or ignored [measure theoretically speaking it is of measure 0] and
A is self-adjoint if and only if z (t ) ∈ R for almost all. [Here also “almost all”
means the measure of the set of points for which the property does not holds
is of measure 0].
***

Example 5: Let H be a separable Hilbert space and u1 , u 2 ,... constitute an


orthonormal basis for H . Let (k n ) be a bounded sequence of scalars and

A( x ) =  k n x, u n un , x ∈ H .
n

A ∈ BL(H ) and

A* ( x ) =  k n x, un un , x ∈ H .
n

Since for all x ∈ H ,

A* ( A( x)) =  k n A( x), un un =  k n
2
x, un u n ,
n n

A( A* ( x)) =  k n A* ( x ), un un =  k n
2
x, u n un ,
n n

we see that A is normal. Further, A is unitary if and only if k n = 1 for each n


and A is self-adjoint if and only if k n ∈ R for each n.
***

More generally, let A ∈ BL(H ) be defined by the matrix M = ( ki , j ) with


respect to the orthonormal basis u1 , u 2 ,.... Then ki , j = A(u j ), ui , i = 1, 2,... We
t
have seen in earlier that A* ∈ BL( H ) is defined by the matrix M = ( k j ,i ) with
respect to u1 , u 2 ,... . Hence A( x ) =  (
i k
j i, j )
x, u j ui and
A* ( x ) =  i ( j )
k j ,i x, u j ui for all x ∈ H . Thus we see that A is self-adjoint
t
if and only if M = M , that is, the matrix M is conjugate-symmetric. Next,
A* A ∈ BL( X ) is defined by the matrix M M =
t
( n )
k n, i k n, j with respect to
u1 , u 2 ,...,

since
65
Block 3 Hilbert Spaces

A* A(u j ), ui = A(u j ), a(ui )

=  n
A(u j ), u n un ,  A(ui ), u m um
m

=  A(u j ), un A(ui ), un
n

=  k n, j k n ,i .
n

for all i, j = 1, 2,... . Similarly, A A* ∈ BL( H ) is defined by the matrix


MM =
t
( n )
ki , n k j , n with respect to u1 , u 2 ,... . Hence we see that A is unitary
t t
if and only if M M = I = M M , that is,

k
n
k
n, i n , j = δ i , j =  ki , n k j , n
n

for all i, j = 1, 2,... . This says that the columns of M form an orthonormal set
2 t t
in l , and so do its rows. Next, A is normal if and only if M M = M M . This
is certainly the case if M is a diagonal matrix. For connections between the
normality of A and the representability of A by a diagonal matrix see following
result. The proof is omitted.
***

Theorem 4: Let u1 , u2 , … constitute an orthonormal basis for H . Suppose that


A ∈ BL(H ) is defined by a matrix M with respect to u1 , u 2 ,...
a) Assume that M is triangular. Then A is normal if and only if M is
diagonal.
t
b) Assume that M = N D N , where the matrices N and D define
bounded operators on H , the columns of N form an orthonormal set in
l 2 and D is diagonal. Then A is normal. [The converse holds if K = C
and H is finite dimensional.

Example 6: Let H = K 2 and A ∈ BL(H ) be given by

A( x (1), x(2)) = (ax(1) + bx (2), cx(1) + dx(2)),

where a, b, c and d are fixed scalars. The following gives precise conditions
on a, b, c, d for A to be self-adjoint, unitary or normal.

2 2
a) If K = C (resp., R ) , then A is normal if and only if b = c and
(a − d ) c = (a − d ) b (resp., either b = c, or b = −c and a = d ).
2 2 2 2
b) A is unitary if and only if a + b = 1 = c + d and ac + b d = 0 (resp.,
either a = cos θ = d and b = − sin θ = −c, or a = cos θ = − d and
b = sin θ = c for some 0 ≤ θ < 2π).

66
c) A is self-adjoint if and only if a, d ∈ R and c = b .
Unit 10 Operators on Hilbert Spaces
We now consider some important properties of self-adjoint, unitary and normal
operators.

Theorem 5: Let H be a Hilbert space and A ∈ BL(H ).

a) Let A be self-adjoint. Then

A = sup { }
A( x), x : x ∈ H , x ≤ 1 .

In particular, A = 0 if and only if A( x), x = 0 for all x ∈ H .

b) A is unitary if and only if A( x) = x for all x ∈ H and A is surjective.


In that case, for all x ∈ H and

A = 1 = A −1

c) A is normal if and only if A( x ) = A* ( x ) for all x ∈ H . In that case,

2
A2 = A* A = A

Proof: a) We have seen in earlier that for every A ∈ BL(H ),

A = sup { }
A( x ), y : x, y ∈ H , x ≤ 1, y ≤ 1 .

Let α = sup { }
A( x ), x : x ∈ H , x ≤ 1 . Clearly, α ≤ A . To prove
A ≤ α, we note that for x, y ∈ H ,

A( x + y ), x + y − A( x − y ), x − y = 2 A( x), y + 2 A( y ), x

= 4 Re A( x), y ,
since A is self-adjoint. Hence

( 2
4 Re A( x), y ≤ α x + y + x − y
2
) = 2α( x 2
+ y
2
)
by the parallelogram law. Let x ≤ 1 and y ≤ 1. Then it follows that
Re A( x ), y ≤ α. If A( x ), y = re iθ for r ≥ 0 and θ ∈ R, then let
x0 = e −iθ x, so that x0 = x ≤ 1 and

A( x), y = r = A( x0 ), y = Re A( x0 ), y ≤ α.

Taking supremum over all x, y ∈ H with x ≤ 1, y ≤ 1, we obtain


A ≤ α, as desired.

In particular, if A( x), x = 0 for all x ∈ H , then A = α = 0, that is,


A = 0. The converse is obvious.

b) For x ∈ H , we have 67
Block 3 Hilbert Spaces
2 2
A( x ) − x = A( x), A( x) − x, x

= A* A( x), x − x, x

(
= A* A − I x, x)
Since A* A − I is self-adjoint, it follows from part (a) that A* A = I if and
only if A( x) = x for all x ∈ H .

Thus if A( x) = x for all x ∈ H and A is surjective, then A* A = I and


A is bijective, so that

AA* = ( AA* ) ( AA−1 ) = A( A* A) A−1 = AA−1 = I ,

that is, A is unitary.

Conversely, if A is unitary, then A* A = I and A−1 = A* , so that


A( x) = x for all x ∈ H and A is surjective. In that case, it follows that
A−1 ( x) = x for all x ∈ H . Taking supremum over all x ∈ H with x ≤ 1,
we obtain A = 1 = A−1 .

c) For x ∈ H , we have
2 2
A( x ) − A* ( x) = A( x), A( x ) − A* ( x), A* ( x )

= A* A( x), x − AA* ( x), x

= ( A* A − AA* ) ( x ), x .

Since B = A* A − AA* is self-adjoint, it follows from part (a) that B = 0


(that is, A is normal) if and only if A( x ) = A* ( x) for all x ∈ H . In that
case, it follows that

A2 ( x) = A( A( x) = A* A( x)

2
for all x ∈ H . Hence A 2 = A* A = A by Theorem 1.

We now investigate whether sums, compositions and limits of self-adjoint,


unitary and normal operators are respectively self-adjoint, unitary and normal.

Theorem 6: Let H be a Hilbert space.

a) Let A and B be self-adjoint. Then A + B is self-adjoint. Also, AB is


self-adjoint if and only if A and B commute.

b) Let A and B be unitary. Then AB is unitary. Also, A + B is unitary if


and only if it is surjective and Re A( x), B ( x) = −1 / 2 for every x ∈ H
with x = 1.

c) Let A and B be normal. If A commutes with B * and B commutes with


68 A* , then A + B and AB are normal.
Unit 10 Operators on Hilbert Spaces
d) Let ( An ) be a sequence in BL(H ), and A ∈ BL(H ) be such that
An − A → 0 as n → ∞. If each An is self-adjoint, unitary or normal,
then A is self-adjoint, unitary or normal, respectively.

* * *
Proof: a) Since ( A + B ) = A + B = A + B, we see that A + B is self-adjoint.
* * *
Also, since ( AB) = B A = BA, we see that AB is self-adjoint if and
only if BA = AB.

b) Since ( AB)* AB = B * A* AB = B * B = I and AB ( AB)* = ABB* A* = I , we


see that AB is unitary. Also, for all x ∈ H , we have

( A + B ) ( x), ( A + B ) ( x) = A( x), A( x ) + B ( x), B ( x)


+ A( x ), B ( x ) + B ( x), A( x)
= 2 x, x + 2 Re A( x), B ( x) .

Hence Theorem 5(b) implies that A + B is unitary if and only if it is


surjective and x, x + 2 Re A( x), B ( x) = 0 for all x ∈ H .

c) Let AB * = B * A and A* B = BA* . Then

( A + B)* ( A + B) = A* A + B* B + A* B + B* A

= AA* + BB* + BA* + AB*

= ( A + B ) ( A + B )* ,

( AB)* AB = B* A* AB = B* AA* B
= AB* BA* = ABB* A*
= AB(AB)* .

Hence A + B and AB are normal.

d) Since A* = A , An* − A* = An − A → 0. If each An is self-adjoint, that


is, An* = An , then ( An ) converges to both A and A* in BL(H ), so that
A* = A, that is, A is self-adjoint. Similar reasoning holds if each An is
unitary or if each An is normal. Note that An* An − A* A → 0 and
An An* − AA* → 0.

In view of the theorem above, let us ask the following question. If A and B
are normal operators and A commutes with B then must A + B and AB be
normal? This amounts us to find out whether A commutes with B * , and B
commutes with A* . This is indeed the case. In fact, has been proved that if
K = C , A is a normal operator on a complex Hilbert space H and A
commutes with a bounded B on H , then A commutes with B * as well.

Have you observed an analogy between the complex numbers and the
bounded operators on a complex Hilbert space H , with the adjoint operation
69
Block 3 Hilbert Spaces
playing the role of complex conjugation? You are right. Then self-adjoint
operators correspond to real numbers and unitary operators correspond to
complex numbers of absolute value 1. Since a normal operator commutes (at
least) with its adjoint, it seems appropriate to let normal operators correspond
to complex numbers. This is made precise in the following result.

Theorem 7: Let K = C and A ∈ BL(H ). Then there are unique self-adjoint


operators B and C on H such that

A = B + iC

Further, A is normal if and only if BC = CB , A is unitary if and only if


BC = CB and B 2 + C 2 = I , and A is self-adjoint if and only if C = 0.

A + A* A − A*
Proof: Let B = and C =
2 2i

Then B and C are self-adjoint operators and A = B + iC. If we also have


A = B1 + iC1 , where B1 and C1 are self-adjoint, then A* = B1 − iC1 , so that

A + A* A − A*
B1 = = B and C1 = = C.
2 2i

Thus B and C are unique.

Now A is normal if and only if


( B − iC ) ( B + iC ) = A* A = AA* = ( B + iC ) ( B − iC ), that is, BC = CB. Similarly,
A is unitary if and only if ( B − iC ) ( B + iC ) = I = ( B + iC ) ( B − iC ), that is,
( B 2 + C 2 ) + i( BC − CB) = I = ( B 2 + C 2 ) − i ( BC − CB). It can be easily seen
that this is equivalent to B 2 + C 2 = I and ( BC − CB ) = 0. Finally, A is self-
adjoint if and only if B − iC = B + iC , that is, C = 0.

Here we have considered three important chains of operators: Self-adjoint


operators, normal operators and unitary operators.

You can observe that for any A ∈ BL( X ) , the operators A* A and A* A are
self-adjoint operators. Also, the identity operator is self-adjoint, and self-adjoint
operators and unitary operators are normal. But a normal operator need not be
unitary or self-adjoint. Here is an example to illustrate this.

Example 7: Let X = C 2 . For α1 , α 2 ∈ C , let A : X → X be given by

A( z1 , z 2 ) = (α1 z1 , α1 z 2 ), z1 , z 2 ∈ C .

Then it can be shown that

i) A is a normal operator.
ii) A is self-adjoint if and only if α1 , α 2 ∈ R.

iii) A is unitary if and only if α1 = 1 = α 2 .

70 ***
Unit 10 Operators on Hilbert Spaces
You can now try the following exercises.

E6) Let { An } be a sequence of unitary operators in BL(H ). Prove that if


An − A → 0, A ∈ BL( H ), then A is unitary.

E7) If A ∈ BL(H ) is normal operator on a Hilbert space H and λ ∈ k , prove


that A − λI is also normal.

E8) If H is a finite dimensional vector space, then show that every isometric
isomorphism of H into itself is unitary.

With this we come to the end of this unit.

10.4 SUMMARY
In this unit we have covered the following:

1. We have introduced the notion of the adjoint of an operator.

2. We have defined the following operators using adjoint of an operator


*
i) self-adjoint operator: A = A
ii) normal operator: AA* = A* A.
iii) unitary operator: AA* = I = A* A.

We have studied some special properties of these operators.

10.5 SOLUTIONS/ANSWERS

E1) This follows from the innerproduct in L2 [a, b] given by

f , g =  f g dx

Af , g =  Af g dx

=  φf g dx

f , Ag =  f φ g dx

=  φ f g dx

A* f , g = f , Ag =  φ f g dx

Therefore A* f = φ f

E2) If φ : BL( H ) → BL( H ) is a map


φ( A) = A* , then
for A1 , A2 ∈ BL( H ) ,
71
Block 3 Hilbert Spaces
φ( A1 ) = φ( A2 )  φ( A1 ) = φ( A2 )
* *

 A1** = A2**
 A1 = A2

Also, if A ∈ BL(H ) then A* ∈ BL( H ) and

φ( A* ) = A** = A
Thus, φ is both one-to-one and onto.

E3) Ox, y = O, y = 0
Also Ox, y = x, O * y = 0
By uniqueness O * = 0
Also Ix, y = x, I * y
But Ix = x.
∴ x, y = x, I * y
*
Therefore I = A (By uniqueness of adjoint)

E4) Firstly An* − A* = ( An − A)* = An − A . This implies that An − A → 0 if


* *
and only if An − A → 0. So

An* An − A* A = An* An − A* An + A* An − A* A

≤ An An* − A* + A* An − A

→ 0.
Similarly it can be shown that An An* − A A* → 0 as n → ∞.

E5) For x, y ∈ H

( A1 + A2 ) x, y = A1 x, y + A2 x, y

= x, A1* y + x, A2* y

= x, ( A1* + A2* ) ( y )

∴By uniqueness in the definition of an adjoint, we have


( A1 + A2 )* = ( A1* + A2* )

Similarly, ( k A) x, y = k Ax, y

= x, (kA)* y

= x, k A* y

∴ By uniqueness (k A) ⊥ = k A*

( A1 A2 ) x, y = A1 ( A2 x ), y
72
Unit 10 Operators on Hilbert Spaces
*
= A2 x, A y 1

= x, A2* A1* y

∴ ( A1 A2 )* = A2* A1* .

E6) An* An − A* A → 0 as n → ∞. Since An is unitary for every n, An* An = I .


Therefore A* A = I . Similarly AA* = I .

E7) Since A is normal A* A = AA*

Now,
( A − λI )* = A* − λ I *

= A* − λ I
Hence
( A − λI )* ( A − λ I ) = ( A* − λ I )( A − λ I )
2
= A* A − λ A + λ I − λA*

( A − λI ) ( A − λI )* is equal to the same expression since A* A = AA*.

E8) Since H is finite dimensional, any infective homomorphism is also


surjective. Thus, if A ∈ BL(H ) is unitary of L , A is unitary.

73

You might also like