Lecture Notes-Functional Analysis

Dr. Sumit Chandok

School of Mathematics
TIET

2020
Contents

0.1 Vector Space or Linear Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

0.2 Linear Subspace . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
0.3 Linear Dependent/Linear Independent . . . . . . . . . . . . . . . . . . . . . . . . . . 8
0.4 Some Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
0.5 Metric Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
0.6 Basic Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
0.7 Some properties in metric space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
0.8 Normed Vector Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
0.9 Convex Set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
0.10 Quotient Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
0.11 Completeness property in normed linear space . . . . . . . . . . . . . . . . . . . . . 31
0.12 Convergence of series in Banach space . . . . . . . . . . . . . . . . . . . . . . . . . 37
0.13 Cantor intersection theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
0.14 Separable Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
0.15 Finite Dimensional Normed Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
0.16 Equivalent Norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
0.17 Compactness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
0.18 Banach Fixed Point Theorem and Applications . . . . . . . . . . . . . . . . . . . . . 50
0.18.1 Linear Algebraic Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
0.18.2 Existence Theorem on Differential equation (Picard’s theorem) . . . . . . . . 52
0.19 Inner Product and Hilbert Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
0.20 Orthogonal and Orthonormal vectors . . . . . . . . . . . . . . . . . . . . . . . . . . 62
0.21 Best Approximation in Hilbert Space . . . . . . . . . . . . . . . . . . . . . . . . . . 65
0.22 Some results in Inner Product Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 70
0.23 Dual of a Hilbert space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
0.24 Orthogonal Basis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
0.24.1 Gram-Schmidt Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
0.25 Some Examples (IPS) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
0.26 Total Orthonormal sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
0.27 Adjoint Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
0.28 Self Adjoint, Normal and Unitary operators . . . . . . . . . . . . . . . . . . . . . . . 88
0.29 Linear Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
0.30 Bounded Linear Transformation-I . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
0.31 Bounded and Continuous Linear Operators-2 . . . . . . . . . . . . . . . . . . . . . . 96
0.32 Extension of Linear Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
0.33 Linear Functionals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
0.34 Dual Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
0.35 Examples of Dual Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
0.36 Linear Functionals and Hahn-Banach Theorem . . . . . . . . . . . . . . . . . . . . . 110
0.37 Application of Hahn Banach Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 116

1
CONTENTS

0.38 Uniform Boundedness Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119

0.39 Open Mapping Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
0.40 Closed Graph Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
0.41 Projection Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
0.42 Compact linear operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
0.43 Strong Convergence and Weak Convergence . . . . . . . . . . . . . . . . . . . . . . . 131
0.44 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134

2
0.1. VECTOR SPACE OR LINEAR SPACE

0.1 Vector Space or Linear Space

Vector spaces play a role in many branches of mathematics and its applications. In fact, in
various practical (and theoretical) problems we have a set X whose elements may be vectors
in three-dimensional space, or sequences of numbers, or functions, and these elements can
be added and multiplied by constants (numbers) in a natural way, the result being again an
element of X. Such concrete situations suggest the concept of a vector space. The definition
will involve a general field K, but in functional analysis, K will be R or C.
Definition 0.1.1. A field is a set F with two operations called addition and multiplication which
satisfy the following field axioms:
Axioms for addition:
(a1) if x ∈ F and y ∈ F, then x + y ∈ F
(a2) x + y = y + x, ∀x, y ∈ F
(a3) (x + y) + z = x + (y + z), ∀x, y, z ∈ F
(a4) ∃0 ∈ F such that 0 + x = x = x + 0, ∀x ∈ F
(a5) To every x ∈ F corresponds an element −x ∈ F such that x + (−x) = 0.
Axioms for multiplication
(m1) if x, y ∈ F, then xy ∈ F
(m2) xy = yx for all x, y ∈ F
(m3) (xy)z = x(yz)∀x, y, z ∈ F
(m4) F contains an element 1 6= 0 such that 1x = x = x1, ∀x ∈ F
1
(m5) If x ∈ F and x 6= 0, then there exists an element x ∈ F such that x( 1x ) = 1 = ( 1x )x.
Distributive law (d1) x(y + z) = xy + xz, ∀x, y, z ∈ F.
For example, Q is a field. The set of all rational numbers satisfy all these axioms.
Definition 0.1.2. A number system is said to be algebraic structure if it is closed with respect
to addition and multiplication, that is, for every x, y ∈ A, x + y ∈ A and xy ∈ A.
Example 0.1.1. • N is an algebraic structure.
• Z is an algebraic structure.
• Q is an algebraic structure.
• R is an algebraic structure.
• Irrational number R√− Q is√not algebraic structure because it is not closed under ’+’ and
’.’, For example: ( 2).(− 2) = −2 ∈ / R − Q.
4 17
• R − N is not an algebraic structure, as 17 . 4 =1∈
/ R − N.
Definition 0.1.3. Let K = R or C (or more generally a field). A vector space over K is a
nonempty set of elements called vectors X on which two operations, vector addition + : X ×
X → X and scalar multiplication . : X × X → X are defined, such that the following assertions
are true for all x, y, z ∈ X and r, s ∈ K.

3
0.1. VECTOR SPACE OR LINEAR SPACE

1. x + y ∈ X
2. x + y = y + x
3. x + (y + z) = (x + y) + z
4. There exists an element 0 (called zero vector) such that x + 0 = 0 + x = x
5. For every x ∈ X, there exists an element −x, such that x + (−x) = 0 = (−x) + x
6. rx ∈ X
7. r(x + y) = rx + ry
8. (r + s)x = rx + sx
9. r(sx) = (rs)x
10. for all x ∈ X, 1.x = x.
Example 0.1.2. • R is a vector space over R with vector addition being the usual addition
of real numbers, and scalar multiplication being the usual multiplication of real numbers.
• Rn is a vector space over R, with addition and scalar multiplication defined as follows:
x1 y1 x1 y1 x1 + y1 x1
           
.
. .
. n
if . , . ∈ R , then . + . = .
. .
. .
. , if α ∈ R and ...  ∈ Rn , then
         .  
x y xn yn xn + yn xn
 n n 
x1 αx1
.
α .. = ..  .
   .
xn αxn
• The function space C[a, b]. Let a, b ∈ R and a < b. Consider the vector space comprising
functions f : [a, b] → K that are continuous on [a, b], with addition and scalar multipli-
cation defined as follows. If f , g ∈ C[a, b], then f + g ∈ C[a, b] is the function given
by
( f + g)(x) = f (x) + g(x), x ∈ [a, b].
If α ∈ K and f ∈ C[a, b], then α f ∈ C[a, b] is the function given by (α f )(x) = α f (x),
x ∈ [a, b].
C[a, b] is referred to as a ‘function space’, since each vector in C[a, b] is a function (from
[a, b] to K).
• The sequence space s. Let s denote the set of all sequences x = {xn }∞ 1 of real or complex
numbers. Define the operations of addition and scalar multiplication pointwise: For
all x = (x1 , x2 , ...), y = (y1 , y2 , ...) ∈ s and all α ∈ K, define x + y = (x1 + y1 , ...) and
αx = (αx1 , ...). Then s is a linear space over K.
• The sequence space ` p , 1 ≤ p < ∞. Let ` p denote the set of all sequences x = {xn }∞1 of
real or complex numbers satisfying the condition ∑∞ |x | p < ∞, with addition and scalar
i=1 i
multiplication defined as follows:
1 + (yn )1 = (xn + yn )1 , (xn )1 , (yn )1 ∈ ` p ;
(xn )∞ ∞ ∞ ∞ ∞

1 = (αxn )1 , α ∈ K, (xn )1 ∈ ` p . Then ` p is a vector space.

α(xn )∞ ∞ ∞

4
0.1. VECTOR SPACE OR LINEAR SPACE

Proof. Let x = (x1 , x2 , ...), y = (y1 , y2 , ...) ∈ ` p . Now, we show that x + y ∈ ` p . Since for
each i ∈ N, we have

|xi + yi | p ≤ [2 max{|xi |, |yi |}] p

≤ 2 p max{|xi | p , |yi | p }
≤ 2 p (|xi | p + |yi | p ).

It follows that x + y ∈ ` p . Also if x = (xn ) ∈ ` p and α ∈ K, then

∞ ∞
∑ |αxi| p = |α| p ∑ |xi| p < ∞,
i=1 i=1

that is αx ∈ ` p .
• The sequence space `∞ . Let `∞ denote the vector space of all bounded sequences with
values in K, that is supi∈N |xi | < ∞ and with addition and scalar multiplication defined as
follows:
(xn )n∈N + (yn )n∈N = (xn + yn )n∈N , (xn )n∈N , (yn )n∈N ∈ `∞ ;
α(xn )n∈N = (αxn )n∈N , α ∈ K, (xn )n∈N ∈ `∞ .
• The sequence space c = c(N) Let c denote the set of all convergent sequences x = (xn )∞ 1
of real or complex numbers. That is, c is the set of all sequences x = (xn )∞ 1 such that
lim xn exists. Define the operations of addition and scalar multiplication pointwise as in
n→∞
previous example. Then c is a linear space over K.
• The sequence space c0 = c0 (N). Let c0 denote the set of all sequences x = (xn )∞
1 of real
or complex numbers which converge to zero. That is, c0 is the space of all sequences x =
(xn )∞ lim xn = 0. Define the operations of addition and scalar multiplication
1 such that n→∞
pointwise as in previous example. Then c0 is a linear space over field K.

5
0.2. LINEAR SUBSPACE

0.2 Linear Subspace

A subspace of a vector space X is a nonempty subset Y of X such that for all y1 , y2 ∈ Y and all
scalars α, β we have
αy1 + β y2 ∈ Y.
Hence Y is itself a vector space, the two algebraic operations being those induced from X.
Clearly, a subset M of a linear space X is a linear subspace if and only if M + M ⊂ M and
αM ⊂ M for all α ∈ K.
Example 0.2.1. • Every linear space X has at least two distinguished subspaces: M = {0}
and M = X. These are called the improper subspaces of X. All other subspaces of
X(6= {0}) are called the proper subspaces.
• Let X = R2 . Then the nontrivial linear subspaces of X are straight lines through the
origin.
• M = {x = (0, x2 , x3 , ..., xn ) : xi ∈ R, i = 1, ..., n} is a subspace of Rn .
• M = {x : [−1, 1] → R, i = 1, ..., n, x continuous and x(0) = 0} is a subspace of C[−1, 1].
• M = {x : [−1, 1] → R, i = 1, ..., n, x continuous and x(0) = 1} is not a subspace of C[−1, 1].
• The set of vectors in Rn of the form (x1 ; x2 ; x3 ; 0; ...; 0) forms a three-dimensional linear
subspace.
• The set of polynomials of degree ≤ r forms a linear subspace of the the set of polynomials
of degree ≤ n for any r ≤ n.
Example 0.2.2. 1. Suppose that V3 is a vector space. Check which of the following is a
subspace of V3 ?
[i] W = {(a, b, c) : ab < 0} is not subspace as W1 = (1, −1, 2),W2 = (−1, 1, 3) ∈ W
But W1 +W2 ∈ / W.
[ii] W = {(a, b, c) : a ≥ b ≥ c} is not subspace as W1 = (3, 2, 1) ∈ W take α = −1
But αW1 ∈ / W.
[iii] W = {(a, b, c) : b = a2 } is not subspace as W1 = (1, 1, 2),W2 = (2, 4, 3) ∈ W But
W1 +W2 ∈ / W.
√
[iv] W = {(a, b, c) : b = a} is not subspace as W1 = (1, 1, 2),W2 = (4, 2, 3) ∈ W But
W1 +W2 ∈ / W.
[v] W = {(a, b, c) : a2 + b2 + c2 ≤ 1} is not subspace as W1 = (1/2, 1/3, 1/4),W2 =
(1/2, 1/4, 1/3) ∈ W But W1 +W2 ∈ / W.
[vi] W = {(a, b, c) : b+4c = 0} is a subspace as W1 = (a1 , b1 , −b1 /4),W2 = (a2 , b2 , −b2 /4) ∈
W . Here W1 +W2 = (a1 + a2 , b1 + b2 , −(b1 + b2 )/4) ∈ W and αW1 ∈ W for any α.
[vii] W = {(a, 2a, a+1) : a ∈ R} is not subspace as W1 = (1, 1, 2),W2 = (−1, −2, 0) ∈
W But W1 +W2 = (0, −1, 2) ∈ / W.
Example 0.2.3. Let V be a vector space of all the polynomials of degree ≤ 6. Which of the
following are subspaces of V ?
[i] W = { f (x) : deg f (x) = 4} is not a subspace of V as

f1 (x) = −x4 + bx3 + cx2 + dx + e ∈ W,

6
0.2. LINEAR SUBSPACE

and
f2 (x) = x4 + b1 x3 + c1 x2 + d1 x + e1 ∈ W
but f1 + f2 ∈ / W.
[ii] W = { f (x) : coeff of x2 = 2 or − 2} is not a subspace of V as f1 (x) = ax4 + bx3 +
2x2 + dx + e, f2 (x) = a1 x4 + b1 x3 − 2x2 + d1 x + e1 ∈ W but f1 + f2 ∈
/ W.
[iii] W = { f (x) : f (0) = 1} is not a subspace of V . Take f1 , f2 ∈ W such that f1 (0) =
1, f2 (0) = 1. Here ( f1 + f2 )(0) = 2 ∈/ W.
[iv] W = { f (x) : f (x) has negative coefficients} is not a subspace of V take α = −1
then α f ∈ / W.
[v] W = { f (x) : deg f (x) ≤ 3} is a subspace of V . Take f1 , f2 ∈ W such that degree of
f1 , f2 is ≤ 3. So α f1 + β f2 ∈ W for any α, β .
Definition 0.2.1. A linear combination of vectors x1 , x2 ..., xm of a vector space X is an expres-
sion of the form α1 x1 + α2 x2 + ... + αm xm , where the coefficients α1 , ..., αm are any scalars.
Definition 0.2.2. For any nonempty subset M ⊂ X, the set of all linear combinations of vectors
of M is called the span of M, written span M. [M] = {α1 x1 +α2 x2 +...+αm xm : αi ∈ F, ui ∈ M}
Obviously, this is a subspace Y of X, and we say that Y is spanned or generated by M.
Example 0.2.4. i) Let V3 = {(x1 , x2 , x3 ) : xi ∈ R, i = 1, 2, 3} be a vector space and M =
{(1, 0, 0), (1, 1, 1)} be any subset of V3 .
The span of set M is given by

[M] = {α(1, 0, 0) + β (1, 1, 1) : α, β are scalars}

= {(α, 0, 0) + (β , β , β ) : α, β are scalars}
= {(α + β , β , β ) : α, β are scalars}.

ii) Let P4 be the set of all polynomials of degree ≤ 4. Consider M = {x3 + 1, x + x2 , x4 } be

any subset of P4 . Then

[M] = {α(x3 + 1) + β (x + x2 ) + γx4 : α, β , γ are scalars}

= {γx4 + αx3 + β (x2 + x) + α : α, β , γ are scalars}

Remark 0.2.1. 1. [M] is always subspace of vector space.

2. [M] is the smallest subspace of vector space containing M.
3. If M is a subspace of a vector space, then [M] = M.
Example 0.2.5. Check whether (1, 3, −5) ∈ [M],
where M = {(1, 2, 1), (1, 1, −1), (4, 5, −2)}.

(1, 3, −5) = a(1, 2, 1) + b(1, 1, −1) + c(4, 5, −2)

a + b + 4c = 1,
2a + b + 5c = 3, subtracting − a − c = −2,
a − b − 2c = −5, adding 3a + 3c = −2

But these not satisfied, so (1, 3, −5) do not belong to linear span generated by S.

7
0.3. LINEAR DEPENDENT/LINEAR INDEPENDENT

0.3 Linear Dependent/Linear Independent

Definition 0.3.1. Linear independence and dependence of a given set M of vectors x1 , ..., xr ,
(r ≥ 1) in a vector space X are defined by means of the equation

α1 x1 + α2 x2 + ... + αr xr = 0, (0.3.1)

where the coefficients α1 , ..., αr are any scalars.

Clearly, equation (0.3.1) holds for α1 = ... = αr = 0. If this is the only r-tuple of scalars for
which (0.3.1) holds, then set M is said to be linearly independent.
M is said to be linearly dependent if M is not linearly independent, that is, if (0.3.1) also
holds for some r-tuple of scalars, not all zero scalars.
An arbitrary subset M of X is said to be linearly independent if every nonempty finite
subset of M is linearly independent. M is said to be linearly dependent if M is not linearly
independent.
A motivation for this terminology results from the fact that if M = {x1 , ..., xr } is linearly
dependent, at least one vector of M can be written as a linear combination of the others; for
instance, if (0.3.1)) holds with an αr 6= 0, then M is linearly dependent and we may solve
(0.3.1) for x, to get
αj
xr = β1 x1 + ... + βr−1 xr−1 , (β j = − ).
αr
Example 0.3.1. i) Let V3 be a vector spce and M = {(1, 0, 0), (1, 1, 0)} be any subset of V3 .
Check whether M is linearly independent or not.
Solution. Consider a(1, 0, 0) + b(1, 1, 0) = (0, 0, 0). It implies that

a+b = 0

and
b = 0.
So a = 0 = b. Therefore M is linearly independent.
ii) Check whether set M = {x, |x|} is linearly independent or not in C (−1, 1) and C (1, 4).
Solution. Consider ax + b|x| = 0.
Case 1. If x ∈ (0, 1), then ax + bx = 0. As x 6= 0 so

a + b = 0. (0.3.2)

Case 2. If x ∈ (0, 1), then ax − bx = 0. As x 6= 0 so

a − b = 0. (0.3.3)

On solving these two equations, we get a = 0 = b. Therefore M is linearly independent

in C (−1, 1).
Now, we have to check whether M = {x, |x|} is linearly independent or not in C (1, 4).
Here M is linearly dependent because in this case |x| will take only +ve value. Thus
a + b = 0, as x 6= 0.

8
0.3. LINEAR DEPENDENT/LINEAR INDEPENDENT

iii) If u, v, w are linearly independent vectors of vector space, then {u + v, v + w, w + u} is also

linearly independent.
Solution. Consider
a(u + v) + b(v + w) + c(w + u) = 0
u(a + c) + v(a + b) + v(b + c) = 0.u, v, ware L.I. a+c = 0
a+b = 0 c−b = 0
b+c = 0 b = 0 = c, a = 0.
Thus {u + v, v + w, w + u} is linearly independent.
We can use the concepts of linear dependence and independence to define the dimension of
a vector space.
Definition 0.3.2. A subset M of a vector space X is said to be basis for X if
(1) M is linearly independent set.
(2) [M] = X, that is M generates X.
Example 0.3.2. M = {(1, 0, 0), (0, 1, 0), (0, 0, 1)} is basis for V3 .
Solution. Here M is L.I. set. As a(1, 0, 0) + b(0, 1, 0) + c(0, 0, 1) = 0, implies a = 0 = b = c.
Take x = (x1 , x2 , x3 ) ∈ V3 . So (x1 , x2 , x3 ) = a(1, 0, 0) + b(0, 1, 0) + c(0, 0, 1) and x1 = a, x2 =
b, x3 = c. So [M] = V3 . Thus M is a basis for V3 .
Definition 0.3.3. (Finite and infinite dimensional linear spaces). A vector space X is said to be
finite dimensional if there is a positive integer n such that X contains a linearly independent set
of n vectors whereas any set of n + 1 or more vectors of X is linearly dependent.
The number n is called the dimension of X, written n = dim X.
By definition, X = {0} is finite dimensional and dim X = 0. If X is not finite dimensional,
it is said to be infinite dimensional.
Example 0.3.3. Let Pn be the set of all polynomials of degree ≤ n. Pn = {a0 + a1 x + . . . +
an xn : ai ∈ R} = [{1, x, . . . , xn }]. It has n + 1 elements which are L.I. So dim Pn = n + 1.
In analysis, infinite dimensional vector spaces are of greater interest than finite dimensional
ones.
For instance, C[a, b] and `2 are infinite dimensional, whereas Rn and Cn are n-dimensional.
If dim X = n, a linearly independent n-tuple of vectors of X is called a basis for X (or a
basis in X). If {e1 , ..., en } is a basis for X, every x ∈ X has a unique representation as a linear
combination of the basis vectors:
x = α1 e1 + ... + αn en .
For example, a basis in Rn is
e1 = (1, 0, 0, , 0)
e2 = (0, 1, 0, ..., 0)
...............
en = (0, 0, 0, ..., 1).
The space Rn has dimension n. This is sometimes called the canonical basis for Rn .
More generally, if X is any vector space, not necessarily finite dimensional, and B is a
linearly independent subset of X which spans X, then B is called a basis (or Hamel basis) for
X. Hence if B is a basis for X, then every nonzero x ∈ X has a unique representation as a linear
combination of (finitely many!) elements of B with nonzero scalars as coefficients.

9
0.4. SOME INEQUALITIES

0.4 Some Inequalities

This section deals with some fundamental inequalities that appears frequently in functional
analysis.

Lemma 0.4.1. If a and b are real or complex numbers and p ≥ 1. Then

|a + b| p ≤ 2 p (|a| p + |b| p ).

Proof. Case 1: Take |a| > |b|. So we have |a + b| ≤ |a| + |b| < 2|a| and |a + b| p < 2 p |a| p ≤
2 p (|a| p + |b| p ).
Case 2. |a| ≤ |b|. So we have |a + b| ≤ |a| + |b| ≤ 2|b| and |a + b| p ≤ 2 p |b| p . Hence
|a + b| p ≤ 2 p (|a| p + |b| p ).

Lemma 0.4.2. (Cauchy-Schwarz Inequality) If ai , bi are nonnegative real numbers then

n n 1
n 1
∑ |aibi| ≤ ( ∑ |ai|2) 2 ( ∑ |bi|2) 2 .
i=1 i=1 i=1

Here equality occurs if ai and bi are proportional, that is, ai = cbi , where c is constant.

Proof. Taking identical terms, rearranging and expanding, we get

n n
0≤ ∑ ∑ (aib j − a j bi)2
i=1 j=1
n n n n n n
=∑ ∑ a2i b2j + ∑ ∑ b2i a2j − 2 ∑ ∑ (aib j )(a j bi)
i=1 j=1 i=1 j=1 i=1 j=1
n n n
2
∑ 2
= 2( ai )( bi ) − 2( ai bi )2 .
∑ ∑
i=1 i=1 i=1
1 1
Hence ∑ni=1 |ai bi | ≤ (∑ni=1 |ai |2 ) 2 (∑ni=1 |bi |2 ) 2 .

Definition 0.4.1. Let p and q be positive real numbers. If 1 < p < ∞ and 1p + 1q = 1, or if p = 1
and q = ∞, or if p = ∞ and q = 1, then we say that p and q are conjugate exponents.

Lemma 0.4.3. (Young’s Inequality) Let p and q are conjugate exponents with 1 < p and α, β ≥
0. Then
αp βq
αβ ≤ + .
p q
Here, equality occurs iff β = α p−1 .

Proof. If p = 2 = q, then the inequality follows from the fact that (α − β )2 ≥ 0. Notice also,
that if α = 0 or β = 0, then the inequality follows trivially.
p q
If p 6= 2, then consider the function f : [0, ∞) → R given by f (α) = αp + βq − αβ , for fixed
β > 0.
1 q
Then, f 0 (α) = α p−1 − β = 0 when α p−1 = β , that is, when α = β p−1 = β p > 0.
q
We now apply the second derivative test to the critical point α = β p . f 00 (α) = (p −
q
1)α p−2 > 0, for all α ∈ (0, ∞). Thus, we have a global minimum at α = β p .

10
0.4. SOME INEQUALITIES

q p q p q
It is easily verified that 0 = f (β p ) ≤ f (α) = αp + βq − αβ ⇐⇒ αβ ≤ αp + βq , for each
α ∈ [0, ∞).
Or
The result follows directly if α = 0 or β = 0. Without the loss of generality, assume that
α > 0 and β > 0. Take p = mn + 1 and q = 1 + mn such that 1p + 1q = 1, where m and n are
positive integers. Assume α = x1/p and β = y1/q , we get
αp βq x y mx + ny
+ = n + m = .
p q m +1 1+ n m+n
We know that A.M.≥ G.M., we get
mx + ny 1
≥ (xm yn ) m+n = x1/p y1/q = αβ .
m+n
q
αp
Thus, αβ ≤ p + βq .

Lemma 0.4.4. (Holder’s inequality) If ai , bi are nonnegative real numbers then

∞ ∞ 1 ∞ 1
∑ |aibi| ≤ ( ∑ |ai| p) p ( ∑ |bi|q) q ,
i=1 i=1 i=1

1
where p + 1q = 1, p > 1 and ∑∞ p ∞ q
i=1 |ai | < ∞, ∑i=1 |bi | < ∞.

Proof. Suppose that ∑∞ p ∞ p

i=1 |ai | = 1, ∑i=1 |bi | = 1 and due to Young’s inquality,

1 1
|ai bi | ≤ |ai | p + |bi |q .
p q
Summing over i, we get
∞
1 ∞ 1 ∞ 1 1
∑ |aibi| ≤ p ∑ (|ai| ) + q ∑ (|bi|q) = p + q = 1.
p
i=1 i=1 i=1

Now construct sequence (xn ) and (yn ) as

ai
xn = 1
(∑∞ p p
i=1 |ai | )

and
bi
yn = 1 .
(∑∞ q q
i=1 |bi | )
Consider
|ai bi | |ai | p |bi |q
|xi yi | = 1 1 ≤ p
+ q
.
(∑∞ |a | p ) p (∑∞ |b |q ) q p ∑∞
i=1 |ai | q ∑∞
i=1 |bi |
i=1 i i=1 i
Summing over i, and using Young’s inequality we get
∞
|ai | |bi | 1 1
∑ |xiyi| = p
1
q
1 ≤ + = 1.
p q
i=1 (∑∞ ∞
i=1 |ai | ) (∑i=1 |bi | )
p q

1 1
Hence ∑∞ ∞ p p ∞ q q
i=1 |ai bi | ≤ (∑i=1 |ai | ) (∑i=1 |bi | ) .

11
0.4. SOME INEQUALITIES

Lemma 0.4.5. (Minkowski inequality) If p > 1 (fixed) and ai , bi are nonnegative real numbers
and ∑∞ p ∞ p
i=1 |ai | < ∞, ∑i=1 |bi | < ∞, then
∞ 1 ∞ 1 ∞ 1
( ∑ (|ai + bi |) p ) p ≤ ( ∑ |ai | p ) p + ( ∑ |bi | p ) p .
i=1 i=1 i=1

Proof. Take cn = an + bn . Consider

|cn | p = |an + bn | p = |an + bn ||an + bn | p−1 ≤ |an ||cn | p−1 + |bn ||cn | p−1 .

Further by summing, we have

m m m
∑ |cn| p ≤ ∑ |an||cn| p−1 + ∑ |bn||cn| p−1.
n=1 n=1 n=1

Applying Holder’s inequality,

!1/q " #
m m m 1 m 1
∑ |cn| p ≤ ∑ |cn|q(p−1) ( ∑ |ai | p ) + ( ∑ |bi | p )
p p ,
n=1 n=1 i=1 i=1

where q is such that 1/p + 1/q = 1 and q(p − 1) = p, 1 − 1/q = 1/p. Therefore, we have
m m m 1 m 1
( ∑ |cn | p )1/p = ( ∑ |an + bn | p )1/p ≤ ( ∑ |an | p ) p + ( ∑ |bn | p ) p .
n=1 n=1 n=1 n=1

Now note that this inequality holds for all m, and the series in the terms on the right-hand
side converge by assumption, so we can let m → ∞ to obtain the inequality we wanted.

12
0.5. METRIC SPACE

0.5 Metric Space

Definition 0.5.1. If X is a nonempty set and a function d : X × X → R satisfies the following
postulates:
(M1) d(x, y) ≥ 0, ∀x, y ∈ X
(M2) d(x, y) = 0, iff x = y
(M3) d(x, y) = d(y, x) [symmetry]
(M4) d(x, y) ≤ d(x, z) + d(z, y) [triangle inequality]
Then the function d on set X is called a metric or a distance function on set X. A non-empty
set X equipped withg a metric d is known as metric space and it is denoted by (X, d).
Example 0.5.1. • Real Line R, d(x, y) = |x − y| in R.
• usual metric on R2 , that is,
p
d : R2 × R2 → R is d((x1 , y1 ), (x2 , y2 )) = (x1 − x2 )2 + (y1 − y2 )2 .
• usual metric on Rn , that is,
p n
d : Rn × Rn → R is d((x1 , x2 , ..., xn ), (y1 , y2 , ..., yn )) = ∑i=1 (xi − yi )2 .
• trivial metric or discrete metric: let X be a nonempty set and d : X × X → R defined as
(
1; x 6= y
d(x, y) =
0; x = y.

Here d(x, y) ≥ 0, d(x, y) = 0 iff x = y, d(x, y) = d(y, x). Also, for any case triangle
inequality holds.
(
|x − y|; x = cy; for somec > 0
• Paris metric. Let x, y ∈ R, the metric is defined as d(x, y) =
|x| + |y|, else.
Example 0.5.2. • ` p space. Let p ≥ 1 be a fixed real number. By definition, each element
in the space ` p is a sequence x = (ξi ) = (ξ1 ; ξ2 , 0, 0, 0) of numbers such that the series
1
∑|ξi | p converges; thus ∑|ξi | p < 1, and the metric is defined by d(x, y) = (∑|ξi − ηi | p ) p ,
where y = (ηi ) and ∑|ηi | p < ∞. We prove that ` p is a metric space. First three properties
of metric are trivial. To prove triangle inequality, we take x; y; z ∈ ` p and using triangle
inequality, we get

1
d(x, y) = ∑|ξi − ηi | p p

1
≤ ∑[|ξi − φi | + |φi − ηi |] p p

1
1
≤ ∑|ξi − φi | p p + ∑|φi − ηi | p p using Minkowski’s inequality
= d(x, z) + d(z, y).
Hence ` p is a metric space.

1/2
• Euclidean space Rn . d(x, y) = ∑ni=1 (ξi − ηi )2 in Rn , where x = (ξ1 , ξ2 , ..., ξn ) and
y = (η1 , ..., ηn ).
• Unitary space Cn . The n-dimensional unitary space is the space of all ordered n-tuples

1/2
of complex numbers with metric defined by d(x, y) = ∑ni=1 (ξi − ηi )2 in Cn , where
x = (ξ1 , ξ2 , ..., ξn ) and y = (η1 , ..., ηn ).

13
0.5. METRIC SPACE

• sequence space `∞ . Let X we take the set of all bounded sequences of complex numbers;
that is, every element of X is a sequence x = (ξ1 , ξ2 , ..., ξn ) written as x = (ξi ) such that
for all i = 1; 2; ..., we have |ξi | ≤ C; where C is a real number which may depend on x,
but does not depend on i. We choose the metric defined by d(x, y) = supi∈N |ξi − ηi |.
• Function space C[a, b]. Let X we take the set of all real-valued functions f , g, ... which
are functions of an independent real variable t and are defined and continuous on a given
closed interval I = [a, b]. Choosing the metric defined by d(x, y) = maxt∈[a,b] | f (t)−g(t)|.
Rb
We can also define another metric on same space as d1 (x; y) = a | f (t) − g(t)|dt.

Example 0.5.3. Which of the following function d : R × R → R is a metric on R?

• d(x, y) = |x − y| (yes)
• d(x, y) = |x3 − y3 | (yes) Here d(x, y) ≥ 0.
Now d(x, y) = 0 =⇒ x3 = y3 =⇒ x2m−1 = y2m−1 =⇒ y = x, xω, xω 2 =⇒ x = y only
over R.
d(x, y) = d(y, x).
Consider

|x2m−1 − y2m−1 | = |x2m−1 − z2m−1 + z2m−1 − y2m−1 |

≤ |x2m−1 − z2m−1 | + |z2m−1 − y2m−1 |

Therefore d(x, y) ≤ d(x, z) + d(z, y), for any m = 1, 2, 3...

Note: d(x, y) = |x2m−1 − y2m−1 | is a metric on R.
• d(x, y) = |x − 2y| (no) as d(x, y) = |x − 2y| and d(y, x) = |y − 2x|, so d(x, y) 6= d(y, x).
• d(x, y) = |x2 − y2 | (no) as d(x, y) = 0 =⇒ |x2 − y2 | = 0 =⇒ y = x or y = −x, so the
second postulate fails.
• d(x, y) = max(x, y) or d(x, y) = min(x, y)[both not] Both don’t satisfy the postulate (i),
that is non-negative property as max(x, y) or min(x, y) will give negative values if x, y are
negative. e.g. min(−2, −5) = −5 < 0.
• d(x, y) = |x3 − y5 | [no] not symmetry
• d(x, y) = |x3001 − y3001 | [yes]
• Is d : R × R → R defined by d(x, y) = |x| + |y| metric on R? [no]
(i) d(x, y) = |x| + |y| ≥ 0
(ii) d(x, y) = |x| + |y| = 0 =⇒ x 6= y. So it is not metric
• d : R × R → R defined as
(
|x| + |y|; x 6= y
d(x, y) =
0; x = y.

Is d metric over R. [yes]

14
0.5. METRIC SPACE

d(x,y)
• If (X, d) is a metric space. Consider d1 (x, y) = 1+d(x,y) Is d1 a metric on X? [yes]
d(x,y)
(i) d1 (x, y) = 1+d(x,y) ≥ 0.
d(x,y)
(ii) d1 (x, y) = 1+d(x,y) = 0 =⇒ d(x, y) = 0 iff x = y.
d(x,y) d(y,x)
(iii) d1 (x, y) = 1+d(x,y) = 1+d(y,x) [as d is metric]. so d1 (x, y) = d1 (y, x).
(iv) Now, we have to prove the triangle inequality. Consider

d(x, y) ≤ d(x, z) + d(z, y)

1 + d(x, y) ≤ 1 + d(x, z) + d(z, y)
1 1
≥
1 + d(x, y) 1 + d(x, z) + d(z, y)
1 1
− ≤−
1 + d(x, y) 1 + d(x, z) + d(z, y)
1 1
1− ≤ 1−
1 + d(x, y) 1 + d(x, z) + d(z, y)
d(x, y) d(x, z) + d(z, y)
≤
1 + d(x, y) 1 + d(x, z) + d(z, y)
d(x, z) d(z, y)
≤ +
1 + d(x, z) 1 + d(z, y)
d1 (x, y) ≤ d1 (x, z) + d1 (z, y).

Hence d1 (x, y) is metric on X.

• If (X, d) is a metric space. Consider d1 (x, y) = min{1, d(x, y)} Is d1 a metric on X? [yes]
Here d1 (x, y) ≥ 0. Also d1 (x, y) = 0 ⇐⇒ d(x, y) = 0 ⇐⇒ x = y. d1 (x, y) = d1 (y, x).
If d1 (x, y) = 1, then d1 (x, z) = 1 and d1 (z, y) = 1 Hence d1 (x, y) ≤ d1 (x, z) + d1 (z, y).
If d1 (x, y) = d(x, y), then d1 (x, z) = d(x, z) and d1 (z, y) = d(z, y) Hence d1 (x, y) = d(x, y) ≤
d(x, z) + d(z, y) = d1 (x, z) + d1 (z, y).
Hence (X, d1 ) is a metric space.

Euclidean Spaces Let Rn = {(x1 , x2 , ..., xn ) : xi ∈ R∀1 ≤ i ≤ n}. This is a vector space over
R, where sum of two elements x ∈ Rn , y ∈ Rn is x + y = (x1 + y1 , ..., xn + yn ) ∈ Rn and the
product of x ∈ Rn with a scalar λ ∈ R is λ x = (λ x1 , ..., λ x2 ) ∈ Rn .
Define scalar product on Rn by (x, y) = x1 y1 q + ... + xn yn ∈ R.
n
p
Define the norm on R by ||x|| = (x, x) = x12 + ... + xn2 .
Define the map d : Rn × Rn → R by
q
d(x, y) = ||x − y|| = (x1 − y1 )2 + ... + (xn − yn )2 .

The metric space (Rn , d) is called the Euclidean space. The metric d is called the Euclidean
metric.

15
0.6. BASIC PROPERTIES

0.6 Basic Properties

Definition 0.6.1. Let (X, d) be a metric space E ⊆ X.
• If a ∈ X, then an open ball centered at 0 a0 and radius r is denoted by Br (a) and it will
consists of Br (a) = {x ∈ X : d(x, a) < r}.
In R, it is an interval. In R2 , it will become an open disc centered at ’a’ and radius r. In
R3 , it will become an open sphere. In Rn , n > 3, it will become a hyper sphere.
• Similarly, closed ball centered at 0 a0 and radius r is denoted by Br [a] and it will consists
of Br [a] = {x ∈ X : d(x, a) ≤ r}.
• A neighborhood of a point p is a set Nr (p) consisting of all points q such that d(p, q) < r.
The number r is called the radius of Nr (p).
• A point p is called a limit point of the set E if every neighborhood of p contains a point
q 6= p such that q ∈ E.
• The set of all limit points of E is known as derived set of E, denoted by E 0 .
• If p ∈ E and p is not a limit point of E, then p is called an isolated point of E, that is, p
is an isolated point of E if p ∈ E and there is a neighborhood of p that contains no other
point of E.
• p is a boundary point of E if every neighborhood of p contains at least one point in E
and one not in E. The set of boundary points is denoted by ∂ E.
• E is closed if every limit point of E is a point of E.[E 0 ⊆ E.]
A subset of R is closed iff it contains all its limit points.
Proof. Let F be a closed set in R and x be its limit point. We’ll show that x ∈ F. If
not then x ∈ F c and F c is open set. So there exists Nr (x), r > 0 such that Nr (x) ⊆ F c .
Consequently, Nr (x) ∩ F = 0,/ which contradicts that x is a limit point of F.
Conversely, Let F be a subset of R contains all its limit points. We’ll show that F c is
open. If y ∈ F c , then y is not a limit point of F. So there exists Nr (y), r > 0 that does not
contain a point of F (except possible y). But y ∈ F c so Nr (x) ⊆ F c , so F c is open in R.
Hence F is closed.
• A point is an interior point of E if there is a neighborhood N of p such that N ⊂ E. It is
denoted by E int or E o or intE.
• E is open if every point of E is an interior point of E. [A metric space is said to be open
if it is itself is an open set.], that is intS = S.
For example: Consider metric space ([0, 1], d). Which of the following sets are open?
(a) ( 21 , 1] it is open set as int( 12 , 1]= ( 12 , 1].
(b) (0, 12 ] it is not open as int(0, 21 ]= (0, 21 ).
(c) The set Q is not open in R with the usual metric but it is open in (R, d), where d
p metric in R. Indeed, if√x ∈ Q and r > 0, then√for large n ∈ N, we have
is a discrete
x < x + (2)/n < x + r so that x + 2/n ∈ B(x, r) but x + 2/n ∈ / Q. In the case of
discrete metric, for every x ∈ Q, B(x, 1/2) = {x} ⊂ Q, so Q is open in (R, d).
(d) R − N is an open set as int(R − N) = R − N.

16
0.6. BASIC PROPERTIES

• The complement of E [E c ] is the set of all points p ∈ X such that p ∈

/ E.
• p is exterior to E if p is in the interior of E c .
• E is bounded if there is a real number M and a point q ∈ X such that d(p, q) < M, for all
p ∈ E.
Closure of E is the set E = E ∪ E 0 . [Since E ⊂ E and E is closed, E 0 ⊆ (E)0 ⊆ E and
hence E ∪ E 0 ⊆ E. But E ∪ E 0 is a closed set containing E. So E ⊆ E ⊆ E ∪ E 0 . Thus
E = E ∪ E 0 .]
Example 0.6.1. Let E = (−∞, −1] ∪ (1, 2) ∪ {3}. Then
(a) The set of limit points of E is (−∞, −1] ∪ [1, 2].
(b) ∂ E = {−1, 1, 2, 3} and E = (−∞, −1] ∪ [1, 2] ∪ {3}.
(c) 3 is the only isolated point of E.
(d) The exterior point of E is (−1, 1) ∪ (2, 3) ∪ (3, ∞).
1 1
Example 0.6.2. For n ≥ 1, let In = [ 2n+1 , 2n ] and S = ∪∞ n=1 In . Then
(a) The set of limit points of S is S ∪ {0}.
(b) ∂ S = {x : x = 0orx = 1/n(n ≥ 2)} and S = S ∪ {0}
(c) S has no isolated points.
1 1 1
(d) The exterior of S is (−∞, 0) ∪ [ ∞ n=1 ( 2n+2 , 2n+1 )] ∪ ( 2 , ∞).
S

• E is dense in X if every point of X is a limit point of E, or a point of E or both, that is

E = E ∪ E 0 = X.
or
A set E is said to be dense in itself if each point of the set is a limit point of the set, that
is no isolated point in the set., that is, E ⊆ E 0 .
Is R − [a, b] dense in R, where a and b are two distinct real numbers.
Here E = R − [a, b] = (−∞, a) ∪ (b, ∞) and E 0 = (−∞, a] ∪ [b, ∞). E ∪ E 0 == E 0 =
(−∞, a] ∪ [b, ∞) = R − (a, b). So it is not dense in R.
For example, Q; R − Q; R; Q − N; [a, b], 0/ are dense in itself beacuse their derived sets
contained in R.
• E is perfect if E is closed and if every point of E is a limit point of E. [E is closed and
dense in itself, that is [E = E 0 .] E 0 ⊆ E and E ⊆ E 0 . Hence E = E 0 .]
For example, 0, / R are perfect set because they are closed and their derived set is itself,
that is dense in itself.
Any open interval (x, y) : x < y ∈ R necessarily contains an Irrational number, therefore the
interior of the Rationals is empty because an open set containing any Rational is not wholly
contained in the Rationals. The closure of the interior is also empty since the empty set is both
open and closed in any topology.
Theorem 0.6.3. Every neighborhood is an open set.
Proof. Consider a neighborhood E = Nr (p) and q be any point of E such that d(p, q) < r. Then
there is a positive real number h such that d(p, q) = r − h. For all points s, Nh (q), such that
d(q, s) < h, we have d(p, s) ≤ d(p, q) + d(q, s) < r − h + h = r, so s ∈ E. Hencs Nh (q) ⊂ E.
Thus q is an interior point of E.

17
0.7. SOME PROPERTIES IN METRIC SPACE

0.7 Some properties in metric space

Definition 0.7.1. (Convergent Sequence). We say that the sequence (xn ) is convergent in the
metric space X = (X, d) if lim d(xn , x) = 0 for some x ∈ X.
We shall also use the notation xn → x.
Definition 0.7.2. (Diameter). For a set M ⊂ X we define the diameter δ (M) by δ (M) =
supx,y∈M d(x, y).
M is said to be bounded if δ (M) is finite.

Lemma 0.7.1. Let {xn } be a sequence in a metric space (X, d).Then

(a) there is at most one point x ∈ X such that {xn } converges to x, that is it is unique.
(b) If {xn } is convergent then it is bounded.
(c) if {xn } converges to x and {yn } converges to y then d(xn , yn ) converges to d(x, y).

Proof. (a) Assume that xn → x and xn → y and x 6= y. Then d(x, y) > 0. Then we find a positive
integer k, l such that d(xn , x) < ε/2 for n ≥ k and d(xn , y) < ε/2, for n ≥ l. For n ≥ max{k, l},
we have

d(x, y) ≤ d(x, xn ) + d(xn , y)

< ε/2 + ε/2 = ε,

d(x, y) = 0. Hence, we conclude that it is impossible for a sequence {xn } to converge to two
different points.
(b) If {xn } converges to x, then we know that there is a positive integer k such that d(xn , x) <
ε for n > k.
Choose M = max{ε, d(x1 , x), ..., d(xk , x)}. Then d(xn , x) ≤ M, for n = 1, 2, 3.... So the
sequence is bounded.
(c) Assume that xn → x and yn → y. Consider

d(xn , yn ) − d(x, y) ≤ d(xn , x) + d(x, y) + d(y, yn ) − d(x, y) = d(xn , x) + d(y, yn ).

Interchanging xn and x, yn and y and multiplying with −1, we get

d(x, y) − d(xn , yn ) ≤ d(xn , x) + d(y, yn ).

Thus |d(x, y) − d(xn , yn )| ≤ d(xn , x) + d(y, yn ). Now, taking lim n → ∞, we have d(xn , yn ) →
d(x, y).
Definition 0.7.3. Let (X, d) be a metric space and {xn } be a sequence of points of X. We
say that {xn } is Cauchy if for every ε > 0 there exists an integer k such that d(xn , xm ) < ε if
n, m ≥ k.

Proposition 0.7.2. If {xn } is a Cauchy sequence, then {xn } is bounded.

Proof. Take ε = 1. Since {xn } is Cauchy, there exists an integer k such that d(xn , xm ) < 1 if
n, m ≥ k. Take m = k, then d(xn , xk ) < 1 for n ≥ k.
Let t > 1 be such that d(xi , xk ) < t for 1 ≤ i ≤ k − 1. Then xn ∈ B(xk ,t) for all n, so {xn } is
bounded.

Proposition 0.7.3. If {xn } is convergent, then {xn } is a Cauchy sequence.

18
0.7. SOME PROPERTIES IN METRIC SPACE

Proof. Assume that xn → x. Then for a given ε > 0, there exists an integer k such that d(xn , x) <
ε/2 if n ≥ k. By taking n, m ≥ k,
d(xn , xm ) ≤ d(xn , x) + d(x, xm ) < ε/2 + ε/2 = ε.
So {xn } is a Cauchy sequence.
Proposition 0.7.4. If {xn } is a Cauchy and it contains a convergent subsequence, then {xn }
converges.
Proof. Assume that {xn } is Cauchy and {xnk } is subsequence of {xn } such that xnk → x. Let
ε > 0. Since {xn } is Cauchy, there exists k0 such that d(xn , xnk ) < ε/2 for all n ≥ k0 . Also, since
xnk → x, there exists k00 such that d(xnk , x) < ε/2 for all n ≥ k00 . Set k = max{k0 , k00 }. Then for
n ≥ k,
d(xn , x) ≤ d(xn , xnk ) + d(xnk , x) < ε/2 + ε/2 = ε.
Hence xn → x.
A Cauchy sequence need not converge. For example, consider sn = {1/n} in the metric
space (X, d), where X = (0, 1), d = |x − y|. Here the sequence is Cauchy in (0,1) but does not
converge to any point of interval as sn → 0 ∈
/ X.
We shall prove that {sn } is a Cauchy sequence.
For ε > 0, choose an integer k such that k > 2/ε. Let m, n ≥ k such that |pm | ≤ 1/m < ε/2
and |pn | ≤ 1/n < ε/2. Now, consider
d(pm , pn ) = |pm − pn |
≤ |pm | + |pn |
≤ 1/m + 1/n
< ε/2 + ε/2 = ε, ∀m, n ≥ k.
Hence {sn } is a Cauchy sequence.
Definition 0.7.4. A metric space (X, d) is called complete if every Cauchy sequence {xn } in X
converges to some point of X.
A subset A of X is called complete if A as a metric subspace of (X, d) is complete, that is,
if every Cauchy sequence {xn } in A converges to a point in A.
Theorem 0.7.5. Let (X, d) be a complete metric space and A ⊆ X. Then A is complete iff A is
closed.
Proof. Suppose that A is a complete subspace of X. We claim that A is closed, that is A = A.
We know that A ⊆ A. Let x ∈ A = A ∪ A0 . So x ∈ A or x ∈ A0 . If x ∈ A0 , x is a limit point of
A ao there exists a sequence {pn } in A such that lim pn = x. Since every convergent sequence
is Cauchy, so {pn } is a Cauchy sequence in A and A is complete. Therefore x ∈ A.
Therefore in both the cases x ∈ A. Hence A ⊆ A. Therefore, A = A. Hence A is closed.
Conversely, Let A be a closed subspace of X. We claim that A is complete.
Let {pn } be a Cauchy sequence in A. As A ⊆ X, so {pn } be a Cauchy sequence in X. Since
X is complete, lim pn = p ∈ X. This implies p ∈ A. Since A is closed, A = A. Hence p ∈ A.
Therefore A is complete.
Or
Let A be complete. Then for every x ∈ A, there is a sequence (xn ) which converges to x.
Since (xn ) is Cauchy and A is complete xn → x. Therefore A is closed.
Conversely, Let A be closed and (xn ) be a Cauchy sequence in A. Then xn → x ∈ A, which
implies x ∈ A, so x ∈ A. Hence A is complete.

19
0.7. SOME PROPERTIES IN METRIC SPACE

Corollary 0.7.6. (i) Every closed subset of a complete metric space is complete.
(ii) Every complete subset of a metric space is closed.

Definition 0.7.5. Let (X, d) and (Y, d1 ) be metric spaces and f : X → Y be a mapping. The
function f is said to be continuous at the point x0 ∈ X if the following holds:
For every ε > 0, there exists a δ > 0 such that for all x ∈ X if d(x, x0 ) < δ , then d1 ( f (x), f (x0 )) <
ε.

Theorem 0.7.7. Let f : X → Y be a function from a metric space (X, d) to (Y, d1 ) and x0 ∈ X.
Then f is continuous at x0 iff for every sequence {xn } such that xn → x0 , f (xn ) → f (x0 ).
Also, f is continuous iff for every convergent sequence {xn } in X, limn→∞ f (xn ) =
f (limn→∞ xn ).

Proof. Suppose that f is continuous at x0 and xn → x0 . We have to prove that f (xn ) → f (x0 ).
Let ε > 0 be given. By definition of continuity at x0 , there exists δ > 0 such that for all x ∈ X,
if d(x, x0 ) < δ , then d1 ( f (x), f (x0 )) < ε.
Since xn → x0 , there exists an integer k such that for all n ≥ k, d(xn , x0 ) < δ .
Hence d1 ( f (xn ), f (x0 )) < ε, for all n ≥ k, that is, f (xn ) → f (x0 ).
Conversely, assume that f is not continuous at x0 . Then there exists some ε > 0 such that for
every δ > 0 there exists a point x ∈ X satisfying d(x, x0 ) < δ but d1 ( f (x), f (x0 )) ≥ ε. For each
n, take δ = 1/n, n=1,2,... and choose xn so that d(xn , x0 ) < 1/n but d1 ( f (xn ), f (x0 )) ≥ ε. Hence
xn → x0 but { f (xn )} does not converge to f (x0 ), which is not true as per our assumption.

20
0.8. NORMED VECTOR SPACE

0.8 Normed Vector Space

Motivation. In many cases a vector space X may at the same time be a metric space because a
metric d is defined on X. However, if there is no relation between the algebraic structure and
the metric, we cannot expect a useful and applicable theory that combines algebraic and metric
concepts. To guarantee such a relation between ”algebraic” and ”geometric” properties of X
we define on X a metric d in a special way as follows. We first introduce an auxiliary concept,
the norm (definition below), which uses the algebraic operations of linear space. Then we
employ the norm to obtain a metric d that is of the desired kind. This idea leads to the concept
of a normed space.
Let X be a vector space over R or C. A norm is a function ||.|| : X → R having the following
properties:
(N1) ||x|| ≥ 0 and ||x|| = 0 iff x = 0. (positive definiteness)
(N2) ||αx|| = |α| ||x||, for all x ∈ X and α ∈ R or C. (scalar)
(N3) ||x + y|| ≤ ||x|| + ||y|| for all x, y ∈ X. (triangle inequality)
The pair (X, ||.||) is called a normed vector space. The pair (X, ||.||) is called a semi-normed
linear space if it satisfies (N2) and (N3), but not necessarily (N1).
If x, y ∈ X, then the number ||x − y|| provides a notion of closeness of points x and y in X,
that is, a distance between them. Thus ||x|| = ||x − 0|| is the distance of x from the zero vector
in X.
Example 0.8.1. 1. R or C is a vector space over R or C, and if we define ||.|| : R → [0, +∞)
by kxk = |x|, x ∈ R or C, then it becomes a normed space.
It is easy to verify N1 and N2, we now verify N3, for x, y ∈ R or C, we have

kx + yk2 = |x + y|2 = (x + y)x + y = (x + y)x + y = xx + yx + xy + yy

= |x|2 + xy + xy + |y|2 = |x|2 + 2 Re(xy) + |y|2
≤ |x|2 + 2 |xy| + |y|2 = |x|2 + 2 |x||y| + |y|2
= |x|2 + 2 |x||y| + |y|2 = (|x| + |y|)2 = (kxk + kyk)2 .

Hence the result.

2. Let n be a natural number and X = Rn or Cn . For each x = (x1 , x2 , ...xn ) ∈ X, define
1
kxk p = (∑ni=1 |xi | p ) p , for 1 ≤ p < ∞ and
kxk∞ = max1≤i≤n |xi | Then (X, k.k p ) and (X, k.k∞ ) are normed spaces.
1
N1) For ach 1 ≤ i ≤ n, we have |xi | ≥ 0 =⇒ ∑ni=1 |xi | p ≥ 0 =⇒ (∑ni=1 |xi | p ) p ≥ 0 =⇒
kxk p ≥ 0.
1
For any x ∈ X, kxk p = 0 ⇔ (∑ni=1 |xi | p ) p = 0 ⇔ |xi | p = 0 ⇔ xi = 0, for all i = 1, ...n ⇔
x = 0.
1 1 1
N2) For any x ∈ X and α ∈ F, kαxk p = (∑ni=1 |αxi | p ) p = (|α| p ∑ni=1 |xi | p ) p = |α| (∑ni=1 |xi | p ) p =
|α| kxk p

21
0.8. NORMED VECTOR SPACE

N3) For any x, y ∈ X,

!1
n p

kx + yk p = ∑ |xi + yi| p
i=1
!1 !1
n p n p

≤ ∑ |xi| p + ∑ |yi| p
i=1 i=1
= kxk p + kyk p .

3. Let X = B[a, b] be the set of all bounded real-valued functions on [a, b]. For each x ∈ X,
define kxk∞ = supa≤t≤b |x(t)|. Then (X, k.k∞ ) is a normed linear space.
N1 and N2 are simple to prove. Now, we prove N3, For any t ∈ [a, b] and x, y ∈ X, we
have
|x(t) + y(t)| ≤ |x(t)| + |y(t)| ≤ sup |x(t)| + sup |y(t)| = kxk∞ + kyk∞ .
a≤t≤b a≤t≤b

Since this is true for all t ∈ [a, b], we have that kx + yk∞ = supa≤t≤b |x(t) + y(t)| ≤ kxk∞ +
kyk∞ .
R 1
b 2 2
4. Let X = C[a, b]. For each x ∈ X, define kxk∞ = supa≤t≤b |x(t) and kxk2 = a |x(t)| dt .
Then (X, k.k2 ) and (X, k.k∞ ) are normed linear spaces.
p |
1 ∈ X, define kxk p = (∑i∈N |xi | ) f rac1p. Then
5. Let X = ` p , 1 ≤ p < ∞. For each x = (xi )∞
(X, k.k p ) and is a normed linear space.

1 ∈ X, define kxk = kxk∞ = supi∈N |xi |. Then (X, k.k∞ ) and

6. Let X = `∞ . For each x = (xi )∞
is a normed linear space.

Proposition 0.8.2. Let X be a normed space. Then d(x, y) = ||x − y||, x, y ∈ X defines a metric
on X.
Proof. The axioms M1, M2, M3 are easy to see. Now, for x, y, z ∈ X, we have
d(x, z) ≤ ||x − z||
= ||(x − y) + (y − z)||
≤ ||x − y|| + ||y − z||
= d(x, y) + d(y, z).
Hence it is a metric.
Theorem 0.8.3. (a) If (X, k.k) is a normed linear space, then d(x, y) = kx − yk defines a metric
on X. Such a metric d is said to be induced or generated by the norm k.k. Thus, every normed
linear space is a metric space, and unless otherwise specified, we shall henceforth regard any
normed linear space as a metric space with respect to the metric induced by its norm.
(b) If d is a metric on a linear space X satisfying the properties: For all x; y; z ∈ X and for
all α ∈ F,
i) d(x, y) = d(x + z, y + z) (Translation Invariance)
ii) d(αx, αy) = |α|d(x, y) (Absolute Homogeneity); then kxk = d(x, 0) defines a norm on
X.

22
0.8. NORMED VECTOR SPACE

Proof. (a) We show that d(x, y) = kx − yk defines a metric on X. To that end, let x; y; z ∈ X.
M1. d(x, y) = kx − yk ≥ 0 and d(x, y) = 0 ⇔
M2. kx − yk = 0 ⇔ x = y by N1.
M3. d(x, y) = kx − yk = k(−1)(y − x)k = | − 1| ky − xk = ky − xk = d(y, x)
M4.

d(x, z) ≤ ||x − z||

= ||(x − y) + (y − z)||
≤ ||x − y|| + ||y − z||
= d(x, y) + d(y, z).

(b) d(x + z, y + z) = kx + z − (y + z)k = kx − yk = d(x, y)

d(αx, αy) = kαx − αyk = kα(x − y)k = |α| kx − yk.

It is clear from Theorem 0.8.3, that a metric d on a linear space X is induced by a norm on
X if and only if d is translation-invariant and positive homogeneous.
Metric for Rn : We define a norm of a vector n n 2 21
p Xn = (x1 , x2 , ..., xn ) ∈ R as ||X|| = (∑i=1 xi ) .
The metric induce by this norm is d(X,Y ) = ∑i=1 (xi − yi )2 .
1
For p ≥ 1 and X = (x1 , x2 , ..., xn ). Let ||X|| p = (∑ni=1 xi2 ) p . The metric induce by this norm
1
is d p (X,Y ) = (∑ni=1 (xi − yi ) p ) p .
For p = ∞, we have ∞ norm on Rn ||X||∞ = max{|xi | : 1 ≤ i ≤ n}. The metric induce by
this norm is d∞ (X,Y ) = max{|xi − yi | : 1 ≤ i ≤ n}.
Diameter If (X, d) is a metric space then for any subset A of X, Diameter of A is denoted
by dia(A) = d(A) = supx,y∈A d(x, y).
d({a}) = 0, that is diameter of singleton set is zero.
Diameter of a set containing more than one element will be positive.
A metric space is said to be bounded if diameter of X is finite.
|x−y|
Example 0.8.4. Consider metric space (R, d), where (a) d(x, y) = |x − y| (b) d(x, y) = 1+|x−y|
What is the diameter of metric space (R, d)? Is (R, d) bounded?
(a) d(x, y) = |x − y|. Here dia(R) = ∞, because dia(R) = sup d(x, y) = sup[0, ∞) = ∞.
Hence metric (R, d) is not bounded.
|x−y|
(b) d(x, y) = 1+|x−y| < 1. dia(R) = sup d(x, y) = sup[0, 1) = 1. Hence metric (R, d) is
bounded.

23
0.9. CONVEX SET

0.9 Convex Set

Let L be a real linear space and let x, y ∈ L. The closed segment determined by x and y is the
set
[x, y] = {(1 − a)x + ay : 0 ≤ a ≤ 1}.
The half-closed segments determined by x and y are the sets
[x, y) = {(1 − a)x + ay : 0 ≤ a < 1},
and
(x, y] = {(1 − a)x + ay : 0 < a ≤ 1}.
The open segment determined by x and y is
(x, y) = {(1 − a)x + ay : 0 < a < 1}.
The line determined by x and y is the set
`x,y = {(1 − a)x + ay : a ∈ R}.
Definition 0.9.1. Let X be a linear space over a field F and let x1 , x2 ∈ X. The set of all elements
of the form y = α x1 + (1 − α)x2 , 0 ≤ α ≤ 1, is called the line segment joining the points x1
and x2 . A set C ⊂ X is called convex if all segments joining any two points of the set C are
contained in the set.
Clearly every subspace M of X is a convex set.

(convex set) (nonconvex set)

In Euclidean space the cube, circle, sphere, triangular area in Euclidean 2-space, Euclidean
plane and line segments are convex sets.
Thus a set C is convex in X if for all scalars α in 0 ≤ α ≤ 1, αC +(1−α)C ⊂ C, equivalently
for any two reals α, β with 0 ≤ α, β < 1, αC + βC ⊂ C.
Theorem 0.9.1. Let x, y, z be three distinct points in the real linear space L such that z ∈ `x,y .
Then, one of these points belongs to the open segment determined by the remaining two points.
Proof. Since z ∈ `x,y , we have z = (1 − a)x + ay for some a ∈ R.

y 
a>1
u
 
x  
u 0<a<1


 a<0

1
We have a 6∈ {0, 1} because the points x, y, z are distinct. If a > 1 we have y = a−1
a x + a z,
a−1 1
so y ∈ (x, z) because a , a ∈ (0, 1). If 0 < a < 1 we have z ∈ (x, y). Finally, if a < 0, since
a −a


x = 1 + 1−a z + 1−a y, we have x ∈ (z, y).

24
0.9. CONVEX SET

Theorem 0.9.2. Let L be a real linear space. A subset C of L is convex if and only if any convex
combination of elements of C belongs to C.
or
Let L be a linear space and C ⊂ L be a convex set. If x1 , x2 , · · · , xn ∈ C then all elements of
the form α1 x1 + α2 x2 + · · · + αn xn ∈ C where αi ≥ 0 and α1 + α2 + · · · + αn = 1.

Proof. The sufficiency of this condition is immediate. To prove its necessity consider x1 , . . . , xk ∈
C and the convex combination
y = a1 x1 + · · · + ak xk .
We prove by induction on k ≥ 1 that y ∈ C. The base case, m = 1 is immediate since in this
case y = a1 x1 and a1 = 1.
For the inductive step, suppose that the statement holds for k and let y be given by y =
a1 x1 + · · · + ak xk + ak+1 xk+1 , where a1 + · · · + ak + ak+1 = 1, ai ≥ 0 and xi ∈ C for 1 ≤ i ≤ k + 1.
We have
k
ai
y = (1 − ak+1 ) ∑ xi + ak+1 xk+1 .
i=1 1 − ak+1

Since z = ∑ki=1 1−aaik+1 xi is a convex combination of k vectors, we have z ∈ C by the inductive

hypothesis, and the equality y = (1 − ak+1 )z + ak+1 xk+1 implies y ∈ C. Hence C is convex.

Theorem 0.9.3. Arbitrary intersection of any number of convex sets in a linear space is a
convex set.

Proof. Let M = {Ci : i ∈ I} be a collection of convex sets and let C = M. Suppose that
T

x1 , . . . , xk ∈ C, ai ≥ 0 for 1 ≤ i ≤ k, and a1 + · · · + ak = 1. Since x1 , . . . , xk ∈ Ci , it follows that

a1 x1 + · · · + ak xk ∈ Ci for every i ∈ I. Thus, a1 x1 + · · · + ak xk ∈ C, which proves the convexity
of C.
Note: Union of two convex sets may not be convex.

Theorem 0.9.4. Let C be a convex subset of a real linear space L. If r1 , r2 ∈ R≥0 , then we have

(r1 + r2 )C = r1C + r2C.

Proof. If at least one of r1 , r2 is 0 the equality obviously holds; therefore, assume that both r1
and r2 are positive.
Let z ∈ r1C + r2C. There exists x, y ∈ C such that z = r1 x + r2 y, and therefore,
 
r1 r1
z = (r1 + r2 ) x+ y .
r1 + r2 r1 + r2

Since C is convex, r1r+r

1
2
x + r1r+r
1
2
y ∈ C, which implies z ∈ (r1 +r2 )C, so r1C +r2C ⊆ (r1 +r2 )C.
The reverse inclusion is immediate. Suppose that (r1 + r2 )C = r1C + r2C holds for all
r1 , r2 ≥ 0. If 0 ≤ α ≤ 1, take r1 = α and r2 = 1 − α, then we see that αC + (1 − α)C = C ⊂ C.

Theorem 0.9.5. If A and B are two convex sets in a linear space X then A + B is also a convex
set.

25
0.9. CONVEX SET

Proof. Now, A + B = {a + b : a ∈ A, b ∈ B}. Let x, y ∈ A + B. So x = a1 + b1 , for some a1 ∈

A, b1 ∈ B and y = a2 + b2 for some a2 ∈ A, b2 ∈ B. Let 0 ≤ α ≤ 1. Now
αx + (1 − α)y = α(a1 + b1 ) + (1 − α)(a2 + b2 )
= (αa1 + (1 − α)a2 ) + (αb1 + (1 − α)b2 ) ∈ A + B,
as A and B are convex sets; αa1 + (1 − α)a2 ∈ A, αb1 + (1 − α)b2 ∈ B. So A + B is a convex
set in X.
Theorem 0.9.6. Let (L, k·k) be a normed linear space. An open sphere B(x0 , r) ⊆ L is convex.
Proof. Indeed, suppose that x, y ∈ B(x0 , r), that is, kx − x0 k < r and kx0 − yk < r.
Let a ∈ [0, 1] and let z = (1 − a)x + ay. We have
kx0 − zk = kx0 − (1 − a)x − ayk
= ka(x0 − y) + (1 − a)(x0 − x)k
≤ a kx0 − yk + (1 − a) kx0 − xk < r.
so z ∈ B(x0 , r).
Similarly, a closed sphere B[x0 , r] is a convex set.
Definition 0.9.2. Let S be a non-empty subset of a linear space X, the convex hull of S denoted
by conv. hull(S) is the intersection of all convex sets in X containing S.
So, conv. hull (S) in the smallest convex set in X containing S.
Theorem 0.9.7. Conv. hull(S) consists of all vectors

x = α1 x1 + α2 x2 + · · · + αn xn ,

i.e. x = ∑ni=1 αi xi , where x1 , x2 , · · · , xn ∈ S, αi ≥ 0 with ∑ni=1 αi = 1 and index n is not fixed.

Proof. The proof follows from the principle of mathematical induction.
Let T be the collection of all vectors x = ∑ni=1 αi xi , where x1 , x2 , · · · , xn ∈ S, αi ≥ 0 with
∑ni=1 αi = 1 and index n is not fixed.
For n = 1 every element of S is an element of T , i.e S ⊂ T .
We shall now prove that T is convex. Let u, v ∈ T , where u = ∑ni=1 αi xi , where x1 , x2 , · · · , xn ∈
S, αi ≥ 0 with ∑ni=1 αi = 1.
Take v = ∑m m
i=1 βi xi , where x1 , x2 , · · · , xm ∈ S, βi ≥ 0 with ∑i=1 βi = 1.
Let 0 ≤ γ ≤ 1 . Then
n m
γu + (1 − γ)v = γ( ∑ αi xi ) + (1 − γ)( ∑ βi xi )
i=1 i=1
n m
= ∑ γαi xi + ∑ (1 − γ)βi xi .
i=1 i=1

Now, γαi ≥ 0 and (1 − γ)βi ≥ 0. Also,

n m
∑ γαi + ∑ (1 − γ)βi
i=1 i=1
n m
= γ ∑ αi + (1 − γ) ∑ βi
i=1 i=1
= γ + (1 − γ) = 1.

26
0.9. CONVEX SET

So γu + (1 − γ)v ∈ T. Therefore T is convex. S ⊂ T and consequently conv. hull(S) ⊂ T .

Now, we have to prove that T ⊂ conv. hull(S).
Let W be any convex set containing S. For n = 1, every element of T is an element of S and
consequently every element of T is an element of W .
Let us consider that the result is true for n = m, i.e. it is true for m number of elements, i.e.,
if x = ∑m m
i=1 αi xi where x1 , x2 , · · · , xm ∈ S, αi ≥ 0 with ∑i=1 αi = 1, then x is an element of W .
Let, n = m + 1 and x = ∑m+1 m+1
i=1 αi xi ∈ T , where αi ≥ 0 with ∑i=1 αi = 1.
Take, β = α1 + α2 + · · · + αm . Then β > 0. Put βi = β , i = 1, 2, · · · , m. Then, ∑m
αi
i=1 βi = 1.
As W is convex and β1 x1 + β2 x2 + · · · + βm xm , xm+1 ∈ W , we have

x = α1 x1 + α2 x2 + · · · + αm xm + αm+1 xm+1
= β β1 x1 + β β2 x2 + · · · + β βm xm + (1 − β )xm+1
= β (β1 x1 + β2 x2 + · · · + βm xm ) + (1 − β )xm+1 ∈ T ⊂ W.

So x = α1 x1 + α2 x2 + · · · + αm xm + αm+1 xm+1 ∈ W .
Thus the result is true by the Principle of Mathematical Induction, i.e. every elements of T
is an element of W with S ⊂ W . As W is arbitrary it follows that T ⊂ ∩W , where S ⊂ W , that
is T ⊂ conv hull(S).
Therefore T = conv hull(S). Hence the result.

27
0.10. QUOTIENT SPACE

0.10 Quotient Space

Let M be a linear subspace of a linear space X over field F. For all x, y ∈ X, define x ≡ y(
mod M) ⇐⇒ x − y ∈ M.
It is easy to verify that ≡ defines an equivalence relation on X.
For x ∈ X, denote by [x] = {y ∈ X : x ≡ y (mod M)} = {y ∈ X : x − y ∈ M} = x + M the
coset of x with respect to M. The quotient space X/M consists of all the equivalence classes
[x], x ∈ X. The quotient space is also called a factor space.

Proposition 0.10.1. Let M be a linear subspace of a linear space X over F. For x, y ∈ X and
α ∈ F, define the operations [x + y] = [x] + [y] and [αx] = α[x]. Then X/M is a linear space
with respect to these operations.

Hint: For all x; y ∈ X and α ∈ F, (x + M) + (y + M) = x + y + M and α(x + M) = αx + M.

Let Y be a subspace of a linear space X. The coset of an element x ∈ X is denoted by x +Y
and is defined to be the set x +Y = {v ∈ Y : v = x + y; y ∈ Y }.
These distinct cosets form a partition of X. The quotients of normed spaces may be formed.
Notice that we need both the algebraic structure (subspace of a linear space) and a topolog-
ical property (closed) to make it all work.
The linear space X/Y (the quotient or factor space) is formed as follows. The elements of
X/Y are cosets of Y - sets of the form x +Y for x ∈ X. The set of cosets is a linear space under
the operations
(x1 +Y ) + (x2 +Y ) = (x1 + x2 ) +Y
α(x +Y ) = αx +Y.
Notice that this makes sense precisely because Y is itself a linear space, so for example
Y +Y = Y and αY = Y for α 6= 0.
Two cosets x1 +Y and x2 +Y are equal if as sets x1 +Y = x2 +Y , which is true if and only
if x1 + x2 ∈ Y .
Example 0.10.2. 1. Let X = R2 , and let Y be a one-dimensional subspace of R2 , that is,
Y = {(y1 , 0) : x ∈ R}.
Then for a given vector x = (x1 , x2 ) ∈ R2 , the coset x +Y is the set
x +Y = {x + y : y ∈ Y }
= {(x1 + y1 , x2 + 0) : y1 ∈ R}
= {(y1 , x2 ) : y1 ∈ R},
which is the horizontal line at height x2 .
The quotient space R2 /Y = {x + Y : x ∈ R2 } = {(y1 , 0) + Y : y1 ∈ R}; that is, quotient
space R2 /Y is the set of all horizontal lines in R2 .
We now want to introduce a norm on a quotient space. Let M be a closed linear subspace
of a normed linear space X over F. For x ∈ X, define k[x]k := infy∈[x] kyk.
If y ∈ [x], then y−x ∈ M and hence y = x+M, for some m ∈ M. Hence k[x]k = infy∈[x] kyk =
infm∈M kx + mk = infm∈M kx − mk = d(x, M).

Proposition 0.10.3. Let M be a closed linear subspace of a normed linear space X over F.
The quotient space X/M is a normed linear space with respect to the norm k[x]k = infy∈[x] kyk,
where [x] ∈ X/M.

28
0.10. QUOTIENT SPACE

Proof. N1. It is clear that for any x ∈ X, k[x]k = d(x, M) ≥ 0.

For any x ∈ X, k[x]k = 0 ⇔ d(x, M) = 0 ⇔ x ∈ M = M ⇔ x + M = M = [0].
N3. For any x; y ∈ X and α ∈ F/{0},
y
kα[x]k k[αx]k = d(αx, M) = inf kαx − yk = inf α(x − )
y∈M y∈M α
|α| inf kx − zk = |α|d(x, M) = |α| k[x]k .
z∈M

N4. Let x, y ∈ X. Then

k[x] + [y]k = k[x + y]k = d(x + y, M) = inf kx + y − zk

z∈M
= inf kx + y − (z1 + z2 )k
z1 ,z2 ∈M
= inf k(x − z1 ) + (y − z2 )k
z1 ,z2 ∈M
≤ inf k(x − z1 )k + k(y − z2 )k
z1 ,z2 ∈M
= inf k(x − z1 )k + inf k(y − z2 )k
z1 ∈M z2 ∈M
= d(x, M) + d(y, M) = k[x]k + k[y]k .

The norm on X/M as defined in Proposition 0.10.3 is called the quotient norm on X/M.

Theorem 0.10.4. If X is a normed space, and Y is a closed normed linear subspace of X, then
X/Y is a normed space under the norm kx +Y k = infz∈x+Y kzk.

Proof. Let x + Y be any coset of X, and (xn ) ⊂ z + Y be a convergent sequence with xn → x.

Then for any fixed n, xn − xm → xn − x is a sequence in Y converging in X. Since Y is closed,
we must have xn − x ∈ Y , so x +Y = xn +Y = z +Y . That is, the limit of the sequence defines
the same coset as does the sequence - the set z +Y is a closed set.
Assume now that kx +Y k = 0. Then there is a sequence (xn ) ⊂ x +Y with kxn k → 0. Since
x +Y is closed and xn → 0, we must have 0 ∈ x +Y , so x +Y = Y , the zero element in X/Y .
Homogeneity: kα(x +Y )k = infz∈x+Y kαzk = |α| infz∈x+Y kzk = |α| kx +Y k .
Triangle inequality:

k(x1 +Y ) + (x2 +Y )k = inf kz1 + z2 k

z1 ∈x+Y,z2 ∈x+Y
≤ inf kz1 k + inf kz2 k
z1 ∈x+Y z2 ∈x+Y
= kx1 +Y k + kx2 +Y k .

Proposition 0.10.5. (a) Show that k.k is continuous, that is, k.k : X → R defined by x 7→ kxk is
a continuous mapping.
(b) Show that the non trivial mapping f (x) = Ax from Rn to Rm (with Euclidean norms)
defined by the m × n matrix A = (ai j ) is uniformly continuous.

29
0.10. QUOTIENT SPACE

Proof. (a) Using triangle inequality, |kxk − kyk| ≤ kx − yk ≤ δ = ε.

(b) Using the Cauchy-Schwartz inequality, we can show that f is uniformly continuous.
Take a ∈ Rn and b = Aa. Then for any x ∈ Rn , we have
m n
kAx − Aak2 = ∑ | ∑ ai j (x j − a j )|2
i=1 j=1
m n n
< ∑ ( ∑ |ai j |2 )( ∑ |(x j − a j )|)2
i=1 j=1 j=1
2
= C2 kx − ak ,

where C2 = ∑nj=1 |ai j |2 > 0. It follows that f is uniformly continuous

By defnition, a subspace Y of a normed space X is a subspace of X considered as a vector
space, with the norm obtained by restricting the norm on X to the subset Y . This norm on Y is
said to be induced by the norm on X.
If Y is closed in X; then Y is called a closed subspace of X.
By definition, a subspace Y of a Banach space X is a subspace of X considered as a normed
space. Hence we do not require Y to be complete.

30
0.11. COMPLETENESS PROPERTY IN NORMED LINEAR SPACE

0.11 Completeness property in normed linear space

Definition 0.11.1. Let (xn )∞ n=1 be a sequence in a normed linear space (X, k.k)
∞
(a) (xn )n=1 is said to converge to x if given ε > 0 there exists a natural number N = N(ε)
such that kxn − xk < ε for all n ≥ N.
Equivalently, (xn )∞ lim kxn − xk = 0.
n=1 converges to x if n→∞
If this is the case, we shall write xn → x or lim xn = x.
n→∞
Convergence in the norm is called norm convergence or strong convergence.
(b) (xn )∞n=1 is called a Cauchy sequence if for given ε > 0 there exists a natural number
N = N(ε) such that kxn − xm k ≤ ε for all m, n ≥ N.
Equivalently, (xn )∞ n=1 is Cauchy if lim kxn − xm k = 0.
m,n→∞

Lemma 0.11.1. Let X be a normed linear space and A be a nonempty subset of X.

1. |d(x, A) − d(y, A) ≤ kx − yk, for all x, y ∈ X.

2. | kxk − kyk | ≤ kx − yk, for all x, y ∈ X.
3. If xn → x, then kxn k → kxk.
4. If xn → x and yn → y, then xn + yn → x + y.
5. If xn → x and αn → α, then αn xn → αx.
6. Every Cauchy sequence is bounded.
7. Every convergent sequence is a Cauchy sequence.

Proof. (1) For any a ∈ A, d(x, A) ≤ kx − ak ≤ kx − yk + ky − ak, so d(x, A) ≤ kx − yk + d(y, A)

or d(x, A) − d(y, A) ≤ kx − yk.
Interchanging x and y, we get the result.
(2) follows from (1) by taking A = {0}.
(3) follows from (2).
(4), (5) easy to prove using triangle inequality.
(6) Let (xn ) be a Cauchy sequence in X. Choose n1 so that kxn − xn1 k ≤ 1 for all n ≥ n1 .
By (2), kxn k ≤ 1 + kxn1 k for all n ≥ n1 . Thus kxn k ≤ max{kx1 k , ... kxn−1 k , 1 + kxn1 k} for all
n.
(7) Let (xn ) be a Cauchy sequence in X and converges to x ∈ X. Suppose that for a given ε >
0, there is a natural number N such that kxn − xk < ε2 for all n ≥ N. For n, m ≥ N, kxn − xm k ≤
kxn − xk + kx − xm k < ε2 + ε2 = ε. Thus {xn } is a Cauchy sequence in X.

Proposition 0.11.2. Let (X, k.k) be a normed linear space over F. A Cauchy sequence in X
which has a convergent subsequence is convergent.

Proof. Let (xn ) be a Cauchy sequence in X and {xnk } its subsequence which converges to x ∈ X.
Suppose that for a given ε > 0, there is a natural numbers N1 , N2 such that kxn − xm k < ε2 for
all n, m ≥ N1 and kxnk − xk < ε2 for all k ≥ N2 .
Let N = max{N1 , N2 }. If k ≥ N, then since nk ≥ k, kxk − xk ≤ kxn − xnk k + kxnk − xk <
ε ε
2 + 2 = ε. Hence xn → x.

Definition 0.11.2. A metric space (X, d) is said to be complete if every Cauchy sequence in X
converges in X.

31
0.11. COMPLETENESS PROPERTY IN NORMED LINEAR SPACE

Definition 0.11.3. A normed linear space that is complete with respect to the metric induced
by the norm is called a Banach space.

Theorem 0.11.3. Let (X, k.k) be a Banach space and M be a linear subspace of X. Then M is
complete if and only if M is closed in X.

Proof. Assume that M is complete. We have to show that M is closed. To that end, let x ∈ M.
Then there is a sequence (xn ) in M such that kxn − xk → 0. Since (xn )converges, it is Cauchy.
Completeness of M guarantees the existence of an element y ∈ M such that kxn − yk → 0. By
uniqueness of limits, x = y. Hence x ∈ M and, consequently, M is closed.
Assume that M is closed. We show that M is complete. Let (xn ) be a Cauchy sequence in
M. Then (xn ) is a Cauchy sequence in X. Since X is complete, there is an element x ∈ X such
that kxn − xk → 0. But then x ∈ M since M is closed. Hence M is complete.

Proposition 0.11.4. (a) Show that X = Rn or Cn is complete if we choose metric

!1/2
n
d(x, y) = ∑ (ξi − ηi)2 ,
i=1

where x = (ξi ), y = (ηi ) ∈ X.

(b) Show that X = `∞ is complete if we choose metric

d(x, y) = sup|(ξi − ηi )|,

i∈N

where x = (ξi ), y = (ηi ) ∈ X.

(c) Show that X = ` p , where p ≥ 1 is complete if we choose metric
!1/p
n
d(x, y) = ∑ (ξi − ηi) p ,
i=1

where x = (ξi ), y = (ηi ) ∈ X.

1/2
Proof. (a) Here d(x, y) = ∑ni=1 (ξi − ηi )2 , where x = (ξi ), y = (ηi ).
(m) (m)
Suppose that (xn ) is a Cauchy sequence in X = Rn , where xm = (ξ1 , ...ξn ) so for given
ε > 0 there exists an integer N such that for m, n > N, we have
!1/2
n
(m) (r)
d(xm , xr ) = ∑ (ξi − ξi )2 < ε. (0.11.1)
i=1

(m) (r) (m) (r)

Squaring, for m, r > N and i = 1, ..., n we get ∑ni=1 (ξi − ξi )2 < ε 2 . Hence |(ξi − ξi )| <
ε.
(1) (2)
This shows that for each fixed i = 1, ..., n, the sequence (ξi , ξi , ...) is a Cauchy sequence
(m)
of real numbers. Since R is complete, every Cauchy sequence in R converges, say ξi → ξi
as m → ∞.
As x = (ξ1 , ξ2 , ..., ξn ) ∈ Rn , so take n → ∞ in (0.11.1), we get d(xm , x) ≤ ε, for m > N.
Hence the result.

32
0.11. COMPLETENESS PROPERTY IN NORMED LINEAR SPACE

(m) (m)
(b) Let (xm ) be a Cauchy sequence in `∞ , where xm = (ξ1 , ξ2 , ...) ∈ `∞ . Since the metric
defined on `∞ is given by d(x, y) = supi |ξi − ηi |, where x = (ξi ) and y = ηi . For a given ε > 0,
(m) (n)
there is an N such that for all m, n ≥ N, we have d(xm , xn ) = supi |ξi − ξi | < ε.
(m) (n)
Therefore, for every fixed j, and m, n ≥ N we have |ξi − ξi | < ε. Hence for every fixed
(1) (2)
j, the sequence (ξ j , ξ j , ...) is a Cauchy sequence in R. As R is complete, so it converges, say
(m) (m) (m) (m)
ξj → ξ j as m → ∞, we have |ξ j − ξ j | < ε, whenever m > N. Since xm = (ξ1 , ξ2 , ...) ∈
(m)
`∞ , there is a real number km such that |ξ j | ≤ km for all j. Hence by triangle inequality
(m) (m)
|ξ j | ≤ |ξ j − ξ j | + |ξ j | ≤ ε + km for m > N. This inequality holds for every j, and the right
hand side does not involve j. Hence (ξ j ) is a bounded sequence of numbers. This implies
(m)
that x = (ξ j ) ∈ `∞ . Also, we obtain d(xm , x) = supi |ξ j − ξ j | ≤ ε, for m > N. This show that
xm → x. Hence `∞ is complete.
(c) Let (xn ) be a a Cauchy sequence in ` p . Suppose that xn = (xn(1) , xn(2) ...). Then for given

1
p p < ε for all m, n ≥ N.
ε > 0, there exists N(ε) ∈ N such that d(xn , xm ) = ∑∞ |x
i=1 n(i) − x m(i) |
For each fixed index i, we have |xn(i) − xm(i) | < ε for all m, n ≥ N, that is for each fixed
index i, (xn(i) )∞ 1 is a Cauchy sequence in R. Since R is complete, there exists x(i) ∈ R such that
xn(i) → x(i) as n → ∞, that is for each fixed index i, (xn(i) )∞ 1 is a Cauchy sequence in R. Since
R is complete, there exists x(i) ∈ R such that xn(i) → x(i) as n → ∞. Define x = (x1 , x2 , ...).
Now, we show that x ∈ ` p and xn → x. For each k ∈ N, ∑ki=1 |xn(i) − xm(i) | p ≤ [d(xn , xm )] p =


p < ε p , that is k |x
∑i=1 n(i) − xm(i) | p < ε p , for all k = 1, 2, .... Keep k and n ≥


∑∞ i=1 |xn(i) − xm(i) |
N fixed and let m → ∞, we have ∑ki=1 |xn(i) − xi | p < ε p . Now, letting k → ∞, then for all n ≥ N,
we have ∑∞ p p
i=1 |xn(i) − xm(i) | < ε , which means that xn − x ∈ ` p . Since xn ∈ ell p , we have to
show that x = (x − xn ) + xn ∈ ` p . Consider |xi | = |xi − xn(i) + xn(i) |, using Minkowski inequality,
we have (∑i |xi | p )(1/p) = (∑i |xi − xn(i) + xn(i) | p )(1/p) ≤ (∑i |xi − xn(i) | p )1/p + (∑i xn(i) | p )(1/p) ).
Now, both the series in right converges so series in left converges. Hence x = (x − xn ) + xn ∈ ` p
Also, xn → x. Hence the result.

1
Example 0.11.5. 1. The space Rn equipped with k.k p defined as kxk2 = ∑i∈N |xi |2 2 , is a
Banach space.

1
(In metric d(x, y) = ∑ni=1 (χi − ηi )2 2 , where 1 ≤ p < ∞, x = (χi ) and y = (ηi )). It
is easy to see that kxk2 (respectively, d(x, y)) defined above is a normed linear space
(respectively, metric space).
Now, we have to show that it is complete. Let (xn )n∈N be a sequence in Rn . Then we have
kxk − xm k < ε for all m, k ≥ N, that is, ∑ni=1 | kxki − xmi k |2 < ε 2 for all m, k ≥ N. Thus
it follows that for every i ∈ {1, 2, ..., n}, |xki − xmi | < ε for all m, k ≥ N, that is sequence
(xmi ) is a Cauchy sequence in R and consequently it is convergent. Let xi = lim xki . Then
k→∞
x = (x1 , ..., xn ) ∈ Rn . Now, take k → ∞, we have ∑ni=1 | kxi − xmi k |2 < ε 2 for all m ≥ N,
that is kx − xm k2 ≤ ε for all m ≥ N, and so x = lim xm in the normed linear space.
m→∞

2. Let 1 ≤ p < ∞. The sequence space ` p is a Banach space. Because of the importance of
this space, we give a detailed proof of its completeness.
The classical sequence space ` p is complete.
Proof. Let (xn ) be a a Cauchy sequence in ` p . Suppose that xn = (xn(1) , xn(2) ...). Then

1
p p <ε
for given ε > 0, there exists N(ε) ∈ N such that kxn − xm k p = ∑∞ |x
i=1 n(i) − xm(i) |

33
0.11. COMPLETENESS PROPERTY IN NORMED LINEAR SPACE

for all m, n ≥ N.
For each fixed index i, we have |xn(i) − xm(i) | < ε for all m, n ≥ N, that is for each
fixed index i, (xn(i) )∞ 1 is a Cauchy sequence in R. Since R is complete, there exists
x(i) ∈ R such that xn(i) → x(i) as n → ∞, that is for each fixed index i, (xn(i) )∞ 1 is a
Cauchy sequence in R. Since R is complete, there exists x(i) ∈ R such that xn(i) →
x(i) as n → ∞. Define x = (x1 , x2 , ...). Now, we show that x ∈ ` p and xn → x. For

1
1
each k ∈ N, ∑ki=1 |xn(i) − xm(i) | p p ≤ kxn − xm k p = ∑∞ i=1 |xn(i) − xm(i) |
p p < ε, that is

∑ki=1 |xn(i) − xm(i) | p < ε p , for all k = 1, 2, .... Keep k and n ≥ N fixed and let m → ∞, we
have ∑ki=1 |xn(i) − xi | p < ε p . Now, letting k → ∞, then for all n ≥ N, we have ∑∞ i=1 |xn(i) −
xm(i) | p < ε p , which means that xn − x ∈ ` p . Since xn ∈ ` p , we have x = (x − xn ) + xn ∈ ` p .
Also, xn → x. Hence the result.
3. Space `∞ is complete.
(m) (m)
Proof. Let (xm ) be a Cauchy sequence in `∞ , where xm = (ξ1 , ξ2 , ...) ∈ `∞ . Since the
metric defined on `∞ is given by d(x, y) = supi |ξi − ηi |, where x = (ξi ) and y = ηi . For
(m)
a given ε > 0, there is an N such that for all m, n ≥ N, we have d(xm , xn ) = supi |ξi −
(m)
ξi | < ε.
Therefore, for every fixed j, and m, n ≥ N we have
(m) (m)
|ξi − ξi | < ε.
(1) (2)
Hence for every fixed j, the sequence (ξ j , ξ j , ...) is a Cauchy sequence in R. As R is
(m) (m)
complete, so it converges, say ξ j → ξ j as m → ∞, we have |ξ j − ξ j | < ε, whenever
(m) (m) (m)
m > N. Since xm = (ξ1 , ξ2 , ...) ∈ `∞ , there is a real number km such that |ξ j | ≤ km
(m) (m)
for all j. Hence by triangle inequality |ξ j | ≤ |ξ j − ξ j | + |ξ j | ≤ ε + km for m > N.
This inequality holds for every j, and the right hand side does not involve j. Hence (ξ j )(
is a bounded sequence of numbers. This implies that x = (ξ j ) ∈ `∞ . Also, we obtain
(m)
d(xm , x) = supi |ξ j − ξ j | ≤ ε, for m > N. This show that xm → x. Hence `∞ is complete.

4. Completeness of c. The space c consists of all convergent sequences x = (ξi ) of complex

numbers with metric induced from `∞ . The space c is complete.
Proof. c is a subspace of `∞ and if c is closed in `∞ , so it is complete.
(n)
Consider x = (ξ j ) ∈ c, the closure of c. So there are xn = ξ j ∈ c such that xn → x.
(n)
Hence for a given ε > 0, there is an N such that for n ≥ N and all j, we have |ξ j − ξ j | ≤
(N)
d(xn , x) < ε/3. Since xN ∈ c, its terms ξ j form a convergent sequence. Such a sequence
(N) (N)
is a Cauchy sequence. hence there is an N1 such that |ξ j − ξk | < ε/3 for j, k ≥ N1 .
Using the triangle inequality, we have
(N) (N) (N) (N)
|ξ j − ξk | ≤ |ξ j − ξ j | + |ξ j − ξk | + |ξk − ξk | < ε.

This shows that the sequence x = (ξ j ) is convergent. Hence x ∈ c. Since x ∈ c, this proves
that closedness of c in `∞ and it is complete.
Some examples of not complete space.

34
0.11. COMPLETENESS PROPERTY IN NORMED LINEAR SPACE

R1
Figure 1: This is the graph of C[a, b] induced with d(x, y) = 0 |x(t) − y(t)|dt

Example 0.11.6. 1. Let X be the set of all continuous real-valued functions on I = [0, 1] and
d(x, y) = 01 |x(t) − y(t)|dt. Then (X, d) is not complete.
R

Proof.
The functions 
 0,t ∈ [0, 1/2]

xm (t) = m(x − 12 ), 12 < t < am = 1/2 + 1/m

 1,t ∈ [a , 1],
m

form a Cauchy sequence because d(xm , xn ) is the area of the triangle in Figure 1 and
for every given ε > 0, d(xm , xn ), when m, n > 1/ε. Now, we’ll show that this Cauchy
sequence does not converge in X. For every x ∈ X, we have
Z 1
d(xm , x) = |xm (t) − x(t)|dt
0
Z 1/2 Z am Z 1
= |x(t)|dt + |xm (t) − x(t)| + |1 − x(t)|dt.
0 1/2 am

Each the integrands are nonnegative, so is each integral on the right. Hence d(xm , x) → 0
would imply that each integral approaches zero and, since x(t) is continuous, we should
have (
0,t ∈ [0, 1/2)
x(t) =
1,t ∈ (1/2, 1]
But this is impossible for a continuous function. Hence (xm ) does not converge, that is,
does not have a limit in X. This proves that X is not complete.
Example 0.11.7. Show that (C[a, b], d∞ ) is a complete metric space, where d∞ = sup |x(t) −
y(t)|, t ∈ [a, b].

Proposition 0.11.8. Show that the set of all real numbers constitutes an incomplete metric
space if we choose the metric

d(x, y) = |tan−1 x − tan−1 y|.

35
0.11. COMPLETENESS PROPERTY IN NORMED LINEAR SPACE

Proof. Consider a sequence (xn ) on real line, where xn = n. Then (xn ) is Cauchy but not
convergent. We know that for a given ε > 0, there exists an integer N such that for n > N,

|tan−1 n − (π/2)| < ε/2.

Thus for n, m > N, we get

d(xm , xn ) = |tan−1 m − tan−1 n| ≤ |tan−1 m − (π/2)| + |(π/2) − tan−1 n| < ε.

On the other hand, (xn ) is not a convergent sequence. If not there exists x ∈ R such that
d(xn , x) → 0 as n → ∞. But (using |x − y| ≥ ||x| − |y||)

d(xn , x) = |tan−1 n − tan−1 x| ≥ ||tan−1 n − (π/2)| − |(π/2) − tan−1 x||.

So tan−1 x = π/2. But this a contradiction because for any x ∈ R, tan−1 x < π/2.

36
0.12. CONVERGENCE OF SERIES IN BANACH SPACE

0.12 Convergence of series in Banach space

Definition 0.12.1. Let (xn ) be a sequence in a normed linear space (X, k.k). If the sequence
(sn ) of partial sums converges to s, then we say that the series ∑∞ i=1 xi converges and that its
sum is s. In this case we write ∑∞ x
i=1 i = s. The series ∑∞
x
i=1 i is said to be absolutely convergent
∞
if ∑i=1 kxi k < ∞.

Theorem 0.12.1. A normed linear space (X, k.k) is a Banach space if and only if every abso-
lutely convergent series in X is convergent.

Proof. Let X be a Banach space and ∑∞j=1 x j < ∞. Now, we have to show that ∑∞j=1 x j
converges. Suppose for a given ε > 0, and for each n ∈ N, let sn = ∑nj=1 x j . Let K be a positive
integer such that ∑∞j=K+1 x j < ε. Then for all m > n > K, we have

m n m
ksm − sn k = ∑ xj − ∑ xj = ∑ xj
j=1 j=1 j=n+1
m
≤ ∑ xj
j=n+1
∞
.
≤ ∑ xj
j=n+1
∞
≤ ∑ x j < ε.
j=K+1

Hence the sequence (sn ) of partial sums form a Cauchy sequence in X. SInce X is complete,
the sequence (sn ) converges to some element s ∈ X, that is, the series ∑∞j=1 x j converges.
Conversely, assume that (X, k.k) is a normed linear space in which every absolutely con-
vergent series converges. We have to show that X is complete. Let (xn ) be a Cauchy se-
quence in X. Then there exists n1 ∈ N such that kxn1 − xm k < 12 , where m > n1 . Simi-
larly, there is an n2 ∈ N with n2 > n1 such that kxn2 − xm k < 212 whenever m > n2 . Contin-
uing in this way, we get natural numbers n1 < n2 < ... such that kxnk − xm k < 21k whenever
m > nk . In particular, we have that for each k ∈ N, xnk+1 − xnk < 2−k . For each k ∈ N, let
yk = xnk+1 − xnk . Then ∑nk=1 kyk k = ∑nk=1 xnk+1 − xnk < ∑nk=1 2−k . Hence ∑nk=1 kyk k < ∞, that
is series ∑nk=1 yk is absolutely convergent and hence by our assumption, the series ∑nk=1 yk
j
is convergent in X, that is, there is an x ∈ X such that s j = ∑k=1 yk → s. It follows that
j j
s j = ∑k=1 yk = ∑k=1 (xnk+1 − xnk ) = xn j+1 − xn1 → s as j → ∞. Hence xn j+1 → s + xn1 . Thus the
subsequence (xnk ) of (xn ) converges in X. But if a Cauchy sequence has a convergent subse-
quence, then the sequence itself also converges (to the same limit as the subsequence). It thus
follows that the sequence (xn ) also converges in X. Hence X is complete.

Theorem 0.12.2. Let a, b ∈ R and a < b. The space (C[a, b], k.k∞ ) is a Banach space.

Proof. It is clear that linear combinations of continuous functions are continuous, so that
C[a, b] is a vector space. Define k f k∞ = supx∈[a,b] | f (x)|, f ∈ C[a, b]. Then C[a, b] is a normed
linear space.

37
0.12. CONVERGENCE OF SERIES IN BANACH SPACE

Now, we have to show the completeness. Let ( fn ) be a Cauchy sequence in C[a, b]. Suppose
for a given ε > 0, there exists a N ∈ N such that for all x ∈ [a, b], we have

| fk (x) − fn (x)| ≤ k fk − fn k∞ < ε, (0.12.1)

for all k, n ≥ N. Thus ( fn ) is a Cauchy sequence in K. Since K = R or C is complete, the limit

f (x) = lim fn (x) exists. Now, we must show that limit is continuous. Take k → ∞ in (0.12.1),
we have for all x ∈ [a, b], | f (x) − fn (x)| < ε, for all n ≥ N.
Now, we’ll show that f is continuous on [a, b]. Let x0 ∈ [a, b]. Given any t > 0, let ε = t/3.
As fN is continuous on [a, b], it follows that there exists a δ > 0 such that for all x ∈ [a, b] such
that |x − x0 | < δ , | fN (x) − fN (x0 )| < ε. Consequently, for all x ∈ [a, b] such that |x − x0 | < δ ,
we have

| f (x) − f (x0 )| ≤ | f (x) − fN (x) + fN (x) − fN (x0 ) + fN (x0 ) − f (x0 )|

≤ | f (x) − fN (x)| + | fN (x) − fN (x0 )| + | fN (x0 ) − f (x0 )|
< ε + ε + ε = t.

Hence f is continuous and it belongs to C[a, b].Hence k f − fn k∞ < ε, for all n ≥ N and so fn
converges to f in the normed space C[a, b].

38
0.13. CANTOR INTERSECTION THEOREM

0.13 Cantor intersection theorem

Definition 0.13.1. Let (X, d) be a metric space and A ⊂ X. The diameter of a set A is δ (A) =
diam(A) = sup{d(x, y) : x, y ∈ A}
Example 0.13.1. For the usual (Euclidean) metric space, A = {y : y = x2 , 0 ≤ x ≤ 5} = [0, 25]
the diameter for this set is 25.
That is, the greatest distance between any two points. If the set were circular in shape, then
this definition would make sense. A set is called bounded if its diameter is finite.
Proposition 0.13.2. In the usual metric space (R, d), a subset A is bounded if and only if it is
bounded above and bounded below.
Proof. We know
diam(A) = sup{d(x, y) : x, y ∈ A} = c < ∞
⇔ d(x, y) ≤ c ⇔ |x − y| ≤ c
for all x, y ∈ A. For all x, y ∈ A, we have
−c ≤ x − y ≤ c ⇔ −b ≤ x ≤ b
for all x, b = y − c.
Definition 0.13.2. Let (X, d) be a metric space and A be a nonempty subset of X. The diameter
of A, denoted by δ (A) = sup{d(x, y) : x, y ∈ A}, that is diameter of A is the supremum of the
set of all distances between points of A.
The distance between a point p ∈ X and A is denoted by d(p, A) and defined by d(p, A) =
inf{d(p, x) : x ∈ A}.
The distance between two nonempty subsets of A and B of a metric space (X, d) defined by
d(A, B) = inf{d(x, y) : x ∈ A, y ∈ B}.
A sequence (An ) of subsets of X is said to be monotonic decreasing iff A1 ⊃ A2 ⊃ ..., such
a sequence is also called a nested sequence.
Remark 0.13.1. If A and B are two non-empty subsets such that A ∩ B = 0/ then d(A, B) > 0
but the converse may not be true,

Theorem 0.13.3. Cantor’s Intersection Theorem Let (X, d) be a metric space and (Fn ) be a
nested sequence of non-empty closed subsets of X such that diam(Fn ) → 0. Then X is complete
iff ∩∞
n=1 Fn consists of exactly one point.

Proof. Let X be complete. For each n, we choose xn ∈ Fn . Since diam(Fn ) → 0, for every
ε > 0, there exists a positive integer m0 such that diam(Fm0 ) < ε. Again since (Fn ) is a nested
sequence, we have for n, m > m0 , Fn , Fm ⊂ Fm0 and hence xn , xm ∈ Fm0 , d(xn , xm ) < ε. Thus
(xn ) is a Cauchy sequence. Since X is complete, xn → x0 for some x0 ∈ X.
Now, we have to show that x0 ∈ ∩∞ n=1 Fn . Assume that m ∈ N, for n > m, we have xn ∈
Fn ⊂ Fm . Since xn → x0 , the sequence (xn ) is in every neighbourhood of x0 . As, xn ∈ Fm , so x0
contains an infinite number of points of Fm . Thus x0 is a limit point of Fm . Since Fm is closed,
x0 ∈ Fm and m is arbitrary, we have x0 ∈ ∩∞ n=1 Fn .
Now, suppose that there is another point x1 (6= x0 ) ∈ ∩∞ n=1 Fn . Therefore, d(x0 , x1 ) < diam(Fn )
for every n. Therefore d(x0 , x1 ) = 0, since diam(Fn ) → 0. Hence x0 = x1 and so ∩∞ n=1 Fn = {x0 }.
Conversely, let ∩∞ F
n=1 n consists of a single point for every nested sequence (F n ) of nonempty
closed subsets of Fn of X such that diam(Fn ) → 0.

39
0.13. CANTOR INTERSECTION THEOREM

Now, we have to prove that X is complete. Let (xn ) be a Cauchy sequence in X. Take
S1 = {x1 , x2 , ...}, S2 = {x2 , x3 , ...} ,..., Sn = {xn , xn+1 , ...}. Since (xn ) is a Cauchy sequence, for
given ε > 0, there exists a positive integer m0 such that n, m > m0 , d(xn , xm ) < ε. It follows
that for n > m0 , diam(Sn ) < ε and consequently, diam(Sn ) → 0.
Also, S1 ⊃ S2 ⊃ ... so that S1 ⊃ S2 ⊃ ... [as A ⊃ B =⇒ A ⊃ B]. Therefore, diam(Sn ) =
diam(Sn ). Hence (Sn ) is a nested sequence of closed subsets of X whose diameter tends to
zero. Then by hypothesis, there exists a unique point x0 ∈ X such that x0 ∈ ∩∞ n=1 Sn . We claim
that (xn ) converges to x0 . Since diam(Sn ) → 0, for given ε > 0, there exists m0 ∈ N such that
diam(Sm0 ) < ε and consequently, for n > m0 , xn , x0 ∈ Sm0 , d(xn , x0 ) < ε. Hence (xn ) converges
to x0 . Thus it is shown that every Cauchy sequence in X converges to a point in X. Hence X is
complete.

40
0.14. SEPARABLE SPACE

0.14 Separable Space

If (xk ) is a sequence in a normed space X, we can associate with (xk ) the sequence (sn ) of
partial sums sn = x1 + x2 + ... + xn .
If sn is convergent, we say ksn − sk → 0; then the infinite series ∑∞ i=1 si = x1 + x2 + ... is
said to converge, s is called the sum of the series.
If kx1 k+kx2 k+... converges, the given series is said to be absolutely convergent. However,
in a normed space X, absolute convergence implies convergence if and only if X is complete.
The concept of convergence of a series can be used to define a ”basis” as follows. If a
normed space X contains a sequence {en } with the property that for every x ∈ X there is a
unique sequence of scalars {α} such that kx − (α1 e1 + ... + αn en )k → 0, then {en } is called a
Schauder basis (or basis) for X.
The series ∑∞ i=1 αi ei , which has sum x is then called the expansion of x with respect to {en }
and we write x = ∑∞ i=1 αi ei
For example, ` p has a Schauder basis, namely {en }, where en = (δn j ); that is, (en ) is the
sequence whose n-th term is 1 and all other terms are zero; thus

e1 = (1; 0; 0; 0; ...)
e2 = (0; 1; 0; 0; ...)
e3 = (0; 0; 1; 0; ...)
.......

If a normed space X has a Schauder basis, then X is separable. Conversely, does every
separable Banach space have a Schauder basis? This is a famous question raised by Banach
himself. Almost all known separable Banach spaces had been shown to possess a Schauder
basis. Nevertheless, the surprising answer to the question is no and answer was given by P.
Enflo (1973) who was able to construct a separable Banach space which has no Schauder
basis.
Definition 0.14.1. A subset M of a metric space X is said to be dense in X if M = X.
X is said to be separable if it has a countable subset which is dense in X.
That is, we can have limit points of A within it and the result can equal to the parent set. One
good example is the set of rationals and the real set. We all know that the set of rationals is not
complete. In Analysis, real numbers can be constructed using Dedekind cuts or the addition of
limits to every Cauchy sequence. That is, Q = R or that the set of rationals are dense in the set
of reals. The complex plane, too, can be separated from the irrational real and imaginary parts
against the rational ones. This has importance in the theory of operators.
More technically, if M is dense in X, then for each x ∈ X and for every ε > 0, B(x0 ; ε)
will contain points of M; or, in other words, in this case there is no point x0 ∈ X which has
a neighbourhood that does not contain points of M. This is the direct consequence of the
definition of a limit point.
Examples [1] The real line R is separable since the set Q of rational numbers is a countable
dense subset of R.
[2] The complex plane C is separable since the set of all complex numbers with rational
real and imaginary parts is a countable dense subset of C.

Theorem 0.14.1. Show that the space ` p is separable for 1 ≤ p < ∞.

41
0.14. SEPARABLE SPACE

Proof. We will proceed as follows: we will first construct a countable subset then basing our
argument on the fact that Q is dense in R, construct limits for every sequence of elements of
` p which will be limit points of sequences in the constructed subset.
Let M be the set of all sequences x of the form
x = (χ1 ; χ2 ; . . . ; χn ; 0; 0; . . .),
where n is any integer. Now, we can assume that χi ∈ Q for all i since we are only discussing
real or complex sequences. Since Q is countable, Qn is therefore countable, leading us to a
countable M. That justifies one part of the definition. To prove that M = ` p . Take y = (ηi ) ∈ ` p .
Since this is convergent, we have ∑∞ p p
k=n+1 |ηk | < ε /2. Since Q = R, for each ηk there is
rational χk close to it. Hence we can find a x ∈ M such that ∑nk=1 |χk − ηk | p < ε p /2. Since
1
p p p p n
d(x, y) = (∑∞ ∞
k=1 |χk − ηk | ) . We therefore have [d(x, y)] = ∑k=n+1 |ηk − 0| + ∑k=1 |χk −
ηk | p < ε p /2 + ε p /2 = ε p or that d(x, y) < ε. That is, every sequence x will have a limit point
y. Hence, M = ` p .
or
Take M to be the set of all sequences with rational entries such that all but a finite number of
the entries are zero. (If the entries are complex, take for M the set of finitely nonzero sequences
with rational real and imaginary parts.) It is clear that M is countable. We show that M is dense
in ` p . Let ε > 0 and x = (xn ) ∈ ` p . Then there is an N such that
∞
∑ |xk | p < ε/2.
k=N+1

Now, for each 1 ≤ k ≤ N, there is a rational number qk such that |xk − qk | p < ε/(2N). Set
q = (q1 , q2 , . . . , qN , 0, 0, . . .). Then q ∈ M and
N ∞
kx − qk pp = ∑ |xk − qk | p + ∑ |xk | p < ε.
k=1 k=N+1
Hence M is dense in ` p .

This should in no way mean that every collection of convergent sequences forms a separable
set.
Theorem 0.14.2. A normed linear space X is separable if and only if it contains a countable
set B such that lin(B) = X.
Proof. Assume that X is separable and let A be a countable dense subset of X. Since the linear
hull of A, lin(A), contains A and A is dense in X , we have that lin(A) is dense in X , that is,
lin(A) = X.
Conversely, assume that X contains a countable set B such that lin(B) = X. Let B = {xn :
n ∈ N}.
Assume first that F = R, and put
n
C = { ∑ λ j x j : λ j ∈ Q; j = 1; 2; . . . ; n; n ∈ N}.
j=1

We first show that C is a countable subset of X . The set Q × B is countable and conse-
quently, the family F of all finite subsets of Q × B is also countable. The mapping
n
{(λ1 , x1 ), . . . , (λn , xn )} 7→ ∑ λ jx j
j=1

42
0.14. SEPARABLE SPACE

maps F onto C . Hence C is countable.

Next, we show that C is dense in X. Let x ∈ X and ε > 0. Since lin(B) = X, we can find an
n ∈ N, points x1 ; x2 ; . . . ; xn ∈ B and λ1 , ..., . . . λn ∈ F such that
n
x − ∑ λ j x j < ε/2.
j=1

Since Q is dense in R, for each λi ∈ R, we can find µi ∈ Q such that

ε
|λi − µi | < , i = 1, 2, . . . , n.
2n(1 + kxi k)
Hence,
n n n n
x − ∑ µ jx j ≤ x − ∑ λ jx j + ∑ λ jx j − ∑ µ jx j
j=1 j=1 j=1 j=1
n
< ε/2 + ∑ |λ j − µ j | x j
j=1
n ε xj
< ε/2 + ∑
j=1 2n(1 + kxi k)
< ε/2 + ε/2 = ε.

This shows that C is dense in X.

If F = C, the set C is that of finite linear combinations with coefficients being those complex
numbers with rational real and imaginary parts.
We now give another argument based on Theorem 0.14.2 to show that the sequence space
` p , where 1 ≤ p < ∞, is separable. Let en = (δnm )m∈N where
(
1, n = m
δnm =
0, n 6= m.

Clearly, en ∈ ` p . Let ε > 0 and x = (xn ) ∈ ` p . Then there is a natural number N such that
p
∑∞
k=n+1 |xk | < ε, for all n ≥ N.
Now, if n ≥ N, then Hence lin({en : n ∈ N}) = ` p . Of course, the set {en : n ∈ N} is
countable.

Theorem 0.14.3. The space `∞ is not separable.

Proof. Let y = (ηi ), where ηi = 0, 1. There are uncountably many y. If we put small balls
with radius 13 at y, they will not intersect. It follows that if M ⊂ `∞ is dense in `∞ , then M is
uncountable. Therefore `∞ is not separable.

43
0.15. FINITE DIMENSIONAL NORMED SPACE

0.15 Finite Dimensional Normed Space

Are finite dimensional normed spaces simpler than infinite dimensional ones? In what respect?
These questions are rather natural. They are important since finite dimensional spaces and
subspaces play a role in various considerations (for instance, in approximation theory and
spectral theory). Quite a number of interesting things can be said in this connection. Hence
it is worthwhile to collect some relevant facts, for their own sake and as tools for our further
work. This is our program in this section and the next one.
A source for results of the desired type is the following lemma. Very roughly speaking it
states that in the case of linear independence of vectors we cannot find a linear combination
that involves large scalars but represents a small vector.
Lemma 0.15.1. Let {x1 , ..., xn } be a linearly independent set of vectors in a normed space X
(of any dimension). Then there is a number c > 0 such that for every choice of scalars α1 , ...αn
we have
kα1 x1 + ... + αn xn k ≥ c(|α1 | + ...|αn |), (c > 0). (0.15.1)
Proof. We write s = |α1 | + ... + |αn |. If s = 0, all α j are zero, so that (0.15.1) holds for any
c. Let s > 0. Then (0.15.1) is equivalent to the inequality which we obtain from (0.15.1) by
dividing by s and writing β j = α j /s, that is,
n
kβ1 x1 + ... + βn xn k ≥ c, ( ∑ |β j | = 1). (0.15.2)
j=1

Hence it suffices to prove the existence of a c > 0 such that (0.15.2) holds for every n-tuple of
scalars β1 , ..., βn with ∑nj=1 |β j | = 1.
Suppose that this is false. Then there exists a sequence (ym ) of vectors
n
(m) (m) (m)
ym = β1 x1 + ... + βn xn = ∑ |β j |xi ,
j=1
(m)
where ∑nj=1 |β j | = 1 such that kym k → 0 as m → ∞.
(m) (m) (m)
Since ∑nj=1 |β j | = 1, we have |β j | ≤ 1. Hence for each fixed j the sequence (β j ) =
(1) (2) (m)
(β j , β j , ...) is bounded. Consequently, by the Bolzano-Weierstrass theorem, (β1 ) has a
convergent subsequence. Let β1 denote the limit of that subsequence, and let
n
(m) (m)
(y1,m ) = γ1 x1 + ∑ |βi |xi
i=2
(m)
denote the corresponding subsequence of (ym ) such that γ1 → β1 .
(m) (m) (m)
By the same argument, (y1,m ) has a subsequence (y2,m ) = γ1 x1 + γ2 x2 + ∑nj=3 |β j |x j
(m)
for which the corresponding subsequence of scalars β2 converges; let β2 denote the limit.
Continuing in this way, after n steps we obtain a subsequence (yn,m ) = (yn,1 , yn,2 , ...) of
(m) (m) (m)
(ym ) whose terms are of the form yn,m = ∑nj=1 γ j x j , where ∑nj=1 |γ j | = 1, with scalars γ j
(m)
satisfying γ j → β j as m → ∞. Hence as m → ∞, we have yn,m → y = ∑nj=1 β j x j , where
∑nj=1 |β j | = 1, so that not all β j can be zero. Since {x1 , ...xn } is a linearly independent set,
we thus have y 6= 0. On the other hand, yn,m → y implies kyn,m k → kyk, by the continuity of
the norm. Since kym k → 0 by assumption and (yn,m ) is a subsequence of (ym ), we must have
kyn,m k → 0. Hence kyk = 0, so that y = 0. This contradicts y 6= 0. Hence we get the result.

44
0.15. FINITE DIMENSIONAL NORMED SPACE

Theorem 0.15.2. Every finite dimensional subspace Y of a normed space X is complete. In

particular, every finite dimensional normed space is complete.

Proof. We consider an arbitrary Cauchy sequence (ym ) in Y and show that it is convergent in
Y ; the limit will be denoted by y. Let dimY = n and {e1 , ..., en } any basis for Y . Then each ym
(m)
has a unique representation of the form ym = ∑nj=1 α j e j .
Since (ym ) is a Cauchy sequence, for every ε > 0 there is an N such that kym − yr k < ε
when m, r > N. From this and Lemma 0.15.1 we have for some c > 0,
n n
(m) (r) (m) (r)
ε > kym − yr k = ∑ (α j − α j )e j ≥ c ∑ |α j − α j |,
j=1 j=1

where m, r > N. Division by c > 0, we have

n
(m) (r) (m) (r)
|α j −αj | ≤ ∑ |α j − α j | < ε/c,
j=1

(m) (1)
m, r > N. This shows that each of the n sequences (α j ) = (α j , ...), j = 1, ..., n is Cauchy in
(m)
R or C. Hence it converges, let α j → α j denote the limit. Using these n limits α1 , ...αn , we
define y = α1 e1 + ... + αn en . Clearly, y ∈ Y . Furthermore,
n n
(m) (r) (m) (r)
kym − yk = ∑ (α j − α j )e j ≤ ∑ |α j −αj | ej .
j=1 j=1

(m)
As α j → α j , kym − yk → 0, that is ym → y. This shows that (ym ) is convergent in Y . Since
(ym ) was an arbitrary Cauchy sequence in Y , this proves that Y is complete.

Theorem 0.15.3. A subspace M of a complete metric space X is itself complete if and only if
the set M is closed in X.
Using above two theorems, we have
Theorem 0.15.4. Every finite dimensional subspace Y of a normed space X is closed in X.
Note that infinite dimensional subspaces need not be closed or complete.
Example 0.15.5. Let X = C[0, 1], k.k = maxt∈[0,1] |x(t)| and Y = span{x0 , x1 , ...}, where x j (t) =
t j , so that Y is the set of all polynomials, that is Y = span{1,t,t 2 , ...}. Y ⊂ C[0, 1]. Y is not
closed as yn (t) = 1 + t + t 2 /(2!) + ... + t n /(n!) → et ∈
/ Y . Hence Y is not complete.

45
0.16. EQUIVALENT NORMS

0.16 Equivalent Norms

Definition 0.16.1. Let k.k and k.k0 be two different norms defined on the same linear space X.
We say that k.k is equivalent to k.k0 if there are positive numbers α and β such that

α kxk ≤ kxk0 ≤ β kxk ,

for all x ∈ X.
Example 0.16.1. Let X = R2 , x = (ξ1 , ξ2 ) ∈ X and define

kxk1 = |ξ1 | + |ξ2 |

and q
kxk2 = |ξ1 |2 + |ξ2 |2 .
Here [By Holder’s inequaity]
2 q
2 1/2 2 1/2
kxk1 = ∑ (1)|ξi | ≤ (∑ 1 ) (∑|ξi | ) = 2 kxk2
i=1

or √1 kxk≤ kxk2 .
2 1
p
Also, kxk2 = |ξ1 |2 + |ξ2 |2 ≤ |ξ1 | + |ξ2 | = kxk1 . Hence
1
√ kxk1 ≤ kxk2 ≤ 1 kxk1 .
2
Therefore kxk1 and kxk2 are equivalent.
Example 0.16.2. Let X = Rn , for each x = (x1 , ..., xn ) ∈ X, let
!1
n n 2

kxk1 = ∑ |xi | , kxk2 = ∑ |xi|2 and kxk∞ = max |xi |

1≤i≤n
i=1 i=1

It is easy to check that kxk1 ; kxk2 and kxk∞ are norms on X.

Now, we show that these norms are equivalent.
Equivalent of kxk1 ; kxk∞ : Let x = (x1 , x2 , ..., xn ) ∈ X. For each k = 1, 2, ..., n,
n
|xk | ≤ ∑ |xi |
i=1
=⇒ max |xk |
1≤k≤n
n
≤ ∑ |xi |
i=1
⇔ kxk∞ ≤ kxk1 .

46
0.16. EQUIVALENT NORMS

Also, for k = 1, ..., n, we have

|xk | ≤ max |xk |

1≤k≤n
n
= kxk∞ =⇒ ∑ |xi|
i=1
n
≤ ∑ kxk∞
i=1
= n kxk∞
⇔ kxk1 ≤ n kxk∞ .

Hence kxk∞ ≤ kxk1 ≤ n kxk∞ .

Equivalent of kxk2 ; kxk∞ :
..., n, |xk | ≤ kxk∞ =⇒ |xk |2 ≤ (kxk∞ )2 =⇒
Let x = (x1 , x2 , ..., xn ) ∈ X. For each k = 1, 2, √
∑ni=1 |xi |2 ≤ ∑ni=1 (kxk∞ )2 = n(kxk∞ )2 ⇔ kxk2 ≤ n kxk∞ .

1
Also, for k = 1, ..., n, we have |xk | ≤ ∑ni=1 |xi |2 2 = kxk2 =⇒ max1≤k≤n |xk | ≤ kxk2 ⇔
kxk∞ ≤ kxk2 . √
Consequently, kxk∞ ≤ kxk2 ≤ n kxk∞ , which proves the equivalence of kxk2 ; kxk∞ .

Lemma 0.16.3. Let {x1 , ..., xn } be a linearly independent set of vectors in a normed space X
(of any dimension). Then there is a number c > 0 such that for every choice of scalars α1 , ...αn
we have
kα1 x1 + ... + αn xn k ≥ c(|α1 | + ...|αn |), (c > 0). (0.16.1)

Theorem 0.16.4. On a finite dimensional normed space X every two norms k.k1 and k.k2 are
equivalent.

Proof. Let n = dim X, and {e1 , e2 , ..., en } be a basis for X with kei k = 1. Assume that x ∈ X
represented as x = ∑ni=1 αi ei . Then by lemma there exists c > 0 such that kxk1 ≥ c ∑ni=1 |αi |.
Also, using triangle inequality, we have kxk2 ≤ ∑ni=1 |αi | kei k2 ≤ k ∑ni=1 |αi |, where k =
max{kei k2 }.
Combining these inequalities, we get kxk2 ≤ (k/c)c ∑ni=1 |αi | ≤ (k/c) kxk1 = (1/a) kxk1 ,
where a = c/k. Hence a kxk2 ≤ kxk1 .
By interchanging the roles of k.k1 and k.k2 and proceeding as above we get the other
inequality. Hence the result.

Proposition 0.16.5. Show that two norms k.k1 and k.k2 on Rn satisfy √1 kxk ≤ kxk2 ≤ kxk1 .
n 1

Proof. Since kxk21 = (∑ni=1 |ξi |)2 ≥ ∑ni=1 |ξi |2 = kxk1 , thus kxk1 ≥ kxk2 .

1/2
  2 1/2
1 1
n n
Also, √n kxk1 = ∑i=1 √n |ξi | ≤ ∑i=1 |ξi | 2
∑i=1 √1n
n
= kxk2 .
Therefore, √1 kxk ≤ kxk2 ≤ kxk1 .
n 1

47
0.17. COMPACTNESS

0.17 Compactness
Definition 0.17.1. A metric space X is said to be compact if every sequence in X has a con-
vergent subsequence. A subset M of X is said to be compact if M is compact considered as a
subspace of X, that is, if every sequence in M has a convergent subsequence whose limit is an
element of M.

Lemma 0.17.1. A compact subset M of a metric space is closed and bounded.

Proof. First, we have to prove that M is closed, that is M = M. As M ⊂ M, it is enough to

d
show that M ⊂ M. Let x ∈ M so there exists a (xn ) in M such that xn → x. Since M is compact
so every subsequence has a convergent subsequence, that is xnk → x ∈ M. Therefore, M ⊂ M.
Hence M = M, that is M is closed.
To show that M is bounded, supoose M is not bounded so there exists (yn ) in M such that
d(yn , y) > n, where y is some fixed point of M. Therefore (yn ) can not have a convergent
subsequence which contradicts the fact that M is compact. Hence the result.
The converse of this lemma is in general false.
Proof. To prove this important fact, we consider the sequence (en ) in `∞ with kxk∞ = sup xi ,
where en = (δn j ) has the nth term 1 and all other terms 0 (e1 = (1, 0, 0, ..), e2 = (0, 1, 0, ..)...).
This sequence is bounded since ken k = 1. Its terms constitute a point set which is closed
because it has no point of accumulation. But the sequence (en ) has no convergent subsequence,
so point set is not compact.

Theorem 0.17.2 (Bolzano-Weistrass Theorem). Every real, bounded sequence has at least one
convergent subsequence.

However, for a finite dimensional normed space we have

Theorem 0.17.3. In a finite dimensional normed space X, any subset M ⊂ X is compact if and
only if M is closed and bounded.

Proof. By Lemma 0.17.1, Compactness implies implies closedness and boundedness.

Let M be closed and bounded. Suppose that dimX = n and {e1 , e2 , ...en } is a basis for X.
(m) (m)
We consider any sequence (xm ) in M such that xm = ξ1 e1 + ... + ξn en .
Since M is bounded, so is (xm ), say kxm k ≤ k for all m. By Lemma 0.15.1, we have
(m) (m)
k ≥ kxm k = ∑nj=1 ξ j e j ≥ c ∑nj=1 |ξ j |, where c > 0. Hence the sequence of numbers
(m)
(ξ j ) (j fixed) is bounded and by Bolzano-Weierstrass theorem, has a point of accumulation
ξi , 1 ≤ i ≤ n. As in proof of Lemma 0.15.1, we conclude that (xm ) has a subsequence (zm )
which converges to z = ∑ ξ j e j . Since M is closed, z ∈ M. This shows that the arbitrary sequence
(xm ) in M has a subsequence which converges in M. Hence M is compact.
Our discussion shows the following. In Rn (or in any other finite dimensional normed
space) the compact subsets are precisely the closed and bounded subsets, so that this property
(closedness and boundedness) can be used for defining compactness. However, this can no
longer be done in the case of an infinite dimensional normed space.

Lemma 0.17.4. F. Riesz’s Lemma. Let Y and Z be subspaces of a normed space X (of any
dimension), and suppose that Y is closed and is a proper subset of Z. Then for every real
number θ in the interval (0, 1) there is a z ∈ Z such that kzk = 1, kz − yk ≥ θ for all y ∈ Y .

48
0.17. COMPACTNESS

Proof. We consider any v ∈ Z −Y and denote its distance from Y by a, that is a = inf kv − yk
y∈Y
(see figure).
Z X
v Y
y0
a = inf kv − yk
y∈Y

Clearly, a > 0 since Y is closed. We now take any θ ∈ (0, 1). By the definition of an
infimum there is a y0 ∈ Y such that

a ≤ kv − y0 k ≤ a/θ , (0.17.1)

a/θ > a since 0 < θ < 1. Let z = c(v − y0 ), where c = 1/ kv − y0 k. Then kzk = 1, and we
show that that kz − yk ≥ θ for every y ∈ Y . We have

kz − yk = kc(v − y0 ) − yk
= c v − y0 − c−1 y
= c kv − y1 k ,

where y1 = y0 + c−1 y. The form of y1 shows that y1 ∈ Y . Hence kv − y1 k ≥ a, by the definition

of a. Writing c out and using (0.17.1), we obtain
a
kz − yk = c kv − y1 k ≥ ca = a/ kv − y0 k ≥ = θ.
a/θ
Since y ∈ Y is arbitrary, hence the result.
In a finite dimensional normed space the closed unit ball is compact. Conversely, Riesz’s
lemma gives the following useful and remarkable.

Theorem 0.17.5. If a normed space X has the property that the closed unit ball M = {x : kxk ≤
1} is compact, then X is finite dimensional.

Proof. We assume that M is compact but dimX = ∞, and show that this leads to a contradiction.
We choose any x1 of norm 1, that is kx1 k = 1. This x1 generates a one dimensional subspace
X1 = {αx1 } of X, which is closed and is a proper subspace of X since dim X = ∞. By Riesz’s
lemma there is an x2 ∈ X of norm 1 such that kx2 − x1 k ≥ θ = 1/2. In particular, kx3 − x1 k ≥
1/2, kx3 − x2 k ≥ 1/2. Proceeding by induction, we obtain a sequence (xn ) of elements xn ∈
M such that kxm − xn k ≥ 1/2. Obviously, (xn ) can’t have a convergent subsequence. This
contradicts the compactness of M. Hence our assumption dimX = ∞ is false and dimX < ∞.

49
0.18. BANACH FIXED POINT THEOREM AND APPLICATIONS

0.18 Banach Fixed Point Theorem and Applications

Banach fixed point theorem states conditions sufficient for the existence and uniqueness of a
fixed point of a mapping (point that is mapped onto itself remains invariant under a mapping).
The theorem also gives an iterative process by which we can obtain approximations to the fixed
point and error bounds. It finds canonical applications in differential and integral equations.
Definition 0.18.1. Let (X, d) be a metric space and T : (X, d) → (X, d) be a mapping. Then T
is said to be a contraction mapping if there exists a constant 0 < α < 1 such that,

d(T (x), T (y)) ≤ αd(x, y),

for all x, y ∈ X.
Example 0.18.1. Let X = R be the space of reals with usual metric. Let T : X → X be defined
by
(a) T (x) = x/3, for all x ∈ X
(b) T (x) = 2x, for all x ∈ X
Then for x, y ∈ X, (a) is contraction but (b) is not.
(a)
1
|T (x) − T (y)| = |x/3 − y/3| = |x − y|.
3
So, T is a contraction mapping.
(b)
|T (x) − T (y)| = |2x − 2y| = 2|x − y|.
So, T is not a contraction mapping.

Theorem 0.18.2. Every contraction mapping on a metric space is uniformly continuous.

Proof. Let (X, d) be a metric space and let T : (X, d) → (X, d) be a contraction mapping. So
there exists a constant 0 < α < 1 such that

d(T (x), T (y)) ≤ αd(x, y),

for all x, y ∈ X.
Let ε > 0 be arbitrary. Choose δ = αε (> 0), such that d(x, y) < δ . So
ε
d(T (x), T (y)) < α = ε.
α
Therefore, T is uniformly continuous on X.

Theorem 0.18.3. Banach Contraction Principle. Every contraction in a complete metric

space has unique fixed point.

Proof. Let (X, d) be a complete metric space and f : X → X is a contraction mapping. For any
x0 ∈ X, define a sequence (xn ) in X such that

xn = f (xn−1 ) = f n (x0 ) n ≥ 1.

50
0.18. BANACH FIXED POINT THEOREM AND APPLICATIONS

Firstly, we shall show that (xn ) is a Cauchy sequence.

Consider

d(xn+1 , xn ) = d( f (xn ), f (xn−1 ))

≤ α d(xn , xn−1 )
≤ α 2 d(xn−1 , xn−2 ).

In general for any positive integer n, we have

d(xn+1 , xn ) ≤ α n d(x1 , x0 ).

For m ≥ n ≥ 1 , using triangular inequality, we have

d(xn , xm ) ≤ d(xn , xn+1 ) + d(xn+1 , xn+2 ) + ... + d(xm−1 , xm )

≤ α n d(x1 , x0 ) + α n+1 d(x1 , x0 ) + ... + α n+m d(x1 , x0 )
α n (1 − α m−n )
= d(x1 , x0 )
1−α
n m−n
Since 0 ≤ α < 1 ⇒ α (1−α
1−α
)
→ 0 as m, n → ∞ This proves that (xn ) is a Cauchy sequence in
a complete metric space X , ∴ converges to a point x (say) in X. Therefore, lim xn = x. Since
n→∞
f is continuous on X, we have lim f (xn ) = f (x).
n→∞
Further, we’ll show that x is fixed point of f .

lim f (xn ) = lim xn+1 = x.

n→∞ n→∞

By uniqueness of limit, f (x) = x. Hence x is fixed point of f .

Uniqueness:Let f has two fixed points x1 and x2 i.e., f (x1 ) = x1 , f (x2 ) = x2 , x1 6= x2 , where
x1 , x2 ∈ X
then d(x1 , x2 ) = d( f (x1 ), f (x2 )) ≤ α d(x1 , x2 ). Since d(x1 , x2 ) > 0 Therefore dividing by d(x1 , x2 ),
we get α ≥ 1 , which contradicts the definition of f . Therefore f has unique fixed point.
Example 0.18.4. Let X = R be the space of reals with usual metric. Let T : X → X be defined
by T (x) = x/3, for all x ∈ X.
Then for x, y ∈ X,
1
|T (x) − T (y)| = |x/3 − y/3| = |x − y|.
3
So, T is a contraction mapping. Also R is complete. Hence 0 is a unique fixed point of T .

0.18.1 Linear Algebraic Equations

Consider a system of linear algebraic equations
n
yi = ∑ ai j x j + bi, i = 1, ..., n,
j=1

where ai j and bi are real coefficients such that ∑nj=1 |ai j | ≤ α < 1 for each i = 1, ..., n.
Take x = (x1 , x2 , ..., xn ) ∈ Rn and y = (y1 , y2 , ..., yn ) ∈ Rn . Define a metric d on Rn by

d(x, y) = max |xi − yi |.

1≤i≤n

51
0.18. BANACH FIXED POINT THEOREM AND APPLICATIONS

Define T : Rn → Rn by T (x) = y.
(1) (1) (1) (2) (2) (2)
Let x(1) , x(2) ∈ Rn such that x(1) = (x1 , x2 , ..., xn ) and x(2) = (x1 , x2 , ..., xn ).
(1) (1) (1) (1) (1) (1)
Take T (x(1) ) = y(1) and T (x(2) ) = y(2) , where y(1) = (y1 , y2 , ..., yn ) and y(2) = (y1 , y2 , ..., yn ).
(1) (1) (2) (2)
Here yi = ∑nj=1 ai j x j + bi , and yi = ∑nj=1 ai j x j + bi , for i = 1, ..., n. Consider

(1) (2)
d(y(1) , y(2) ) = max |yi − yi |
i
n n
(1) (2)
= max | ∑ ai j x j − ∑ ai j x j |
i j=1 j=1
n
(1) (2)
= max | ∑ ai j (x j − x j )|
i j=1
n
(1) (2)
≤ max ∑ |ai j | |x j − x j |
i j=1
(1) (2)
≤ α max |x j − x j |
j

= αd(x(1) , x(2) ).

Hence d(T (x(1) ), T (x(2) )) ≤ αd(x(1) , x(2) ), that is T is a contraction mapping. Hence by Ba-
nach contraction mapping theorem T has a unique fixed x in Rn .

0.18.2 Existence Theorem on Differential equation (Picard’s theorem)

Consider a differential equation
dy
= f (x, y),
dx
where, f (x, y) be a real valued continuous function of two real variables x and y over a closed
rectangle D of R2 such that, f satisfies Lipschitzian condition with respect to y i.e.

| f (x, y1 ) − f (x, y2 )| ≤ L |y1 − y2 |,

for all (x, y1 ), (x, y2 ) ∈ D where, L > 0 is a Lipschitzian constant.

Let (x0 , y0 ) be an interior point of D then there is a positive real t and unique solution
y = φ (x) in the neighbourhood [x0 − t, x0 + t] of x0 and continuous in [x0 − t, x0 + t] such that
d
y0 = φ (x0 ) and dx φ (x) = f (x, φ (x)) in the neighbourhood [x0 − t, x0 + t] of x0 .
d
Proof. We shall proof the existence of continuous function φ (x) such that dx φ (x) = f (x, φ (x))
in the neighbourhood [x0 − t, x0 + t] of x0 such that y0 = φ (x0 ). It is equivalent to find out the
function Z x
φ (x) = y0 + f (t, φ (t))dt, (t, φ (t)) ∈ D
x0

in the neighbourhood [x0 − t, x0 + t] such that y0 = φ (x0 ). Now consider a closed circular disc
U centred at (x0 , y0 ) with some positive radius such that U is a subset of D.
Now, f is continuous over U. Let m = sup | f (x, y)| < ∞.
(x,y)∈U
Now, choose two positive reals t and δ such that
(i) 0 < t < 1/L
(ii) mt ≤ δ

52
0.18. BANACH FIXED POINT THEOREM AND APPLICATIONS

(iii) [x0 − t, x0 + t], [y0 − δ , y0 + δ ] ⊂ U

Now, consider the class of all real valued continuous functions from [x0 − t, x0 + t] to [y0 −
δ , y0 + δ ] and let this class be denoted by E. i.e.

E = {φ : φ : [x0 − t, x0 + t] → [y0 − δ , y0 + δ ]}.

So E ⊂ C[x0 − t, x0 + t]. Note that, C[x0 − t, x0 + t] is a complete metric space.

But E is considered as a subspace metric space of C[x0 − t, x0 + t]. We shall now prove that
E is complete. It is sufficient to show E is closed i.e. E = E. Obviously , E ⊂ E.
To show, E ⊂ E. Let h ∈ E. So there exists a sequence of continuous functions { fn } ∈ E
such that, lim fn = h. Since this limit is the uniform limit of a sequence of continuous function
{ fn } over [x0 −t, x0 +t] it follows that h is continuous over [x0 −t, x0 +t] i.e. h ∈ C[x0 −t, x0 +t].
We shall now show that the range of h lies in [y0 − δ , y0 + δ ]. If possible let, x0 − t ≤ α ≤
x0 + t such that h(α) ∈ / [y0 − δ , y0 + δ ]. But, lim fn (α) = h(α).
So for sufficiently large values of n, fn (α) ∈ / [y0 − δ , y0 + δ ], a contradiction to the fact
that fn (x) ∈ [y0 − δ , y0 + δ ] for all n and for all x ∈ [x0 − t, x0 + t]. So h(α) ∈ [y0 − δ , y0 + δ ].
Hence h : [x0 − t, x0 + t] → [y0 − δ , y0 + δ ] is a continuous function. i.e. h ∈ E. So E ⊂ E.
Consequently E = E. Hence E is closed and therefore E is complete with respect to sup metric
as defined for C[x0 − t, x0 + t].
Now, let, T : E → E be defined by, T (φ ) = ψ, where φ ∈ E, where
Z x
ψ(x) = y0 + f (t, φ (t))dt, ∀x ∈ [x0 − t, x0 + t].
x0

So ψ is a continuous function over [x0 − t, x0 + t]. Now, consider

Z x
|ψ(x) − y0 | = |y0 + f (t, φ (t))dt − y0 |
x0
Z x
=| f (t, φ (t))dt|
x
Z x0
≤ | f (t, φ (t))|dt
x
Z 0x
≤ sup | f (t, φ (t))|dt
x0
Z x
≤m dt
x0
= m|x − x0 |
≤ mt [as |x − x0 | ≤ t]
≤ δ [by (ii)].

So |ψ(x) − y0 | ≤ δ , for all x ∈ [x0 − t, x0 + t]. Hence y0 − δ ≤ ψ(x) ≤ y0 + δ . So, ψ(x) ∈

[y0 − δ , y0 + δ ], that is, ψ ∈ E. Take φ1 , φ2 ∈ E such that, T (φ1 ) = ψ1 , T (φ2 ) = ψ2 where
ψ1 , ψ2 ∈ E defined by, Z x
ψ1 (x) = y0 + f (t, φ1 (t))dt
x0
and Z x
ψ2 (x) = y0 + f (t, φ2 (t))dt.
x0

53
0.18. BANACH FIXED POINT THEOREM AND APPLICATIONS

So,
Z x Z x
|ψ1 (x) − ψ2 (x)| = | f (t, φ1 (t))dt − f (t, φ2 (t))dt|
x x0
Z 0x
=| ( f (t, φ1 (t)) − f (t, φ2 (t)))dt|
x
Z x0
≤ | f (t, φ1 (t)) − f (t, φ2 (t))|dt
x0
Z x
≤L |φ1 (t) − φ2 (t)|dt
x0
Z x
≤ L d(φ1 , φ2 ) dt
x0
= L|x − x0 |d(φ1 , φ2 )
≤ L t d(φ1 , φ2 ), 0 < Lt < 1.

Hence sup |ψ1 (x) − ψ2 (x)| ≤ L t d(φ1 , φ2 ). Therefore

x∈[x0 −t,x0 +t]

d(ψ1 , ψ2 ) ≤ L t d(φ1 , φ2 )

or it can be rewritten as d(T (φ1 ), T (φ2 )) ≤ L t d(φ1 , φ2 ). So, T is a contraction map. Hence
by Banach Contraction Principle Theorem T has a unique fixed point in E say φ ∈ E. i.e.
T (φ ) = ψ. So φ (x) = y0 + xx0 f (t, φ (t))dt, where y0 = φ (x0 ).
R

54
0.19. INNER PRODUCT AND HILBERT SPACES

0.19 Inner Product and Hilbert Spaces

In a normed space we can add vectors and multiply vectors by scalars, just as in elementary
vector algebra. Furthermore, the norm on such a space generalizes the elementary concept of
the length of a vector. However, what is still missing in a general normed space, and what
we would like to have√if possible, is an analogue of the familiar dot product a.b = a1 b1 +
a2 b2 + a3 b3 and |a| = a.a and condition for orthogonality a.b = 0 which are important tool
in many applications. Hence the question arises whether the dot product and orthogonality can
be generalized to arbitrary vector spaces. In fact, this can be done and leads to inner product
spaces and complete inner product spaces, called Hilbert spaces. Inner product spaces are
special normed spaces.
In the real case, an inner product is a symmetric bilinear form and positive definite on
V × V → R, where V is a vector space. In the complex case, it is a Hermitian symmetric,
conjugate bilinear and positive definite on V ×V → C. A (semi) inner product gives rise to a
(semi) norm. Thus, an inner product space is a particular case of a normed space.
Further, because of the inner product structure, Hilbert spaces have extra features than
Banach spaces, which make them more special and hence lead to several interesting properties.
For example, some ideas for use in differential equations like Fourier Analysis do not fit well
in Banach space setting but fits very well in Hilbert space setting.
Definition 0.19.1. Let V be a linear/vector space over a field F (where F is either C or R).
By an inner product (or pre-Hilbert space) on V , we mean a mapping f := h., .i : V ×V → F,
(u, v) 7→ hu, vi = f (u, v), that assigns for each (u, v) ∈ V × V a value in F, denoted by hu, vi,
called a scalar valued expression (or simply the inner product of u and v), such that for each
u, v, w ∈ V and λ ∈ F we have
(I1) hu, vi = hv, ui [Hermitian/Conjugate symmetric]
(I2) hλ u, vi = λ hu, vi [Homogeneous]
(I3) hu, v + wi = hu, vi + hv, wi [Additivity]
(I4) hu, vi ≥ 0 and hu, ui = 0 =⇒ u = 0. [Positivity]
In (I1), the bar denotes the complex conjugation.
A Hilbert space is a complete inner product space (complete in the metric defined by the
inner product).
There are number of observations that can be made about these axioms. The fact that hu, vi
must be real follows from (I1) on taking v = u in (I1). Further, (I2) for λ = 0 shows that
h0, vi = 0 for every v ∈ V , and in particular, we
p have u = 0 =⇒ hu, vi = O. We shall see in
details how the function k.k defined by kuk = hu, ui makes V into a normed space.
A vector space V together with its inner product h., .i, that is, the pair (V, h., .i), is said to
be an inner product space. An inner product space is also called a pre-Hilbert space. As in the
case of normed spaces, there are really two definitions here and these depend on V over the
real field and V over complex field, respectively. Thus, when the underlying field is R, then V
is called a real inner product space; otherwise it is called a complex inner product space. In the
real case the axioms are the same, except that, (I1) will be written without the bar over hv, ui
since complex conjugation has no effect: hu, vi = hv, ui.
When (I1) is combined with (I3) and (I2), we have hu + v, wi = hu, wi + hv, wi, hu, λ vi =
λ hu, vi.
Thus, the inner product is linear in the first variable, and conjugate linear in the second
variable. As a consequence of this, we easily get that hλ u, µvi = λ µ hu, vi for λ , µ ∈ F and in
the sequel, we shall use this formula often. In fact, we deduce the following general properties

55
0.19. INNER PRODUCT AND HILBERT SPACES

of the inner product which are indeed immediate consequences of the axioms (I1)-(I4) of Def-
inition, and can be verified easily. If OV denotes the zero vector in V and 0 denotes the scalar
zero in F, then, for any vector v ∈ V , we have 0.v = OV so that by the linearity property of the
inner product h0V , vi = h0.v, vi = 0 hv, vi = 0.
Given a vector space V over F, a semi-inner product for V is a function on V × V which
satisfies all the axioms of the inner product except (I4) and instead of (I4), satisfies just the
condition hu, ui > O. Thus, a semi-inner product allows the possibility that for some u 6= 0,
hu, ui = O.

Proposition 0.19.1. In an inner product space (V, h., .i), we have for each u, v, w ∈ V and α, β ∈
F
(i) h0, ui = hu, 0i = 0
(ii) hαu + β v, wi = α hu, wi + β hv, wi [Linearity]
(iii) hu, αv + β wi = α hu, vi + β hu, wi [Conjugate Linearity]

Proof. (i) We have h0, ui = h0.0, ui = 0 h0, ui = 0. Also hu, 0i = h0, ui = 0 = 0.

(ii) hαu + β v, wi = α hu, wi + β hv, wi.
(iii) hu, αv + β wi = hαv + β w, ui = hαv, ui + β hw, ui = hαv, ui+β hw, ui = α hu, vi+β hu, wi.

More generally, for all u j , vk ∈ V and α j , βk ∈ F ( j = 1, 2, ..., n; k = 1, 2, ..., m), we have

* +
n n n n
∑ α j u j , ∑ βk vk = ∑ ∑ α j βk u j , vk
j=1 k=1 j=1 k=1

The bar over α, β in (iii) and also the bar over βk in (0.19) are to be omitted in the case of a
real space as it has no effect.
From Definition and Proposition, we say that the key properties that characterize the inner
product are ”Hermitian symmetric, conjugate bilinear and positive definite”. Clearly, for ev-
ery w ∈ V , hu, wi = hv, wi holds iff hu − v, wi = 0, which for the choice w = u − v ∈ V gives
hu − v, u − vi = 0, that is u = v. The simplest example of an inner product space is C itself,
with hu, vi = uv.
Schwarz’s Inequality:
We may observe that the dot product a.b = a1 b1 + a2 b2 + a3 b3 , where a, b ∈ R3 is inner
product on R3 and the absolute value |a| = (a.a)1/2 , where a ∈ R3 , defines a norm on R3 .
Definition 0.19.2. If V is an inner product space equipped with an inner product h., .i, the
norm k.k on V associated with the inner product is the nonnegative real number defined by the
√
formula kuk = u, u for all u ∈ V . Here, it is evident that for every u ∈ V and α ∈ C, we have
the homogeneity condition kαuk2 = hαu, αui = αα hu, ui = |α|2 kuk2 .
Hence kuk ≥ 0 with equality iff u = 0. This property is called positivity condition of the
norm. p
For z = (z1 , z2 , ..., Zn ) ∈ Cn , kzk = hz, zi = ∑nk=1 |zk |2 which we have already noticed
p

as the Euclidean distance from the origin to the point z ∈ Cn . This observation shows that
the definition of inner product space arise naturally as a generalization from the concept of
(Euclidean) length that is familiar for tne set of real or complex numbers.

p naturally arises is: given an inner product h., .i on a vector space V is the
A question that
map x 7→ kxk := hx, xi, x ∈ V , a norm on V ??
To prove this we use Schwarz’s inequality

56
0.19. INNER PRODUCT AND HILBERT SPACES

Theorem 0.19.2. If h., .i is an inner product on a linear space V and x, y ∈ V , then |hx, yi| ≤
kxk kyk.
Proof. The above result is obvious if y = 0. Hence assume that y 6= 0. Then for every α ∈ F,
we have kx − αyk2 ≥ 0. Therefore,
kx − αyk2 = hx − αy, x − αyi
= kxk2 − hx, αyi − hαy, xi + |α|2 kyk2
= kxk2 − 2Re hx, αyi + |α|2 kyk2 .

By taking α = hx, yi / kyk2 , we have |α|2 kyk2 = |hx, yi|2 / kyk2 and hx, αyi = α hx, yi = |hx, yi|2 / kyk2 ,
so that 0 ≤ kx − αyk2 = kxk2 − |hx, yi|2 / kyk2 . Hence, |hx, yi| ≤ kxk kyk.

p p
If x and y are two vectors in an inner product space, then |hx, yi| = hx, xi hy, yi
p
Theorem 0.19.3. If h., .i is an inner product on a linear space X, then x 7→ kxk := hx, xi, x ∈ X
is a norm on X.
Proof. As h., .i is an inner product on a linear space X, using Schwarz inequality, we get
kx + yk2 = hx + y, x + yi
= hx, xi + hx, yi + hy, xi + hy, yi
= kxk2 + kyk2 + 2Re hx, yi
≤ kxk2 + kyk2 + 2|hx, yi|
≤ kxk2 + kyk2 + 2 kxk kyk
= (kxk + kyk)2 .
Thus k.k satisfies the inequality kx + yk ≤ kxk + kyk for every x, y ∈ X. It is easy to verify the
other conditions.
Hence inner product spaces are normed spaces, and Hilbert spaces are Banach spaces. But not
all normed spaces are inner product spaces.
Theorem 0.19.4. If x and y are two vectors in inner product space, then
(i) kx + yk2 + kx − yk2 = 2 kxk2 + 2 kyk2 [Parallelogram law]
(ii) 4 hx, yi = kx + yk2 − kx − yk2 + i kx + iyk2 − i kx − iyk2 .
[Polarisation Identity]
Proof. (i) x and y are two vectors in inner product space
kx + yk2 = hx + y, x + yi
= hx, xi + hx, yi + hy, xi + hy, yi
= kxk2 + kyk2 + hx, yi + hy, xi .
Also, we have
kx − yk2 = hx − y, x − yi
= hx, xi − hx, yi − hy, xi + hy, yi
= kxk2 + kyk2 − hx, yi − hy, xi .

57
0.19. INNER PRODUCT AND HILBERT SPACES

Adding we get the result.

(ii) On subtracting, we get kx + yk2 − kx − yk2 = 2 hx, yi + 2 hy, xi .
On replacing y by iy, we get kx + iyk2 −kx − iyk2 = 2 hx, iyi+2 hiy, xi = 2i hx, yi+2i hy, xi =
−2i hx, yi + 2i hy, xi. Further, multiplying both sides by i, we get i kx + iyk2 − i kx − iyk2 =
2 hx, yi − 2 hy, xi.
Hence kx + yk2 − kx − yk2 + i kx + iyk2 − i kx − iyk2 = 4 hx, yi.

Theorem 0.19.5. Hilbert space or inner product is jointly continuous, that is, xn → x, yn → y,
=⇒ hxn , yn i → hx, yi.

Proof. Consider

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

= |hxn , yn − yi + hxn − x, yi|
≤ |hxn , yn − yi| + |hxn − x, yi|
≤ kxn k kyn − yk + kxn − xk kyk .

Now, xn → x and yn → y, we have kxn − xk → 0 and kyn − yk → 0. Hence |hxn , yn i−hx, yi| → 0,
or hxn , yn i → hx, yi.
Definition 0.19.3. A complete inner product space is called a Hilbert space i.e. an inner product
space X which is complete with respect to a metric d : X × X → R induced by the inner product
1
h, i on X × X i.e. d(x, y) = hx − y, x − yi 2 for all x, y ∈ X.

Theorem 0.19.6. A Banach space X is a Hilbert space if and only if parallelogram law holds
in it.
Proof. We know that every Hilbert space X is a Banach space where parallelogram law holds
in it.
Conversely suppose that X is a Banach space where parallelogram law holds. Without
loss of generality we can assume a function h, i whose range is R. For all x, y ∈ X, define
h, i : X × X → R by
1h 2 2
i
hx, yi = kx + yk − kx − yk .
4
Here hx, yi = hx, yi as hx, yi is real. Also hx, xi ≥ 0 and hx, xi = 0 iff x = 0.
Now, for u, v, w ∈ X, we have

ku + v + wk2 + ku + v − wk2 = 2(ku + vk2 + kwk2 )

and
ku − v + wk2 + ku − v − wk2 = 2(ku − vk2 + kwk2 ).
On subtracting these two, we get

ku + v + wk2 + ku + v − wk2 − ku − v + wk2 − ku − v − wk2

= 2(ku + vk2 − ku − vk2 )
=⇒ 4[hu + w, vi + hu − w, vi] = 2(4 hu, vi)
hu + w, vi + hu − w, vi] = 2 hu, vi .

58
0.19. INNER PRODUCT AND HILBERT SPACES

Take u = w, we get h2u, vi = 2 hu, vi.

Again, choosing x1 = u + w, x2 = u − w, x3 = v, we get

hu + w, vi + hu − w, vi] = 2 hu, vi = h2u, vi , =⇒ hx1 , x3 i + hx2 , x3 i = hx1 + x2 , x3 i .

Similarly for x1 , x2 , x3 , x4 ∈ X, hx1 + x2 + x3 , x4 i = hx1 + x2 , x4 i + hx3 , x4 i = hx1 , x4 i + hx2 , x4 i +

hx3 , x4 i.
Put x1 = x2 = x3 = x and x4 = y, we get h3x, yi = 3 hx, yi. So by principle of mathematical
induction, we get hnx, yi = n hx, yi for any positive integer n, x, y ∈ X.
Now
1h i
h−x, yi = k−x + yk2 − k−x − yk2
4
1h i
= − kx + yk2 + ky − xk2
4
= − hx, yi .

Take n = −m, (m > 0), we get

hnx, yi = h−mx, yi
= m h−x, yi
= −m hx, yi
n hx, yi .

Hence hλ x, yi = λ hx, yi, where λ is wither positive or negative integer.

Take λ = p/q, where gcd(p,q)=1 and p and q are integers. Consider
 
p
hλ x, yi = x, y
q
 
p
q hλ x, yi = q x, y = hpx, yi = p hx, yi
q
p
=⇒ hλ x, yi = hx, yi
q
 
p p
=⇒ x, y = hx, yi
q q
Let, λ be any real. So there exists a sequence of rationals (rn ) such that rn → λ as n → ∞. So,
hrn x, yi = rn hx, yi → λ hx, yi. Now,

| hrn x, yi − hλ x, yi | = | h(rn − λ )x, yi |

≤ ||(rn − λ )x||||y|| (by C-S inequality )
= |rn − λ |||x||||y|| → 0asn → ∞.

Hence hrn x, yi → hλ x, yi. So, hlambdax, yi = λ hx, yi for all x, y ∈ X. So, X is an inner product
space with respect to h, i consequently, X is a Hilbert space.

Proposition 0.19.7. The Euclidean space Rn is a Hilbert space.

Proof. The space Rn is a Hilbert space with inner product defined by

hx, yi = ξ1 η1 + ... + ξn ηn ,

59
0.19. INNER PRODUCT AND HILBERT SPACES

p q
where x = (ξ1 , ..., ξn ) and y = (η1 , ..., ηn ). Also, kxk = hx, xi = ξ12 + ... + ξn2 and Eu-
p p
clidean metric defined by d(x, y) = kx − yk = hx − y, x − yi = (ξ1 − η1 )2 + ... + (ξn − ηn )2 .
With respect to this metric we can at once see that Rn is complete so as to make Rn , a Hilbert
space.

Proposition 0.19.8. The space `2 is a Hilbert space.

Proof. We define the inner product by hx, yi = ∑∞j=1 ξ j η j . Here, we consider kxk = hx, xi1/2 =
q
∑∞j=1 |ξ j |2 . Here it is easily shown that all inner product axioms are satisfied in `2 . The
p
1
metric d of `2 is defined by d(x, y) = kx − yk = hx − y, x − yi = ∑|ξi − η j | 2 is complete so
as to make `2 a Hilbert space.

Proposition 0.19.9. For 1 ≤ p < ∞, ` p with p 6= 2 is not an inner product space, hence not
Hilbert space.

Proof. A norm on an inner product space satisfies the parallelogram equality kx + yk2 +kx − yk2 =
2 kxk2 + 2 kyk2 . Here ` p with p 6= 2 does not satisfies this, for example, x = (1, 1, 0, 0, ...) ∈ ` p
and y = (1, −1, 0, ...) ∈ ` p . Then kxk = kyk = 21/p and kx + yk = kx − yk = 2. Hence parallel-
ogram equality fails if p 6= 2.
But ` p is complete. Hence ` p with p 6= 2 is a Banach space which is not Hilbert space.

Proposition 0.19.10. C[a, b] is not inner product and hence not Hilbert space.

Proof. Define kxk = supt∈[a,b] |x(t)| does not satisfy the parallelogram equality kx + yk2 +
kx − yk2 = 2 kxk2 + 2 kyk2 . Take x(t) = 1 and y(t) = (t − a)/(b − a), we have kx(t)k = 1
and ky(t)k = 1. x(t) + y(t) = 1 + (t − a)/(b − a) and x(t) − y(t) = 1 − (t − a)/(b − a). Hence
kx + yk = 2, kx − yk = 1 and kx + yk2 + kx − yk2 = 5. But 2 kxk2 + 2 kyk2 = 4. Hence the
result.

60
0.19. INNER PRODUCT AND HILBERT SPACES

Definition 0.19.4. Isomorphism. An isomorphism T if an IPS (X1 , hi1 ) onto an IPS (X2 , hi2 )
over the same field K is a bijective linear operator T : X1 → X2 which preserves the inner
product, that is, hT x, Tyi = hx, yi.
Definition 0.19.5. A subspace Y of an inner product space X is defined to be a vector subspace
of X taken with the inner product on X restricted to Y ×Y .
Similarly, a subspace Y of a Hilbert space H is defined to be a subspace of H, regarded as an
inner product space. Note that Y need not be a Hilbert space because Y may not be complete.
Note. Let Y be a subspace of a Hilbert space H. Then:
(a) Y is complete if and only if Y is closed in H.
(b) If Y is finite dimensional, then Y is complete.
(c) If H is separable, so is Y . More generally, every subset of a separable inner product
space is separable. [It has countable subset which is dense in itself]
Definition 0.19.6. An element x of an inner product space X is said to be orthogonal to an
element y ∈ X if hx, yi = O. We also say that x and y are orthogonal, and we write x ⊥ y.
Law of cosines in R2 is ku − vk2 = kuk2 + kvk2 − 2 kuk kvk cos θ , where u, v ∈ R2 and θ is
angle between u and v. hu − v, u − vi = hu, ui + hv, vi − 2 kuk kvk cos θ . On simplifying, we get
−2 hu, vi = −2 kuk kvk cos θ or cos θ = hu, vi /(kuk kvk).
If u ⊥ v, then cos θ = 0, implies that hu, vi = 0. Conversely, if hu, vi = 0, we have cos θ = 0,
θ = π/2, u ⊥ v.
Example 0.19.11. (1) The space Rn is a Hilbert space with inner product definedq by hx, yi =
p
ξ1 η1 +...+ξn ηn , where x = (ξ1 , ..., ξn ) and y = (η1 , ..., ηn ). Also, kxk = hx, xi = ξ12 + ... + ξn2
p p
and Euclidean metric defined by d(x, y) = kx − yk = hx − y, x − yi = (ξ1 − η1 )2 + ... + (ξn − ηn )2 .
It is complete.
In trigonometry, using law of cosines, we have ka − bk2 = kak2 + kbk2 − 2 kak kbk cos θ ,
where θ is acute angle between a and b
”Is it possible to generalize the concept of angle between vectors in an inner product space
”
ku − vk2 = kuk2 + kvk2 − 2Re hu, vi. By Schwarz inequality, |Re hu, vi| ≤ |hu, vi| ≤ kuk kvk,
so that −1 ≤ Rehu,vi Rehu,vi 2 2
kukkvk ≤ 1. Therefore, if we define cos φ = kukkvk , then ku − vk = kuk + kvk −
2

2 kuk kvk cos φ .

Also, hx, yi = x.y = 0, agrees with elementary concept of orthoganility.
(2) `2 is a Hilbert space with inner product defined by hx, yi = ∑∞j=1 ξ j η j . Here, we consider
q
1/2
kxk = hx, xi = ∑∞j=1 |xi j |2 .
(3) ` p with p 6= 2 is not an inner product space, hence not Hilbert space. A norm on an inner
product space satisfies the parallelogram equality kx + yk2 + kx − yk2 = 2 kxk2 + 2 kyk2 . Here
` p with p 6= 2 does not satisfies this, for example, x = (1, 1, 0, 0, ...) ∈ ` p and y = (1, −1, 0, ...) ∈
` p . Then kxk = kyk = 21/p and kx + yk = kx − yk = 2. Hence parallelogram equality fails if
p 6= 2.
But ` p is complete. Hence ` p with p 6= 2 is a Banach space which is not Hilbert space.
(4) C[a, b] is not inner product and hence not Hilbert space.
Define kxk = maxt∈[a,b] |x(t)| does not satisfy the parallelogram equality kx + yk2 +kx − yk2 =
2 kxk2 + 2 kyk2 . Take x(t) = 1 and y(t) = (t − a)/(b − a), we have kx(t)k = 1 and ky(t)k = 1.
x(t) + y(t) = 1 + (t − a)/(b − a) and x(t) − y(t) = 1 − (t − a)/(b − a). Hence kx + yk = 2,
kx − yk = 1 and kx + yk2 + kx − yk2 = 5. But 2 kxk2 + 2 kyk2 = 4. Hence the result.

61
0.20. ORTHOGONAL AND ORTHONORMAL VECTORS

0.20 Orthogonal and Orthonormal vectors

Definition 0.20.1. An element x of an inner product space H is said to be orthogonal to an
element y ∈ H if hx, yi = O. We also say that x and y are orthogonal, and we write x ⊥ y.
Definition 0.20.2. (Orthogonal Complement). Let Y be a closed subspace of X. Then the
orthogonal complement of Y in X is Y ⊥ = {z ∈ X|z ⊥ Y }, which is the set of all vectors
orthogonal to Y.
Let M be a non empty subset of a Hilbert space H. An element x ∈ H is said to be orthogonal
to M if hx, yi = 0 for all y ∈ M and in symbol we write x ⊥ M.
Then the set of all elements of H which are orthogonal to M is called the orthogonal com-
plement of M and denoted by M ⊥ . Similarly we can interpret M ⊥⊥ , M ⊥⊥⊥ , ... .
Two sets M1 and M2 are said to be orthogonal to each other written as M1 ⊥ M2 whenever
hx, yi = 0 for all x ∈ M1 and for all y ∈ M2 .
From definition, it is easy to see that
(i) {θ }⊥ = H, H ⊥ = {θ }.
(ii) if M be a subspace of H then M ∩ M ⊥ = {θ }.

Theorem 0.20.1. (Pythagorean theorem) If x, y ∈ H and x ⊥ y then

(i) ||x + y||2 = ||x||2 + ||y||2 and
(ii) ||x − y||2 = ||x||2 + ||y||2 .

Proof. We have, hx, yi = 0. So

(i)||x + y||2 = hx + y, x + yi = hx, xi + hx, yi + hy, xi + hy, yi
= hx, xi + hy, yi
= ||x||2 + ||y||2
and

(ii)||x − y||2 = hx − y, x − yi = hx, xi − hx, yi − hy, xi + hy, yi

= hx, xi + hy, yi
= ||x||2 + ||y||2 .

Theorem 0.20.2. Let M be a subset of a Hilbert space H. Then

(i) if M1 ⊂ M2 ⊂ H then M1⊥ ⊃ M2⊥ .
(ii) M ⊂ M ⊥⊥ .
(iii) (M ⊥⊥ )⊥ = M ⊥ .
(iv) if M ⊂ H, then M ⊥ is a closed subspace of H.

Proof. (i) Let x ∈ M2⊥ . So hx, yi = 0, for all y ∈ M2 . So hy, xi = 0 for all y ∈ M2 . Therefore,
hy, xi = 0, for all y ∈ M1 as M1 ⊂ M2 . So x ∈ M1⊥ . Hence M1⊥ ⊃ M2⊥ .
(ii) Let x ∈ M. Then hy, xi = 0 for all y ∈ M ⊥ . So x ∈ (M ⊥ )⊥ . Thus M ⊂ M ⊥⊥ .
(iii) By (ii) M ⊂ M ⊥⊥ . So by (i) (M ⊥⊥ )⊥ ⊂ M ⊥ . Again by (ii), M ⊥ ⊂ (M ⊥ )⊥⊥ ⊂ (M ⊥⊥ )⊥ .
Hence (M ⊥⊥ )⊥ = M ⊥ .
(iv) Let x, y ∈ M ⊥ and α be any complex scalar. Therefore x ⊥ M, y ⊥ M, (x − y) ⊥ M and
αx ⊥ M and so x − y, αx ∈ M ⊥ . Consequently M ⊥ is a linear subspace of H.

62
0.20. ORTHOGONAL AND ORTHONORMAL VECTORS

Let x ∈ M ⊥ . So there exists a sequence (xk ) in M such that limk xk = x with xk ⊥ M. By the
continuity of inner product we see that x ⊥ M. Hence M ⊥ is closed.

Theorem 0.20.3. If M is a closed subspace of a Hilbert space H then M = M ⊥⊥ .

Proof. As M ⊂ (M ⊥ )⊥ = M ⊥⊥ . Now we have to shaow that M ⊃ (M ⊥ )⊥ = M ⊥⊥ .

Let x ∈ M ⊥⊥ . Then x = y + z, where y ∈ M ⊂ M ⊥⊥ . Since M ⊥⊥ is a vector space, z =
x − y ∈ M ⊥⊥ . Hence z ⊥ M ⊥ . But z ∈ M ⊥ . z ⊥ z, hence z = 0 so that x = y, that is x ∈ Y . Since
x ∈ M ⊥⊥ was arbitrary so M ⊃ M ⊥⊥ . Hence the result.

Theorem 0.20.4. (Dense set). For any subset M 6= 0/ of a Hilbert space H, the span of M is
dense in H if and only if M ⊥ = {0}.

Proof. Let x ∈ M ⊥ . Assume that V = spanM is dense in H. Then x ∈ V = H, so there exists

a sequence (xn ) ∈ V such that xn → x. Since x ∈ M ⊥ and M ⊥ ⊥ V , we have hxn , xi = 0. The
continuity of the inner product h., .i implies that hxn , xi → hx, xi = kxk2 = 0, so x = 0. Since
x ∈ M ⊥ is arbitrary, we have M ⊥ = {0}.
Conversely, suppose that M ⊥ = {0}. If x ⊥ V , then x ⊥ M so that x ∈ M ⊥ , so x = 0. Hence
V ⊥ = {0} and so V = H. Hence the result.
Definition 0.20.3. (a) A non empty {ei } of a Hilbert space H is said to be an orthonormal
system if
(i) ei , e j = 0 for i 6= j
and
(ii) ||ei || = 1 for every i, that is any vector of the system is non zero unit vector in H.
(b) If an orthonormal system of H is countable then we can enumerate its elements in a
sequence and then we say it is an orthonormal sequence.
Example 0.20.5. 1) Euclidean space R3 . In the space R3 , the three unit vectors (1, 0, 0), (0, 1, 0), (0, 0, 1)
in the direction of the three axes of a rectangular coordinate system form an orthonormal set.
2) In Rn or Cn , the orthonormal set of vectors are of the type
(0, . . . , 0, 1, 0, . . . , 0),
↑
(k)-th
and denoted by ek . Clearly ||ek || = 1 and ei , e j = 0 for i 6= j.
3) In the space `2 , an orthonormal sequence is (en ), where en = (δn j ) has the nth element 1
and all others zero.
4) Let X be the inner product space of all real-valued continuous functions on [0, 2π] with
inner product defined by hx, yi = 02π x(t)y(t) dt.
R

An orthogonal sequence in X is (un ), where un (t) = cos nt, n = 0, 1, ... Another orthogonal
sequence in X is (vn ) , where vn (t) = sin nt,
n = 1, 2, ....
R 2π  0, m 6= n

Now hum , un i = 0 cos mt cos nt dt = π, m = n = 1, 2, ...

 2π, m = 0 = n
√
Hence an orthonormal sequence is (en ), where e0 (t) = 1/ 2π, en (t) = ukun (t) nk
= cos√ nt , n =
π
1, 2, ...
Similarly for vn , we can obtain an orthonormal sequence eˆn as eˆn = vkvn (t)
nk
= sin
√ nt , n = 1, 2, ...
π

63
0.20. ORTHOGONAL AND ORTHONORMAL VECTORS

Theorem 0.20.6. An orthonormal system in a Hilbert space H is linearly independent.

Proof. Let {en }, n = 1, 2, . . . be an orthonormal system in H and let for a finite subset {e1 , e2 , . . . , en }
of the system we have α1 e1 +α2 e2 +· · ·+αn en = θ , where αi ’s are scalars. Then for 1 ≤ j ≤ n,

0 = θ,ej = ∑ αiei, e j = ∑ αi ei , e j = α j e j , e j = α j ,

other terms being zero because of mutual orthogonality. So α1 = α2 = · · · = αn = 0. This

means that subsystem of the given system is linearly independent. Hence the system {en },
n = 1, 2, . . . is linearly independent.
Our goal is a representation of a Hilbert space as a direct sum which is particularly simple
and suitable because it makes use of orthogonality. To understand the situation and the prob-
lem, let us first introduce the concept of a direct sum. This concept makes sense for any vector
space and is defined as follows.
Definition 0.20.4. (Direct sum). A vector space X is said to be the direct sum of two subspaces
Y and Z of X, written X = Y Z, if each x ∈ X has a unique representation x = y + z, y ∈ Y, z ∈
L

Z. Then Z is called an algebraic complement of Y in X and vice versa, and Y , Z is called a

complementary pair of subspaces in X. L
From the elementary property of linear algebra we have if X = Y + Z, then X = Y Z if
and only Y ∩ Z = {θ }.
For example, Y = R is a subspace of the Euclidean plane R2 . Clearly, Y has infinitely many
algebraic complements in R2 , each of which is a real line. But most convenient is a complement
that is perpendicular. We make use of this fact when we choose a Cartesian coordinate system.

Theorem 0.20.7. If M and N are two closed subspace of a Hilbert space H with M ⊥ N then
M + N is closed in H.

Proof. As H is complete and M and N are closed so M and N are complete. Since M ⊥ N, so
M ∩ N = {θ }. So H = M N.
L

Let z ∈ M + N. So there exists a sequence (zk ) where zk ∈ M + N such that zk → z as k → ∞.

Let zk = xk + yk , where xk ∈ M, yk ∈ N, k = 1, 2, . . . . Hence xn − xm ∈ M and yn − ym ∈ N. So
hxn − xm , yn − ym i = 0.
Hence by Pythagorean Theorem,

||(xn − xm ) + (yn − ym )||2 = ||xn − xm ||2 + ||yn − ym ||2 .

So ||zn − zm ||2 = ||xn − xm ||2 + ||yn − ym ||2 . Since zk → z, {zk } is a Cauchy sequence in H. So
||zn −zm || → 0 as n, m → ∞. So ||xn −xm ||2 +||yn −ym ||2 → 0. Hence {xn } and {yn } are Cauchy
sequences in M and N respectively and xn → x ∈ M and yn → y ∈ N (by completeness of M
and N). Since zn = (xn + yn ) → (x + y) ∈ M + N as and so z = x + y ∈ M + N. Consequently
M + N is closed in H.

64
0.21. BEST APPROXIMATION IN HILBERT SPACE

0.21 Best Approximation in Hilbert Space

In a metric space X, the distance δ from an element x ∈ X to a nonempty subset M ⊂ X is
defined to be δ = inf d(x, y).
y∈M
In normed space, this becomes δ = inf kx − yk.
y∈M
It is important to know whether there is a y ∈ M such that δ = kx − yk. that is, intuitively
speaking, a point y ∈ M which is closest to the given x, and if such an element exists, whether
it is unique. This is an existence and uniqueness problem. It is of fundamental importance,
theoretically as well as in applications, for instance, in connection with approximations of
functions.
The segment joining two given elements x and y of a vector space X is defined to be the set
of all z ∈ X of the form z = αx + (1 − α)y, α ∈ R, 0 ≤ α ≤ 1.
A subset M of X is said to be convex if for every x, y ∈ M the segment joining x and y is
contained in M, that is, αx + (1 − α)y ∈ M for all α ∈ [0, 1], x, y ∈ M.
Theorem 0.21.1. (Minimizing vector). Let X be an inner product space and M 6= 0/ a convex
subset which is complete (in the metric induced by the inner product). Then for every given
x ∈ X there exists a unique y ∈ M such that

δ = inf kx − yk = kx − yk .
y∈M

Proof. First we have to show the Existence of y ∈ M.

By the definition of an infimum there is a sequence (yn ) in M such that δn → δ where
δn = kx − yn k.
We show that (yn ) is Cauchy. Writing yn − x = vn we have kvn k = δn and
1
kvn + vm k = kyn + ym − 2xk = 2 (yn + ym ) − x ≥ 2δ ,
2
because M is convex, so that 12 (yn + ym ) ∈ M. Furthermore, we have yn − ym = vn − vm . Hence
by parallelogram equality, we get
kyn − ym k2 = kvn − vm k2 = − kvn + vm k2 + 2(kvn k2 + kvm k2 )
≤ −(2δ )2 + 2(δn2 + δm2 ).
Taking (n, m) → ∞, we get (yn ) is a Cauchy sequence. Since M is complete, yn → y ∈ M. So
kx − yk ≥ δ . Also, we have
kx − yk ≤ kx − yn k + kyn − yk
= δn + kyn − yk
→ δ.
This shows that kx − yk = δ .
Now, we prove the uniqueness of y ∈ M. We assume that y ∈ M and y0 ∈ M both satisfy
kx − yk = δ and kx − y0 k = δ . By the parallelogram equality, we have

ky − y0 k2 = ky − x − (y0 − x)k2
= 2 ky − xk2 + 2 ky0 − xk2 − k(y − x) + (y0 − x)k2
2
1
= 2δ 2 + 2δ 2 − 22 (y + y0 ) − x .
2

65
0.21. BEST APPROXIMATION IN HILBERT SPACE

Here ky − y0 k2 ≤ 0 since 12 (y + y0 ) ∈ M and 12 (y + y0 ) − x ≥ δ . But ky − y0 k2 ≥ 0. There-

fore, ky − y0 k = 0 and y = y0 . Hence the result.

Turning from arbitrary convex sets to subspaces, we obtain a following result which gener-
alizes the familiar idea of elementary geometry that the unique point y in a given subspace Y
closest to a given x is found by ”dropping a perpendicular from x to Y .”
Projection Theorem.
Lemma 0.21.2. (Orthogonality). In Theorem 0.21.1, let M be a complete subspace Y and x ∈ X
fixed. Then z = x − y is orthogonal to Y , that is, z ⊥ Y .
Proof. Suppose that z = x − y is not orthogonal to Y , there would be a y1 ∈ Y such that

hz, y1 i = β 6= 0.

Cleraly, y1 6= 0 since otherwise hz, y1 i = O. Furthermore, for any scalar α,

kz − αy1 k2 = hz − αy1 , z − αy1 i

= hz, zi − α hz, y1 i − α[hy1 , zi − α hy1 , y1 i]
= kzk2 − αβ − α[β − α hy1 , y1 i].

β
Further, if we choose β − α hy1 , y1 i = 0, that is α = hy1 ,y1 i , we have

β |β |2
kz − αy1 k2 = kzk2 − β − α[0] = δ 2 − < δ 2.
hy1 , y1 i hy1 , y1 i
But this is impossible because we have z − αy1 = x − y2 where y2 = y + αy1 ∈ Y so that
kz − αy1 k ≥ δ . Therefore, we must have z ⊥ Y .

66
0.21. BEST APPROXIMATION IN HILBERT SPACE

Our goal is a representation of a Hilbert space as a direct sum which is particularly simple
and suitable because it makes use of orthogonality. To understand the situation and the prob-
lem, let us first introduce the concept of a direct sum. This concept makes sense for any vector
space and is defined as follows.
Definition 0.21.1. (Direct sum). A vector space X is said to be the direct sum of two subspaces
Y and Z of X, written X = Y Z, if each x ∈ X has a unique representation x = y + z, y ∈ Y, z ∈
L

Z. Then Z is called an algebraic complement of Y in X and vice versa, and Y , Z is called a

complementary pair of subspaces in X.
For example, Y = R is a subspace of the Euclidean plane R2 . Clearly, Y has infinitely many
algebraic complements in R2 , each of which is a real line. But most convenient is a complement
that is perpendicular. We make use of this fact when we choose a Cartesian coordinate system.
Definition 0.21.2. (Orthogonal Complement). Let Y be a closed subspace of X. Then the
orthogonal complement of Y in X is Y ⊥ = {z ∈ X|z ⊥ Y }, which is the set of all vectors
orthogonal to Y.
L ⊥
Theorem 0.21.3. Let Y be a closed subspace of H. Then H = Y Y .

Proof. H is complete and Y is closed implies Y is complete. Further, we know Y is convex.

Therefore for all x ∈ H there exists y ∈ Y such that x = y + z, where z ∈ Y ⊥ .
Now to prove uniqueness, we assume that x = y + z = y1 + z1 , where y, y1 ∈ Y and z, z1 ∈ Z.
Then y − y1 = z1 − z. Since y − y1 ∈ Y , whereas z1 − z ∈ Z = Y ⊥ , we see that y − y1 ∈ Y ∩Y ⊥ =
{0}. This implies y = y1 and z = z1 .
Definition 0.21.3. (Orthogonal Projection). Let Y be a closed subspace of H. So H = Y + Y ⊥
L

and for all x ∈ H there exists y ∈ Y such that x = y + z, where z ∈ Y ⊥ . Then the orthogonal
projection onto Y is P : H → Y given by x 7→ y = Px. P is called the (orthogonal) projection
(or projection operator) of H onto Y .
Clearly P is bounded. Further P is idempotent, that is, P2 = P, which we call idempotent.
Definition 0.21.4 (Annihilator). An orthogonal complement is a special annihilator, where, by
definition, the annihilator M ⊥ of a set M 6= 0/ in an inner product space X is the set M ⊥ = {x ∈
X : x ⊥ M}.
Thus, x ∈ M ⊥ if and only if hx, vi = 0 for all v ∈ M.
Remark 0.21.1. M ⊥ is a vector space, since x, y ∈ M ⊥ implies for all v ∈ M and scalars α, β ,
we have hαx + β y, vi = α hx, vi + β hy, vi = 0, hence αx + β y ∈ M ⊥ .
M ⊥ is closed.
M ⊂ (M ⊥ )⊥ = M ⊥⊥ , because x ∈ M =⇒ x ⊥ M ⊥ =⇒ x ∈ M ⊥⊥ .

Lemma 0.21.4. If M is a closed subspace of a Hilbert space H then M = M ⊥⊥ .

Proof. As M ⊂ (M ⊥ )⊥ = M ⊥⊥ . Now we have to shaow that M ⊃ (M ⊥ )⊥ = M ⊥⊥ .

Let x ∈ M ⊥⊥ . Then x = y + z, where y ∈ M ⊂ M ⊥⊥ . Since M ⊥⊥ is a vector space, z =
x − y ∈ M ⊥⊥ . Hence z ⊥ M ⊥ . But z ∈ M ⊥ . z ⊥ z, hence z = 0 so that x = y, that is x ∈ Y . Since
x ∈ M ⊥⊥ was arbitrary so M ⊃ M ⊥⊥ . Hence the result.

Lemma 0.21.5. (Dense set). For any subset M 6= 0/ of a Hilbert space H, the span of M is dense
in H if and only if M ⊥ = {0}.

67
0.21. BEST APPROXIMATION IN HILBERT SPACE

Proof. Let x ∈ M ⊥ . Assume that V = spanM is dense in H. Then x ∈ V = H, so there exists

a sequence (xn ) ∈ V such that xn → x. Since x ∈ M ⊥ and M ⊥ ⊥ V , we have hxn , xi = 0. The
continuity of the inner product h., .i implies that hxn , xi → hx, xi = kxk2 = 0, so x = 0. Since
x ∈ M ⊥ is arbitrary, we have M ⊥ = {0}.
Conversely, suppose that M ⊥ = {0}. If x ⊥ V , then x ⊥ M so that x ∈ M ⊥ , so x = 0. Hence
⊥
V = {0} and so V = H. Hence the result.
Definition 0.21.5. (Orthonormal Set/Sequence). M is said to be an orthogonal set if for all
x, y ∈ M, x 6= y =⇒ hx, yi = 0. M is said to be orthonormal if for all x, y ∈ M,
(
0, x 6= y
hx, yi = δxy =
1, x = y.

If M is countable and orthogonal (respectively, orthonormal) then we can write M = (xn )

and we say (xn ) is an orthogonal (respectively, orthonormal) sequence.
More generally, an indexed set, or family, (xα ) , α ∈ I, is called orthogonal if xα ⊥ xβ for
all α, β ∈ I with α 6= β . The family is called orthonormal(if it is orthogonal and all xα have
0, α 6= β
norm 1, so that for all α, β ∈ I we have xα , xβ = δαβ =
1, α = β .

δαβ is the Kronecker delta.

A set S of vectors in IPS V is called an orthogonal set if every pair of distinct vectors on S
are othogonal. Further an orthogonal set S is called orthonormal if kvk = 1, for all v ∈ S.

Lemma 0.21.6. (Pythagorean Relation, Linear Independence). Let M = {x1 , ..., xn } be an or-
thogonal set. Then
(i) k∑ni=1 xi k2 = ∑ni=1 kxi k2 .
(ii) M is linearly independent.

Proof. (i) Consider

2
n
∑ xi = hx1 + ... + xn , x1 + ...xn i
i=1
= hx1 , x1 i + ... hxn , xn i
= kx1 k2 + ... + kxn k2
n
= ∑ kxi k2 .
i=1

(ii) Let {e1 , ..., en } be orthogonal and consider α1 e1 + ... + αn en = 0. Multiplication by

fixed e j , we have ∑ni=1 αi ei , e j = ∑ni=1 αi e j , e j = α j e j , e j = α j = 0.
Converse is not true, that is not every linearly independent set is orthogonal.
Example 0.21.7. 1) Euclidean space R3 . In the space R3 , the three unit vectors (1, 0, 0), (0, 1, 0), (0, 0, 1)
in the direction of the three axes of a rectangular coordinate system form an orthonormal set.
2) In the space `2 , an orthonormal sequence is (en ), where en = (δn j ) has the nth element 1
and all others zero.
3) Let X be the inner product space of all real-valued continuous functions on [0, 2π] with
inner product defined by hx, yi = 02π x(t)y(t) dt.
R

68
0.21. BEST APPROXIMATION IN HILBERT SPACE

An orthogonal sequence in X is (un ), where un (t) = cos nt, n = 0, 1, ... Another orthogonal
sequence in X is (vn ) , where vn (t) = sin nt,
n = 1, 2, ....
R 2π  0, m 6= n

Now hum , un i = 0 cos mt cos nt dt = π, m = n = 1, 2, ...

 2π, m = 0 = n
√
Hence an orthonormal sequence is (en ), where e0 (t) = 1/ 2π, en (t) = ukun (t) nk
= cos√ nt , n =
π
1, 2, ...
Similarly for vn , we can obtain an orthonormal sequence eˆn as eˆn = vkvn (t)
nk
= sin
√ nt , n = 1, 2, ...
π
Note that we even have um ⊥ vn for all m, n. These sequences appear in Fourier series.

69
0.22. SOME RESULTS IN INNER PRODUCT SPACES

0.22 Some results in Inner Product Spaces

Theorem 0.22.1. Let X be an inner product space and (ek ) be an orthonormal sequence in X.
Then
(i) the function f : Cn → R defined by

f (α1 , α2 , . . . , αn ) = ||x − (α1 x1 + α2 x2 + · · · + αn xn )||,

where x ∈ X, (α1 , α2 , . . . , αn ) ∈ Cn , attains its absolute minimum value at one and only one
point (α1 , α2 , . . . , αn ) ∈ Cn , where hx, xk i = αk , k = 1, 2, . . . , n.
(ii) | hx, x1 i |2 + | hx, x2 i |2 + · · · + | hx, xn i |2 ≤ ||x||2 , x ∈ X.

Proof. (i) Let X be an inner product space and (ek ) be an orthonormal sequence in X. Consider

f (α1 , α2 , . . . , αn ) = ||x − ∑ αi xi ||.

On squaring both sides, we have

[ f (α1 , α2 , . . . , αn )]2 = ||x − ∑ αi xi ||2 .

Further,

[ f (α1 , α2 , . . . , αn )]2
= ||x − ∑ αi xi ||2
= x − ∑ αi xi , x − ∑ αi xi
= hx, xi − x, ∑ αi xi − ∑ αixi, x + ∑ αixi, ∑ α j x j
= kxk2 − ∑ αi hx, xi i − ∑ αi hxi , xi + ∑ ∑ αi α j xi , x j
i j

= kxk2 − ∑ αi hx, xi i − ∑ αi hxi , xi + ∑ |αi |2

i
2
= kxk − ∑ | hx, xi i | + ∑ | hx, xi i | − ∑ αi hx, xi i − ∑ αi hxi , xi + ∑ |αi |2
2 2
i
2 2 2
= kxk − ∑ | hx, xi i | + ∑ | hx, xi i − αi | .

But [ f (α1 , α2 , . . . , αn )]2 ≥ 0. So [ f (α1 , α2 , . . . , αn )]2 ≥ kxk2 − ∑ | hx, xi i |2 .

Thus [ f (α1 , α2 , . . . , αn )]2 attains minimum value kxk2 − ∑ | hx, xi i |2 when hx, xi i = αi = 0,
i = 1, 2, ..., n
If possible let f attains the minimum value at (β1 , β2 , . . . , βn ) ∈ Cn where hx, xi i = βi , i =
1, 2, . . . , n. It is to verify that αi = βi , i = 1, 2, . . . , n. So (α1 , α2 , . . . , αn ) ∈ Cn is unique by which
f attains minimum value at (α1 , α2 , . . . , αn ) ∈ Cn .
(ii) The sums ∑ | hx, xi i |2 have nonnegative terms, so that they form a monotone increasing
sequence. These sequence converges because it is bounded by ||x||2 . This is the sequence of
the partial sums of an infinite series, which thus converges. Therefore ∑ | hx, xi i |2 ≤ kxk2 . This
inequality is known as Bessel’s Inequality.
If ∑ | hx, xi i |2 = kxk2 , then the equality is called Parsival’s Equality.

70
0.22. SOME RESULTS IN INNER PRODUCT SPACES

Theorem 0.22.2. (Bessel’s Inequality). Let X be an inner product space and let (ek ) be an
orthonormal sequence in X. Let x ∈ H then we have
∞
∑ |hx, ek i|2 ≤ kxk2 .
k=1

Proof. Let X be an inner product space and let (ek ) be an orthonormal sequence in X, and
suppose x ∈ span({e1 , ..., en }) where n is fixed. Then we can represent x = ∑nk=1 αk ek as
* +
n
hx, ei i = ∑ αk ek , ei = αi
k=1
n
=⇒ x = ∑ hx, ek i ek .
k=1

Now let x ∈ X be arbitrary and take y ∈ span({e1 , ..., en }), where y = ∑nk=1 hx, ek i ek .
Define z by setting x = y + z =⇒ z ⊥ y because
hz, yi = hx − y, yi
= hx, yi − hy, yi
* +
n
= x, ∑ hx, ek i ek − kyk2
k=1
= ∑ hx, hx, ek i ek i − kyk2
= ∑ hx, ek i hx, ek i − ∑|hx, ek i|2 = 0.

Then, we have kxk2 = kyk2 + kzk2 or kzk2 = kxk2 − ∑|hx, ek i|2 ≥ 0. Therefore, ∑|hx, ek i|2 ≤
kxk2 . Hence the result.
Definition 0.22.1. (Fourier Coefficients). The sequence (hx, ek i) is called the Fourier coeffi-
cients of x with respect to (ek ).

Proposition 0.22.3. (Minimum property of Fourier Coefficients). Let {e1 , ..., en } be an or-
thonormal set in an in an inner product space X (n is fixed). Let x ∈ X be an arbitrary, fixed
element and let y = ∑nk=1 βk ek . Then kx − yk depends on β1 , ..., βn . Show by direct calculation
that kx − yk is minimized iff βi = hx, ei i, for all i = 1, ..., n.

Proof. Let γi = hx, ei i and y = ∑ βi ei . Then

kx − yk2 = x − ∑ βi ei , x − ∑ βi ei
= kxk2 − ∑ βi γi − ∑ βi γi + ∑|βi |2 .
This is minimum for given x and ei ’s iff βi = γi .
Fourier Series of periodic function x(t + 2π) = x(t) is a trigonometric series is a series of
the form
∞
a0 + ∑k=1 (ak cos kt + bk sin kt),
with coefficients ak and bk given by formulas called Eulers formulas
Z 2π Z 2π Z 2π
1 1 1
a0 = x(t)dt, ak = x(t) cos kt dt, bk = x(t) sin kt dt.
2π 0 π 0 π 0

71
0.22. SOME RESULTS IN INNER PRODUCT SPACES

These coefficients are called the Fourier coefficients of x. Obviously, the cosine and sine
functions in Fourier series are those of the sequences (uk ) and (vk ), where uk (t) = cos kt and
vk (t) = sin kt. Hence, we write the Fourier series as
∞
x(t) = a0 u0 (t) + ∑ (ak uk (t) + bk vk (t)).
k=1

We assume that termwise integration is permissible (uniform convergence would suffice) and
use the orthogonality of (uk ) and (vk ) as well as the fact that u j ⊥ vk for all j, k. Then we obtain
x, u j = a0 u0 , u j + ∑[ak uk , u j + bk vk , u j ]
(
2 2πa0 , j = 0
= a j u j, u j = a j u j =
πa j , j = 1, 2, ...

and
2
x, v j = b j v j = πb j , j = 1, 2, ...
−1
Solving for a j and b j , and using the orthonormal sequences (e j ) and (e j ), where e j = u j uj
−1
and e0j = v j v j , we obtain
1 1
aj = 2
x, u j = x, e j
uj uj

1 1
bj = 2
x, v j = x, e0j .
vj vj
Hence
∞
x(t) = hx, e0 i e0 + ∑ (hx, ek i ek + x, e0k e0k ).
k=1

For any orthonormal sequence (ek ) in a Hilbert space H, we may consider series of the
form
∑ αiei
where αi ’s are any scalars. Such a series converges and has the sum s if there exists an s ∈ H
such that the sequence (sn ) of the partial sums sn = ∑ni=1 αi ei converges to s, that is, ksn − sk →
0.
Theorem 0.22.4. Let (ek ) be a orthonormal sequence in a Hilbert space H. Then
(i) The series ∑ αi ei converges if and only if series ∑∞ 2
i |αi | converges.
(ii) If series ∑ αi ei converges, then the coefficients αi are the Fourier coefficients hx, ei i,
where x = ∑ αi ei . Hence in this case x = ∑ αi ei can be written as x = ∑ hx, ei i ei
(iii) for any x ∈ H the series x = ∑ αi ei with αi = hx, ei i converges.

Proof. (i) Let sn = ∑ni=1 αi ei and pn = ∑ni=1 |αi |2 . Then because of the orthonormality, for any
m and n > m,
ksn − sm k2 = kαm+1 em+1 + ... + αn en k2
= |αm+1 |2 + ... + |αn |2
= pn − pm .

72
0.22. SOME RESULTS IN INNER PRODUCT SPACES

Hence {sn } is Cauchy sequence in H if and only if (pn ) is Cauchy in R. Since H and R are
complete, we get the result.
(ii) Now take for j = 1, ...k and k ≤ n, we get sn , e j = α j . By assumption sn → x. Since
the inner product is continuous,

α j = sn , e j → x, e j .

Here we can take k(≤ n) as large as we please because n → ∞, so that we have α j = x, e j for
every j = 1, 2, ....
(iii) From the Bessel’s inequality, we see that the series ∑ | hx, ei i |2 converges. Therefore,
from (i), we get the result.

73
0.23. DUAL OF A HILBERT SPACE

0.23 Dual of a Hilbert space

Consider the set of all bounded linear functionals on a Hilbert space. These form a vector space
by the pointwise operations,
(αT + β S)(x) = αT (x) + β S(x).
We denote this vector space H ∗ it is called dual space or first conjugate space.

Theorem 0.23.1. (Riesz Representation Theorem) Let H be a Hilbert space. Every bounded,
linear functional on H can be represented in terms of the inner product on H, i.e.,

f (x) = hx, zi , (R1)

for all x ∈ H where z is uniquely determined by f and

kzk = k f k . (R2)

Proof. The proof is splitted up in several steps.

(1) First, we prove the existence of such an z ∈ H.
If f = 0, then (R1) and (R2) hold if z = 0.
Assume that there is some w 6= 0, such that f (w) 6= 0, (w ∈ H). The nullspace of f ,
N( f ) = {x ∈ H : f (x) = 0} is a closed linear subspace of H, hence N( f ) ⊕ N( f )⊥ = H. Let w ∈
N( f )⊥ such that w 6= 0 and set v = f (x)w − f (w)x, x ∈ H. Then, we have f (v) = f (x) f (w) −
f (w) f (x) = 0 and v ∈ N( f ). Since w ⊥ N( f ), we have
0 = hv, wi
= h f (x)w − f (w)x, wi
= f (x) hw, wi − f (w) hx, wi
= f (x) kwk2 − f (w) hx, wi .
 
f (w)hx,wi f (w)
Therefore, we have f (x) = . Also, it can be rewrite as f (x) = x, w. Then
kwk2 kwk2
f (w)
z= w.
kwk2
(2) Now, we prove the uniqueness of z.
Suppose that f (x) = hx, z1 i = hx, z2 i. Then we have hx, z1 − z2 i = 0, for all x ∈ H. Take
x = z1 − z2 . Then hz1 − z2 , z1 − z2 i = 0 or kz1 − z2 k2 = 0 or kz1 − z2 k = 0. Hence z1 = z2 .
(3) Now, we prove k f k = kzk.
If f = 0 then z = 0 and k f k = kzk = 0. For f 6= 0, z 6= 0. Consider

kzk2 = hz, zi = f (z) ≤ k f k kzk .

Hence kzk ≤ k f k. Also, | f (x)| = |hx, zi| ≤ kxk kzk[by Cauchy Schwarz]. So k f k ≤ kzk. This
yields k f k = kzk.

Theorem 0.23.2. Let H be a Hilbert space and H ∗ be its first conjugate space. Then for every
f ∈ H ∗ , there exists a unique z ∈ H such that

f (x) = hx, zi ,

for all x ∈ H and kzk = k f k.

74
0.23. DUAL OF A HILBERT SPACE

Example 0.23.3. Let H = L2 [0, 1] and D = C[0, 1] ⊂ L2 [0, 1]. Define linear functional T : D →
R as T f = f (0).
This operator is unbounded (and hence not-continuous), for take the sequence of functions,
(
1 − nx, 0 ≤ x ≤ 1n
fn (x) =
0, otherwise

kT fn k
Then k fn k → 0, where as kT fn k = | fn (0) = 1, that is, lim = ∞.
n→∞ k fn k

Example 0.23.4. Let H be a Hilbert space. Take y ∈ H and define the functional Ty = h., yi.
It is easy to see that it is linear. Also it is bounded as
| hy, yi | | hx, yi | kyk kxk
kyk = ≤ sup ≤ sup = kyk .
kyk x6=0 kxk x6=0 kxk

|Ty (x)|
Therefore, kT k = supx6=0 kxk = kyk .

Example 0.23.5. The Riesz representation does not hold for an incomplete inner product space.
Let X = c00 , the linear space of all scalar sequences each of which has only a finite number
of nonzero entries. For x, y ∈ X define hx, yi = ∑∞j=1 x( j)y( j).
p q
It is easy to see that h., .i is an inner product space and kxk2 = hx, xi = ∑∞j=1 |x( j)|2 , x ∈
X.
x(n)
Define f : X → K by f (x) = ∑∞ n=1 n , x ∈ X.
It is easy to see that f is linear. Also it is bounded by using Holder’s inequality, we have
∞
1 ∞
| f (x)|2 ≤ ( ∑ 2
)( ∑ |x(n)|2 ) ≤ C kxk2 .
n=1 n n=1

Hence f is continuous.
Also f ∈ X 0 has no representer in X.
To see this, let y ∈ X. As f (x) = hx, yi for all x ∈ X, then by taking en = (0, . . . , 0, 1, 0, . . .),
we found that y(n) = hen , yi = f (en ) = 1/n 6= 0 for all n ≥ 1. But this is impossible since
y ∈ c0 0. This shows that The Riesz representation does not hold for X.

Lemma 0.23.6. If hv1 , wi = hv2 , wi for all w in an inner product space X, then v1 = v2 . In
particular, hv1 , wi = 0 =⇒ v1 = 0 for all w ∈ X.

Proof. For all w, we have

hv1 − v2 , wi = hv1 , wi − hv2 , wi = 0.

For w = v1 − v2 , kv1 − v2 k2 = 0. Hence v1 − v2 = 0 so that v1 = v2 .

In particular, hv1 , wi = 0 with w = v1 gives kv1 k2 = 0, so v1 = 0.
The practical usefulness of bounded linear functionals on Hilbert spaces results to a large
extent from the simplicity of the Riesz representation (Rl).
Furthermore, (R1) is quite important in the theory of operators on Hilbert spaces. In par-
ticular, this refers to the Hilbert-adjoint operator T ∗ of a bounded linear operator T .

75
0.23. DUAL OF A HILBERT SPACE

Definition 0.23.1. Let X and Y be vector spaces over the same field K(= R or C). Then a
sesquilinear form (or sesquilinear functional) h on X ×Y is a mapping h : x ×Y → K such that
for all x, x1 , x2 ∈ X, y, y1 , y2 ∈ Y and all scalars α, β , we have
(a) h(x1 + x2 , y) = h(x1 , y) + h(x2 , y)
(b) h(x, y1 + y2 ) = h(x, y1 ) + h(x, y2 )
(c) h(α x, y) = αh(x, y)
(d) h(x, β y) = β h(x, y)
Hence h is linear in the first argument and conjugate linear in the second one.
Theorem 0.23.7. Let H1 , H2 be Hilbert spaces and h : H1 × H2 → K a bounded sesquilinear
form. Then h has a representation

h(x, y) = (Sx, y), (0.23.1)

where S : H1 → H2 is a bounded linear operator. S is uniquely determined by h and has norm

kSk = khk . (0.23.2)

Proof. We consider h(x, y) and this is linear in y because of bar. Keep x fixed. Then by Riesz
theorem on functionals, h(x, y) hy, zi. Hence
h(x, y) = hz, yi . (0.23.3)
here z ∈ H2 is unique but depends on fixed x ∈ H1 . It follows that (0.23.3) with variable x
defines an operator S : H1 → H2 given by Sx = z. So by substituting z = Sx in (0.23.3), we have
(0.23.1).
First, we prove S linear. In fact, its domain is the vector space H1 and from (0.23.1) and the
sesquilinearity we obtain
hS(αx1 + β x2 ), yi = h(αx1 + β x2 , y)
= h(αx1 , y) + h(β x2 , y)
= α hS(x1 ), yi + β hS(x2 ), yi
= hαS(x1 ) + β S(x2 ), yi ,
for all y ∈ H2 , so S(αx1 + β x2 ) = αS(x1 ) + β S(x2 ).
Now, we prove that S is bounded.

|h(x, y)|
khk = sup x, y 6= 0
kxk kyk
| hSx, yi |
= sup x, y 6= 0
kxk kyk
| hSx, Sxi |
≥ sup x, Sx 6= 0
kxk kSxk
||Sx||
= sup x 6= 0 = kSk .
kxk
This proves boundedness. Moreover ||h|| ≥ ||S||. We obtain (0.23.2) by noting ||h|| ≤ ||S||
follows by CS inequality
| hSx, yi | ||Sx||||y||
khk = sup x, y 6= 0 ≤ sup x, y 6= 0 = kSk .
kxk kyk kxk kyk

76
0.23. DUAL OF A HILBERT SPACE

Now, to prove S is unique. In fact, assuming that there is a linear operator T : H1 → H2

such that for all x ∈ H1 and y ∈ H2 we have h(x, y) = hSx, yi = hT x, yi, we see that Sx = T x, for
all x ∈ H1 . Hence S = T.

77
0.24. ORTHOGONAL BASIS

0.24 Orthogonal Basis

Let V be an IPS. A set S of vectors in V is called an orthogonal basis of V , if
(i) S is an orthogonal set
(ii) S is basis for V .
Further, an orthonormal basis S is an orthonormal basis for V if kvk = 1, for all v ∈ S.
Example 0.24.1. (1) The standard basis {(1, 0, . . . , 0), (0, 1, 0, . . . , 0), . . . , (0, 0, . . . , 1)} in Rn is
an orthonormal basis for Rn .
(2) set {(1, 1), (−1, 1)} is an orthogonal set in R2 w.r.t. standard inner product.
√ Here h(1,√ 1), (−1, 1)i = 1(−1)+1.1 = 0 but this set is not orthonormal set because k(1, 1)k =
12 + 12 = 2 6= 1.
(3) One checks that for u = (x1 , x2 ), v = (y1 , y2 ) ∈ R2 , h(u, v)i = x1 y1 − x2 y1 − x1 y2 + 4x2 y2
is an inner product in R2 w.r.t this inner product set {(1, 1), (−1, 1)} is not an orthogonal set
h(1, 1), (−1, 1)i = 3 6= 0.
(4) Consider V = C[−π, π], the set of all real valued continuous functions on [−π, π]. V
1 Rπ
is an IPS w.r.t h f , gi = π −π f (x)g(x)dx. w.r.t this inner product the set {sin nx : n ∈ N} is an
orthonormal set.
For this one can easily checks that
(
1 π 1, m = n
Z
sin nx sin mxdx =
π −π 0, m 6= n.

Theorem 0.24.2. Every orthogonal set of nonzero vectors in an IPS is LI.

Proof. Let S be a finite or infinite set of nonzero orthogonal vectors in the given space.
Let v1 , v2 , ..., vm be a set of m vectors in S, m ≤ dim S. Let v = α1 v1 + ... + αm vm , α ∈ F.
For any k, 1 ≤ k ≤ m, hv, vk i = ∑m i=1 αi hvi , vk i = αk hvk , vk i.
hv,vk i
Since vk 6= 0 hvk , vk i 6= 0. αk = hv ,v i = hv,vk2i , k = 1, 2, ..., m.
k k kvk k
If v = 0, then αk = 0 for k = 1, ..., m. Hence S is a LI set.

Corollary 0.24.3. If S = {v1 , ..., vn } is an orthogonal basis for V , then for v ∈ V the kth co-
ordinate of v is given by hv,vk2i , that is if v = (α1 , ...αn ) ∈ V , then αk hv,vk2i .
kvk k kvk k

Corollary 0.24.4. If V is an n-dimensional IPS then V can have atmost n mutually orthogonal
vectors.

0.24.1 Gram-Schmidt Process

Gram-Schmidt process was designed by E. Schmidt (1907). The orthonormal sequences are
very convenient to work with. The remaining practical problem is how to obtain an orthonor-
mal sequence if an arbitrary linearly independent sequence is given. This is accomplished by
a constructive procedure, the Gram-Schmidt process for orthonormalizing a linearly indepen-
dent sequence (xi ) in an inner product space. The resulting orthonormal sequence (ei ) has the
property that for every n, span{e1 , ..., en } = span{x1 , ..., xn }.
The first element is e1 = kxx1 k .
1
Then x2 can be written as x2 = hx2 , e1 i e1 + v2 . Then v2 = x2 − hx2 , e1 i e1 is not the zero
vector since (xi ) is linearly independent; also v2 ⊥ e1 since hv2 , e1 i = 0, so that we can take
e2 = kvv2 k .
2

78
0.24. ORTHOGONAL BASIS

The vector v3 = x3 − hx3 , e1 i e1 − hx3 , e2 i e2 is not the zero vector, and v3 ⊥ e1 , v3 ⊥ e2 . We

take e3 = kvv3 k and so on. Thus in nth step, we observe that
3

n−1
vn = xn − ∑ hxn , ei i ei ,
i=1
vn
is not the zero vector and is orthogonal to e1 , e2 , ..., en . From it we obtain en = kvn k .
u.v
Definition 0.24.1. If u and v belong to inner product space X and v 6= 0, then vector v is
kvk2
called the vector projection of u along v.

Theorem 0.24.5. Let V be an IPS and {v1 , ..., vn } be a set of LI vectors in V . Then one can
constructa set {u1 , ..., un } of orthogonal vectors from {v1 , ..., vn } such that for any k, 1 ≤ k ≤ n,
{u1 , ..., uk } is a basis for L({v1 , ..., vn }).

Proof. We construct vectors u1 , ..., un in following way: (1) i = 1

(2) ui = vi
(3) i = i + 1
hvi ,u j i
(4) if i > n, then stop, other wise ui = vi − ∑i−1
j=1 u 2 u j
k jk
(5) Go to step 3.
Note that ui 6= 0, then vi be a linear combination of v1 , ..., vi−1 , which is not true.
Next we check {u1 , ..., un } is an orthogonal set. Consider
* +
hv2 , u1 i hv2 , u1 i
hu2 , u1 i = v2 − 2
, u1 = hv2 , u1 i − hu1 , u1 i = 0.
ku1 k ku1 k2

Hence u1 and u2 are orthogonal.

In general for any positive integer m ≤ n and 1 ≤ r ≤ m, we have
* +
m v , u u
m+1 j j
hum+1 , ur i = vm+1 − ∑ 2
, ur .
j=1 uj

On simplifying, we get
m vm+1 , u j
hum+1 , ur i = hvm+1 , ur i − ∑ 2
u j , ur
j=1 uj
= hvm+1 , ur i − hvm+1 , ur i = 0.

Hence {u1 , ..., un } is orthogonal set.

Next for any k, 1 ≤ k ≤ n, u1 , ...uk can be expressed as linear combinations of vectors
v1 , ...vk . Since {u1 , ..., uk } is LI set, we get {u1 , ..., uk } is a basis for set L({v1 , ..., vk }).
Remarks 0.24.1. From {v1 , ..., vk } an orthonormal set of vectors {w1 , ..., wk } can be con-
structed as wi = ui / kui k.

Corollary 0.24.6. Every finite dimensional IPS has an orthonormal basis.

79
0.24. ORTHOGONAL BASIS

Example 0.24.7. Consider R3 with standard inner product. Let v1 = (3, 0, 4), v2 = (−1, 0, 7), v3 =
(2, 9, 11). {v1 , v2 , v3 } is LI set of vectors in R3 .
u1 = v1 = (3, 0, 4)
u2 = v2 − hv2 ,u12i u1 = (−1, 0, 7) − 25/25(3, 0, 4) = (−4, 0, 3)
ku1 k
hv3 ,u1 i
u3 = v3 − u − hv2 ,u22i u2 = (0, 9, 0)
ku1 k2 1 ku2 k
{u1 , u2 , u3 } = {(3, 0, 4), (−4, 0, 3), (0, 9, 0)} is an orthogonal set of vectors in R3 .
Further {ui / kui k : i = 1, 2, 3} = { 51 (3, 0, 4), 15 (−4, 0, 3), (0, 1, 0)} is an orthonormal set ob-
tained from {v1 , v2 , v3 }.

80
0.25. SOME EXAMPLES (IPS)

0.25 Some Examples (IPS)

Example 0.25.1. Let u = (1, 3, −4, 2), v = (4, −2, 2, 1), w = (5, −1, −2, 6) ∈ R4 . Find kuk , kvk
and show that h3u
√− 2v, wi = 3 hu, wi√− 2 hv, wi . √
Here kuk = 1 + 9 + 16 + 4 = 30, kvk = 16 + 4 + 4 + 1 = 5. We normalize u and v
u
to obtain the following unit vectors in the directions of u and v, respectively: û = kuk =
√ √ √ √ v
(1/ 30, 3/ 30, −4/ 30, 2/ 30) and v̂ = kvk = (4/5, −2/5, 2/5, 1/5).
Also hu, wi = 5 − 3 + 8 + 12 = 22, hv, wi = 20 + 2 − 4 + 6 = 24. So 3 hu, wi − 2 hv, wi =
3(22) − 2(24) = 18.
Now, 3u − 2v = (−5, 13, −16, 4) and h3u − 2v, wi = −25 − 13 + 32 + 24 = 18 = 3 hu, wi −
2 hv, wi .
Function Space C[a, b] and Polynomial Space P(t)
The notation C[a, b] is used to denote the vector space of all continuous functions on the
closed interval [a, b] —that is, where a ≤ t ≤ b. The following defines an inner product on
C[a, b], where f (t) and g(t) are functions in C[a, b]:
Z b
h f , gi = f (t)g(t) dt
a

It is called the usual inner product on C[a, b].

The vector space P(t) of all polynomials is a subspace of C[a, b] for any interval [a, b], and
hence, the above is also an inner product on P(t).
Angle between vectors: For any nonzero vectors u and v in an inner product space V ,
the angle between u and v is defined to be the angle y such that 0 ≤ θ ≤ π and cos θ =
hu, vi /(kuk kvk).
Example 0.25.2. Consider f (t) = 3t − 5 and g(t) = t 2 in the polynomial space P(t) with inner
product
Z 1
h f , gi = f (t)g(t) dt.
0
Find h f , gi and k f k , kgk the angle θ between f and g.
Solution. Here f (t) g(t) = 3t 3 − 5t 2 . Hence h f , gi = 01 f (t)g(t) dt = 01 (3t 3 − 5t 2 ) dt =
R R

(3/4)t 4 − (5/3)t 3 |10 = 3/4 − 5/3 = −11/12.

Here | f (t)|2 = f (t) f (t) = 9t 2 − 30t + 25 and |g(t)|2 = t 4 .
k f k2 = h f , f i = 01 (9t 2 − 30t + 25)dt = 13.
R
R1 4
kgk2 = hg, gi = √ 0 t dt = 1/5.p
Hence k f k = 13 and kgk = 1/5.
So cos θ = h f , gi /(k f k kgk) = √ −11/12√ .
( 13)( 1/5)

Example 0.25.3. Consider the functions sint and cost in the vector space C[−π, π] of contin-
uous functions on Rthe closed interval [−π, π]. Find hsint, costi .
π
hsint, costi = −π sint cost dt = 0.
Thus it is orthogonal.
Example 0.25.4. Find a nonzero vector w that is orthogonal to u1 = (1; 2; 1) and u2 = (2; 5; 4)
in R3 .
Let w = (x; y; z). Then we want hu1 , wi = 0 and hu2 , wi = 0. This yields the homogeneous
system

81
0.25. SOME EXAMPLES (IPS)

( (
x + 2y + z = 0 x + 2y + z = 0
2x + 5y + 4z = 0 y + 2z = 0

Here z is the only free variable in the echelon system. Set z = 1 to obtain y = −2 and x = 3.
Thus, w = (3, −2, 1) is a desired nonzero vector orthogonal to u1 and u2 .
Example 0.25.5. Let V be the vector space of polynomials P3 (t) with inner product h f , gi =
R1 2 3
−1 f (t)g(t) dt. Apply the Gram–Schmidt orthogonalization process to {1;t;t ;t } to find an
orthogonal basis { f0 ; f1 ; f2 ; f3 } with integer coefficients for P3 (t).
Solution. Now, we know that (
1 t r+s+1 1 2/(r + s + 1), r + s is even
ht r ,t s i = −1 t r+s dt = r+s+1
R
|−1 =
0, r + s is odd.
(1) First set f0 = 1.
ht,1i
(2) Compute t = h1,1i (1) = t − 0 = t. Set f 1 = t.
ht 2 ,1i ht 2 ,t i
(3) Compute t 2 − h1,1i (1) − ht,ti (t) = t 2 − 2/3 2 1
2 (1) + 0(t) = t − 3 .
Multiply by 3 to obtain f2 = 3t 2 − 1.
(4) Compute

3 t 3, 1 t 3 ,t t 3 , 3t 2 − 1
t − (1) − (t) − 2 (3t 2 − 1)
h1, 1i ht,ti h3t − 1, 3t 2 − 1i
2/5
= t 3 − 0(1) − (t) − 0(3t 2 − 1)
2/3
3
= t 3 − (t).
5
Multiply by 5 to obtain f3 (t) = 5t 3 − 3(t).
Thus {1,t, 3t 2 − 1, 5t 3 − 3t} is the required orthogonal basis.
Example 0.25.6. The vectors u1 = (1; 1; 0), u2 = (1; 2; 3), u3 = (1; 3; 5) form a basis S for Eu-
clidean space R3 . Find the matrix A that represents the inner product in R3 relative to this basis
S.
Solution. First compute ui , u j to obtain

hu1 , u1 i = 1 + 1 + 0 = 2 hu1 , u2 i = 1 + 2 + 0 = 3 hu1 , u3 i = 1 + 3 + 0 = 4

hu2 , u2 i = 1 + 4 + 9 = 14 hu2 , u3 i = 1 + 6 + 15 = 22 hu3 , u3 i = 1 + 9 + 25 = 35.
 
2 3 4
So matrix A = 3 14 22
4 22 35

82
0.26. TOTAL ORTHONORMAL SETS

0.26 Total Orthonormal sets

Definition 0.26.1. A total set (or fundamental set) in a normed space X is a subset M ⊂ X
whose span is dense in X.
Accordingly, an orthonormal set (or sequence or family) in an inner product space X which
is total in X is called a total orthonormal set (or sequence or family, respectively) in X.
M is total if and only if spanM = X.
Definition 0.26.2. A non empty subset M of an inner product space X is said to be total iff the
vector x ∈ X is orthogonal to all the vector in M is the null vector only.
Definition 0.26.3. Let X be an inner product space. Then an orthonormal set M in X is said to
be complete if M is a maximal orthonormal set in X, i.e, if there does not exist any orthonormal
set N in X such that M ( N.
Theorem 0.26.1. Let X be an inner product space and M ⊂ X. Then
(1) If M is total in X then x ⊥ M =⇒ x = 0.
(2) If X is complete, then M is total in X.
Proof. (1) Let H be the completion of X. Then, X regarded as a subspace of H, is dense in H.
By assumption, M is total in X, so span(M) is dense in X, and dense in H. It follows that the
orthogonal complement of M in H is {0}.
(2) If X is a Hilbert space and M ⊥ = {0}, then M is total in X (For any subset M 6= 0/ of a
Hilbert space H, the span of M is dense in M iff M ⊥ = {0}.
Theorem 0.26.2. Any dense subset of an inner product space is complete.
Proof. Let A be dense subset of inner product space X. So there exists a sequence of elements
{xn } in A such that xn → x ∈ X.
Let x ⊥ A. Now, hx, xn i → hx, xi = kxk2 . But hx, xn i = 0 for all n. So kxk2 = 0, or x = 0.
Hence A is total.
Theorem 0.26.3. An orthonormal set A in an inner product space X is complete iff x ⊥ A, then
x = 0 in X.
Proof. Let A be an orthonormal set in X which is complete. Assume that x ∈ X such that x ⊥ A,
x
x 6= 0. Let y = ||x|| . So ||y|| = 1. Then A ∪ {y} is an orthonormal set in X contains A properly,
a contradiction to the fact that A is complete. So x = 0.
Conversely, let x ⊥ A, so x = 0. Suppose that A is not complete. Then there exists an
orthonormal set B in X such that A ( B. If x ∈ B − A, then x ∈ B, x ∈ / A. As x ∈ B, ||x|| = 1
=⇒ x 6= 0 , a contradiction.
So A is a maximal orthonormal set and hence A is complete.
Thus orthonormal set A for X in inner product space X is complete iff it is total in X.

Example 0.26.4. Let X = `2 . Define inner product in X by

hx, yi = ∑ ξi ηi ,
x = (ξ1 , ...), y = (η1 , ....) ∈ `2 .
Let A = {e1 , e2 , ...} be an orthonormal set in X where ei = (0, 0..., 1 , 0, ...). Let x ∈ X
↑
ith place
such that kxk = 1 and x ⊥ A. So hx, ei i = 0. Now {x} ∪ A is an orthonormal set and A ⊂

83
0.26. TOTAL ORTHONORMAL SETS

{x, e1 , ...}. As x = ∑ ξi ei , we have hx, ei i = ξi , which implies ξi = 0 for all i. So x = (0, 0, 0, ...).
Hence A is complete.

Proposition 0.26.5. (Parseval’s equality) Show that an orthonormal set M in a Hilbert space H
is total iff
∑|hx, ek i|2 = kxk2 ,
k
for all x ∈ M.
Proof. If M is not total, there is a nonzero x ⊥ M in H. Since x ⊥ M we have 0 on the left-
hand side of Parseval’s Equality, whereas right hand side, which is equal to kxk2 6= 0. Hence if
Parseval’s equality holds for all x ∈ H, then M must be total in H.
Conversely, assume that M to be total in H. Consider any x ∈ H and its nonzero Fourier
coefficients arranged in a sequence, i.e., hx, e1 i , hx, e2 i .... Define

y = ∑ hx, ek i ek .

First we show that x − y ⊥ M. For every ei occuring in y = ∑ hx, ek i ek and using the orthonor-
mality, we have

hx − y, ei i = hx, ei i − ∑ hx, ek i hek , ei i = hx, ei i − hx, ei i .

Now, for every v ∈ M not in y = ∑ hx, ek i ek , we have hx, vi = 0, so that

hx − y, vi = hx, vi − ∑ hx, ek i hek , vi = 0 − 0 = 0.

It follows that x − y ∈ M ⊥ . Since M is total, we have M ⊥ = {0}. Hence x − y = 0, x = y.

Therefore, * +
kxk2 = ∑ hx, ek i ek , ∑ hx, emi em = ∑ hx, ek i hx, ek i.
k m k

Hence the result.

84
0.27. ADJOINT OPERATOR

0.27 Adjoint Operator

The Representation Theorem of functionals on Hilbert spaces will enable us to introduce the
Hilbert-Adjoint Operator of a bounded linear operator on a Hilbert space. This operator was
suggested by problems in matrices and linear differential and integral equations. We’ll see
that it also helps us to define three important classes of operators (called self-adjoint, normal
operators, unitary operators) which have been studied extensively because they play a key role
in various applications.
Definition 0.27.1. Let H1 and H2 be Hilbert spaces and T : H1 → H2 be a bounded linear
operator. Then adjoint is T ∗ : H2 → H1 such that hT x, yi = hx, T ∗ yi, for all x ∈ H1 , y ∈ H2 .
Existence of Hilbert Adjoint operator:
Theorem 0.27.1. Let H be a complex Hilbert space, BLd (H, H) be the collection of all bounded
linear operator from H onto itself and T ∈ BLd (H, H). Then there exists a T ∗ ∈ BLd (H, H),
such that ||T || = ||T ∗ ||.
Proof. Let T ∈ BLd (H, H) and y ∈ H be fixed. Define fy : H → D by fy (x) = hT x, yi, for all
x ∈ H. It is easy to see that fy is linear. Further,
| fy (x)| = | hT x, yi | ≤ ||T x|| ||y|| ≤ ||T || ||x|| ||y||.
So, fy is bounded i.e. fy ∈ H ∗ . So, by Riesz Representation Theorem, for f y ∈ H ∗ there exist
a unique y∗ ∈ H such that fy (x) = hx, y∗ i, for all x ∈ H. Thus for fy ∈ H ∗ we get, y∗ ∈ H. Let
us now define a map T ∗ : H → H by T ∗ (y) = y∗ . So, fy (x) = hx, y∗ i = hx, T ∗ yi, for all x ∈ H.
So hT x, yi = hx, T ∗ yi, for all x ∈ H. Thus for T ∈ BLd (H, H), we get a map T ∗ : H → H. Thus
T ∗ is called the adjoint of T . Now, T ∗ is well defined. If possible let T1∗ : H → H such that
hT x, yi = hx, T1∗ yi, for all x, y ∈ H.
Also we have hT x, yi = hx, T1∗ yi, for all x, y ∈ H. So hx, T ∗ yi = hx, T1∗ yi, for all x, y ∈ H.
Therefore for all x, y ∈ H, we get
hx, T ∗ y − T1∗ yi = 0
hx, (T ∗ − T1∗ )yi = 0
(T ∗ − T1∗ )y = 0 ∈ Y,
T ∗ − T1∗ = Θ, the zero operator on H.
So T ∗ = T1∗ . Therefore, we have hT x, yi = hx, T ∗ yi .
Consider, for all x, y1 , y2 ∈ H,
hx, T ∗ (y1 + y2 )i = hT x, y1 + y2 i
= hT x, y1 i + hT x, y2 i
= hx, T ∗ y1 i + hx, T ∗ y2 i
= hx, T ∗ y1 + T ∗ y2 i

So T ∗ (y1 + y2 ) = T ∗ y1 + T ∗ y2 .
Now, we have to prove that T ∗ (αy) = α(T ∗ y) for y ∈ H and α ∈ C. Consider
hx, T ∗ (αy)i = hT x, αyi
= α hT x, yi
= α hx, T ∗ yi
= x, αT ∗ y .

85
0.27. ADJOINT OPERATOR

Therefore T ∗ (αy) = α(T ∗ y) and T ∗ is linear. For y ∈ H, we have

kT ∗ yk2 = hT ∗ y, T ∗ yi = hT T ∗ y, yi ≤ kT T ∗ yk kyk ≤ kT k kT ∗ yk kyk .

Thus kT ∗ yk ≤ kT k kyk, that is T ∗ is bounded.

∗
Also, sup kTkykyk ≤ kT k. Hence kT ∗ k ≤ kT k.
y6=0
Now, we have to show that T ∗∗ = T . For all x, y ∈ H, we have

hT x, yi = hx, T ∗ yi .

Replacing T by T ∗ , we get
hT ∗ x, yi = hx, T ∗∗ yi .
Now interchanging x, y here we get hT ∗ y, xi = hy, T ∗∗ xi. Taking conjugates on both sides, we
have hx, T ∗ yi = hT ∗∗ x, yi . Hence hT x, yi = hx, T ∗ yi = hT ∗∗ x, yi , for all x, y ∈ H. Therefore
T = T ∗∗ .
Also, kT ∗ k ≤ kT k, replacing T by T ∗ we get kT ∗∗ k ≤ kT ∗ k. So, we have kT k ≤ kT ∗ k.
Therefore, kT k = kT ∗ k.
Note that the adjoint of identity operator I : H → H is the identity operator itself.
Similarly the adjoint of the Θ : H → H is the null operator itself.
Clearly I ∗ , Θ∗ ∈ BLd (H, H) with ||I|| = ||I ∗ || and ||Θ|| = ||Θ∗ ||.
Let H be a complex Hilbert space and let T1 and T2 be two bounded linear operators
mapping H onto itself i.e. T1 , T2 ∈ BLd (H, H). Then T1 +T2 and αT1 (α any complex scalar)
are also BLd (H, H).

Theorem 0.27.2. Let H be a Hilbert space and T1 , T2 : H → H be bounded linear operators.

Then
(i) (T1 + T2 )∗ = T1∗ + T2∗
(ii) (αT1 )∗ = α T1∗ for α ∈ C
(iii) : (T1 T2 )∗ = T2∗ T1∗

Proof. (i) For all x, y ∈ H, we have

hT1 x, yi = hx, T1∗ yi and hT2 x, yi = hx, T2∗ yi .

Now

hx, (T1 + T2 )∗ yi = h(T1 + T2 )x, yi

= hT1 x + T2 x, yi
= hT1 x, yi + hT2 x, yi
= hx, T1∗ yi + hx, T2∗ yi
= hx, T1∗ y + T2∗ yi
= hx, (T1∗ + T2∗ )yi .

This shows that (T1 + T2 )∗ = T1∗ + T2∗ .

(ii) For x, y ∈ H, we have

hx, (αT1 )∗ yi = h(αT1 )x, yi = α hT1 x, yi = α hx, T1∗ yi = hx, αT1∗ yi .

This being true for all x, y ∈ H. So we have (αT1 )∗ = α T1∗ .

86
0.27. ADJOINT OPERATOR

(iii) To prove (T1 T2 )∗ = T2∗ T1∗ . For x, y ∈ H, we have

hT1 x, yi = hx, T1∗ yi and hT2 x, yi = hx, T2∗ yi .

Now

h(T1 T2 )x, yi = hx, (T1 T2 )∗ yi

= h(T1 T2 )x, yi
= hT2 x, T1∗ yi
= hx, T2∗ T1∗ yi .

Hence (T1 T2 )∗ = T2∗ T1∗

Theorem 0.27.3. For any T ∈ BLd (H, H), ||T ∗ T || = ||T ||2 = ||T T ∗ ||.

Proof. We always have ||T ∗ T || ≤ ||T ∗ ||||T || = ||T || ||T || = ||T ||2 , because T ∗ ∈ BLd (H, H)
and ||T || = ||T ∗ ||. Thus ||T ∗ T || ≤ ||T ||2 .
Again

kT k2 = sup {kT xk2 }

kxk≤1
= sup {| hT x, T xi |}
kxk≤1
= sup {| hT ∗ T x, xi |}
kxk≤1
≤ sup {kT ∗ T xk kxk}( using CS inequality)
kxk≤1
∗
≤ kT T k .

Thus kT ∗ T k = kT k2 .
On replacing T by T ∗ and using T ∗ ∗ = T , we get kT T ∗ k = kT ∗∗ T ∗ k = k(T ∗ )∗ T ∗ k =
kT ∗ k2 = kT k2 . Hence the result.

87
0.28. SELF ADJOINT, NORMAL AND UNITARY OPERATORS

0.28 Self Adjoint, Normal and Unitary operators

Definition 0.28.1. T ∈ BLd (H, H) is called self adjoint if T = T ∗ , that is hT x, yi = hx, Tyi, for
all x, y ∈ H.
Theorem 0.28.1. (a) If T1 and T2 are two self adjoint operators over H then T1 + T2 is self
adjoint.
(b) If T1 is self adjoint and α be any real scalar then αT1 is self adjoint.
(c) For any T ∈ BLd (H, H), T ∗ T, T T ∗ and T + T ∗ are self adjoint.
(d) If T1 and T2 are self adjoint operators then T1 T2 is self adjoint if and only if T1 T2 = T2 T1 .

Proof. (a) (T1 + T2 )∗ = T1∗ + T2∗ = T1 + T2 .

(b) (α T1 )∗ = αT1∗ = αT1 = αT1 , because α is a real scalar.
(c) (T ∗ T )∗ = T ∗ (T ∗ )∗ = T ∗ T and (T T ∗ )∗ = (T ∗ )∗ T ∗ = T ∗∗ T ∗ = T T ∗ . Also (T + T ∗ )∗ =
T ∗ + (T ∗ )∗ = T ∗ + T ∗∗ = T ∗ + T = T + T ∗ .
(d) Finally (T1 T2 )∗ = T2∗ T1∗ = T2 T1 . So (T1 T2 )∗ = T1 T2 if and only if T1 T2 = T2 T1 .

Theorem 0.28.2. Let T ∈ BLd (H, H) be such that for all x, y ∈ H, hT x, yi = 0. Then T equals
to the zero operator and conversely.

Proof. For the zero operator Θ, we always have hΘx, yi = h0, yi = 0.

Conversely, let for all x, y ∈ H, hT x, yi = 0. Let x ∈ H be fixed and consider hT x, yi = 0 for
all y ∈ H. That means T x = 0 in H. Now let x be free and we see T x = 0 for x ∈ H, showing
that T = Θ.
Corollary 0.28.3. Let H be complex and T ∈ BLd (H, H) such that for all x ∈ H, hT x, xi = 0.
Then T equals to the zero operator

Proof. If x, y ∈ H and α, β are any two scalars, we have

0 = hT (αx + β y), αx + β yi
= hαT x + β Ty, αx + β yi
= αα hT x, xi + αβ hT x, yi + β α hTy, xi + β β hTy, yi
= αβ hT x, yi + β α hTy, xi .
Take α = 1 and β = 1, we get hT x, yi + hTy, xi = 0.
Also, choose α = i and β = 1, we get i hT x, yi − i hTy, xi = 0 =⇒ hT x, yi − hTy, xi = 0.
By adding these two, we get hT x, yi = 0. Hence by above theorem T = 0.

Theorem 0.28.4. Let T ∈ BLd (H, H). Then for all x ∈ H

(i) hT x, xi is real, if T is self adjoint.
(ii) T is self adjoint if H is complex and hT x, xi is real.
Proof. (i) If T is self adjoint then for all x ∈ H, we have
hT x, xi = hx, T xi = hT ∗ x, xi = hT x, xi
.
(ii) If hT x, xi is real, then
hT x, xi = hT x, xi = hx, T ∗ xi = hT ∗ x, xi .
Hence hT x, xi − hT ∗ x, xi = 0 h(T − T ∗ )x, xi = 0. Hence T = T ∗ .

88
0.28. SELF ADJOINT, NORMAL AND UNITARY OPERATORS

Theorem 0.28.5. Let (Tn ) be a sequence of bounded self-adjoint linear operators Tn : H → H

on a Hilbert space H. Suppose that (Tn ) converges, say, Tn → T , that is kTn − T k → 0, where
k.k is the norm on the space B(H, H). Then the limit operator T is a bounded self-adjoint linear
operator on H.

Proof. Here we have to show that T ∗ = T . Consider

kTn∗ − T ∗ k = k(Tn − T )∗ k = kTn − T k .

Also, by using triangle inequality, we have

kT − T ∗ k ≤ kT − Tn k + kTn − Tn∗ k + kTn∗ − T ∗ k

= kT − Tn k + 0 + kTn − T k
= 2 kTn − T k → 0.

Hence kT − T ∗ k and T ∗ = T .
Definition 0.28.2. Let H be a Hilbert space and T ∈ BLd (H, H) and T ∗ be adjoint of T . T is
said to be normal if T T ∗ = T ∗ T .
Remark 0.28.1. Every self adjoint operator is normal.
From the fact that T ∗∗ = T , we can say immediately that, if T is normal then its adjoint T ∗
is also normal.
Definition 0.28.3. A bounded linear operator T : H → H is said to be unitary if it satisfies the
condition T T ∗ = T ∗ T = I, where I is the identity operator on H.
If T is unitary, then T is injective. If T x1 = T x2 for x1 , x2 ∈ H then T ∗ T x1 = T ∗ T x2 and
since T ∗ T = I, we get x1 = x2 .
Also if T unitary then T is surjective. Because if y ∈ H then T (T ∗ y) = (T T ∗ )y = Iy = y.
Also it follows that unitary operators on H are precisely those operators whose inverses are
equal to their adjoints, that is T −1 = T ∗ .
Note that every self adjoint operator is normal. Also it is evident that every unitary
operator is normal.
A normal operator need not be self adjoint or unitary.
Example 0.28.6. Let I : H → H be an identity mapping and T = 2i I. Then T ∗ = (2iI)∗ = −2i I
and T −1 = −1 ∗ ∗ ∗ ∗
2 i I. So T T = T T = 4I. But T 6= T . Also T 6= T
−1 . So T is normal which is

neither self adjoint nor unitary.

Theorem 0.28.7. Let H be a Hilbert space, and T : H → H be unitary. Then T is isometry, that
is, hT x, T xi = hx, xi or kT xk = kxk, for all x ∈ H.

Proof. For all x ∈ H, we get

kT xk2 = hT x, T xi
= hx, T ∗ T xi
= hx, Ixi
= kxk2 .

Hence kT xk = kxk.

89
0.28. SELF ADJOINT, NORMAL AND UNITARY OPERATORS

Theorem 0.28.8. A bounded linear operator T on a Hilbert space H into itself is unitary if and
only if T is an isometrically isomorphism of H onto itself.

Proof. Suppose that T is an isometric and onto. Isometry implies injectivity, so T is bijective.
Hence T −1 exists and kT xk = kxk. Consider

hT ∗ T x, xi = hT x, T xi = hx, xi
h(T ∗ T − I)x, xi = 0.
T ∗ T = I.

On the similar lines, we have T T ∗ = I. Hence T is unitary.

Conversely, if T is unitary then T is injective and surjective. But T ∗ = T −1 . Also T T ∗ = I.
So for all x ∈ H, we get

hT x, T xi = hx, T ∗ T xi = hx, Ixi .

So T is an isomorphism of H onto itself.

90
0.29. LINEAR OPERATOR

0.29 Linear Operator

Definition 0.29.1. Let X,Y be vector spaces. A linear operator T is an operator, that is,
into
T : X → Y such that the domain D(T ) of T is a vector space and the range R(T ) lies in a vec-
tor space over the same field and for all x, y ∈ D(T ) and scalars α, β such that T (αx + β y) =
αT x + β Ty.
into
Definition 0.29.2. Let X,Y be vector spaces and T : X → Y be linear operator. Then
(1) The null space of T is N(T ) = {x ∈ D(T )such thatT x = 0}
(2) The domain of T , D(T ), is a vector space
(3) The range of T , R(T ), lies in a vector space
onto
Note that obviously D(T ) ⊂ X. It is customary to write T : D(T ) → R(T ).
Also, T 0 = T (0 + 0) = T 0 + T 0 =⇒ T 0 = 0 =⇒ N(T ) 6= 0. /
Example 0.29.1. 1. (Dot Product) Let T : R3 → R defined by T x = x.a = ξ1 α1 + ξ2 α2 +
ξ3 α3 where a = (αi ) ∈ R3 is fixed.
2. (Matrix) Let A = (ai j )n×m . Define T : RmtoRn by T x = Ax.
onto
3. (Differentiation). Let X be the space of polynomials on [a, b]. Then define T : X → X as
T x(t) = dtd x(t).
into Rt
4. (Integration). Let X = C[a, b]. Then define T : X → X as T x(t) = a x(s) ds.

Theorem 0.29.2. Let T be a linear operator. Then

(1) Range R(T ) is a vector space.
(2) If dim D(T ) = n < ∞, then dim R(T ) ≤ n.
(3) Null space N(T ) is a vector space.

Proof. (1) Let y1 , y2 ∈ R(T ) and α, β be arbitrary scalars. Since y1 , y2 ∈ R(T ), there exists
x1 , x2 ∈ D(T ) such that T x1 = y1 and T x2 = y2 . Since D(T ) is a vector space, αx1 + β x2 ∈
D(T ), so T (αx1 + β x2 ) ∈ R(T ). By the linearity of T , we have T (αx1 + β x2 ) = αT x1 +
β T x2 = αy1 + β y2 . Hence αy1 + β y2 ∈ R(T ), R(T) is a vector space.
(2) Choose n + 1 elements y1 , ..., yn of R(T ) such that yi = T xi for all i = 1, ..., n + 1 for
some x1 , ..., xn+1 ∈ D(T ).
Since dim D(T ) = n, so the set {x1 , ..., xn+1 } is linearly dependent. Hence α1 x1 + ... +
αn+1 xn+1 = 0, for some scalars α1 , ..., αn+1 not all zero. Since T is linear and T 0 = 0, we get
T (α1 x1 + ... + αn+1 xn+1 ) = α1 y1 + ... + αn+1 yn+1 = T (0) = 0.
This shows that {y1 , ..., yn+1 } is a linearly dependent set because the α’s are not all zero.Hence
dim R(T ) ≤ n.
(3) Take x1 , x2 ∈ N(T ). Then T x1 = 0 = T x2 . Since T is linear, for any scalars α, β we
have T (αx1 + β x2 ) = αT x1 + β T x2 = 0 + 0 = 0. This shows that αx1 + β x2 ∈ N(T ). Hence
N(T ) is a vector space.
into
Definition 0.29.3. A mapping T : D(T ) → Y is injective (or one-to-one) if for all x1 , x2 ∈ D(T )
such that x1 6= x2 =⇒ T x1 6= T x2 .
Equivalently, T x1 = T x2 =⇒ x1 = x2 .
So there exists the mapping T −1 : R(T ) → D(T ), y0 7→ x0 , that is y0 = T x0 which maps
every y0 ∈ R(T ) onto that x0 ∈ D(T ) for which T x0 = y0 . The mapping T −1 is called inverse
of T .
Therefore, we have T −1 T x = x for all x ∈ D(T ) and T T −1 y = y for all y ∈ R(T ).

91
0.29. LINEAR OPERATOR

Theorem 0.29.3. (Inverse operator) Let X,Y be vector spaces, both real or both complex. Let
T : D(T ) → Y be a linear operator with domain D(T ) ⊂ X and range R(T ) ⊂ Y . Then
(a) the inverse T −1 : R(T ) → D(T ) exists iff T x = 0 =⇒ x = 0.
(b) if T −1 exists, it is a linear operator.
(c) If dim D(T ) = n < ∞ and T −1 exists, then dim R(T ) = dim D(T ).

Proof. (a) Suppose that T x = 0 =⇒ x = 0. Let T x1 = T x2 . Since T is linear, T (x1 − x2 ) =

T x1 − T x2 = 0, so that x1 − x2 = 0 or x1 = x2 . So T is injective and T −1 exists.
Conversely, if T −1 exists, then T x1 = T x2 =⇒ x1 = x2 . Therefore, by taking x2 = 0, we
have T x1 = T 0 = 0 =⇒ x1 = 0.
(b) Assume T −1 exists, domain of T −1 is a R(T ) which is a vector space. Consider x1 , x2 ∈
D(T ), y1 = T x1 and y2 = T x2 . Then x1 = T −1 y1 and x2 = T 1 y2 .
Since T is linear, so that for any scalars α and β we have αy1 + β y2 = αT x1 + β T x2 =
T (αx1 + β x2 ). Therefore T −1 (αy1 + β y2 ) = αx1 + β x2 = αT −1 y1 + β T −1 y2 . Hence T −1 is
linear.
(c) We know dim R(T ) ≤ dim D(T ) and dim D(T ) ≤ dim R(T ) by the same theorem applied
to T −1 .

92
0.30. BOUNDED LINEAR TRANSFORMATION-I

0.30 Bounded Linear Transformation-I

Definition 0.30.1. Let N and N1 be normed linear spaces with same scalars and T : N → N1 is
a linear transformation.
T is said to be continuous iff for any sequence (xn ) in N converging to x ∈ N, the sequence
(T xn ) in N1 converges to T (x) ∈ N1 .
If there exists a real number k ≥ 0 such that kT (x)k ≤ k kxk for every x ∈ N, then k is called
a bound for T and such a T is called a bounded linear transformation.
Theorem 0.30.1. Let T be a linear transformation of a normed linear space N into another
normed linear space N1 . Then the following statements are equivalent to one another.
(i) T is continuous
(ii) T is continuous at origin, xn → 0, then T (xn ) → 0
(iii) There exists a real number k ≥ 0 such that kT xk ≤ k kxk for all x ∈ N, that is T is
bounded.
(iv) If S = {x : kxk ≤ 1} is the closed unit sphere in N, then its image is bounded set in N1 .

Proof. (i) ⇔ (ii) :

Let T be continuous and (xn ) be a sequence in N such that xn → 0. Then by continuity of
T , for xn → 0, we have T xn → T (0) = 0.
Conversely, Let T be continuous at origin and (xn ) be a sequence in N such that xn → x ∈ N.
Then xn −x → 0, T (xn −x) → 0, so T (xn )−T (x) → 0, that is, T xn = T x. Hence T is continuous
mapping.
(ii) ⇔ (iii): Let T be continuous at origin and suppose that T is not bounded, that is,
there exists no real number k such that kT xk ≤ k kxk for every x ∈ N. Then for each positive
integer n, we can find a vector xn such that kT xn k > n kxn k or kT xn k T xn
nkxn k > 1 or nkxn k > 1 or
 
xn
T nkx nk
> 1.
xn xn kxn k
Take yn = nkx nk
. kyn k = nkxnk
= nkx nk
= 1n → 0 and so yn → 0. But T (yn ) 6→ 0, since
kTyn k > 1. Hence T is not continous at origin which is contradiction. Hence T must be
bounded.
Conversely, suppose that T is bounded such that there exists a real number k ≥ 0 such that
kT xk ≤ k kxk for all x ∈ N.
Let (xn ) be any sequence in N such that xn → 0. Then kxn k → k0k = 0. Also, kT xn k ≤
k kxn k for each n. Therefore kT (xn )k → k0k = 0 or T (xn ) → 0. Hence the result.
(iii) ⇔ (iv): suppose that T is bounded such that there exists a real number k ≥ 0 such
that kT xk ≤ k kxk for all x ∈ N. Assume that x is any point of the closed unit sphere S so that
kxk ≤ 1. Then kT xk ≤ k for all x ∈ S. It follows that T [S] is bounded set in N1 .
Conversly, suppose that T [S] is bounded set in S so there exists a real number k ≥ 0 such
that kT xk ≤ k for all x ∈ S.
If x = 0, then T (x) = 0 and so kT xk ≤ k kxk. If x 6= 0, then x/ kxk ∈ S [as kx/ kxkk = 1],
therefore, kT (x/ kxk)k ≤ k or (1/ kxk) kT (x)k ≤ k or kT xk ≤ k kxk. Hence the result.

Theorem 0.30.2. Let B(N, N1 ) be the set of all bounded (or continuous) linear transformations
of normed linear space N into normed linear space N1 . Then B(N, N1 ) is a normed linear space
with respect to pointwise linear operations (T + U)(x) = T (x) + U(x) and (αT )(x) = αT (x)
and norm defined by kT k = sup{kT (x)k : x ∈ N, kxk ≤ 1}. Further, if N1 is a Banach space,
then B(N, N1 ) is also a Banach space.

93
0.30. BOUNDED LINEAR TRANSFORMATION-I

Proof. We know that the set S of all linear transformations from a linear space to linear space
is also a linear space with pointwise linear operations. So to prove that B(N, N1 ) is linear it is
sufficient to prove that that B(N, N1 ) is a subspace of S.
Let T,U ∈ B(N, N1 ). Then T and U are bounded and so there exist real number k and l
such that kT (x)k ≤ k kxk and kU(x)k ≤ l kxk for all x ∈ N.
If α, β are scalars, then
k(αT + βU)(x)k = k(αT )(x) + (βU)(x)k
= kαT (x) + βU(x)k
≤ kαT (x)k + kβU(x)k
= |α| kT (x)k + |β | kU(x)k
≤ |α|k k(x)k + |β |l k(x)k
= [|α|k + |β |l] k(x)k .
Thus αT + βU ∈ B(N, N1 ). Thus B(N, N1 ) is a linear subspace of S.
Now, to prove that B(N, N1 ) is a normed linear space.
[n1] Since kT k = sup{kT (x)k : x ∈ N, kxk ≤ 1} and kT (x)k ≥ 0, thus kT k ≥ 0.
[n2] kT k = sup{kT (x)k : x ∈ N, kxk ≤ 1} = sup{kT (x)k / kxk : x ∈ N, x 6= 0}.
kT k = 0, =⇒ sup{kT (x)k / kxk : x ∈ N, x 6= 0} = 0 ⇔ kT (x)k = 0 or T = 0.
[n3] If α is any scalar, then kαT k = sup{k(αT )(x)k : x ∈ N, kxk ≤ 1} = sup{|α| kT (x)k :
x ∈ N, kxk ≤ 1} = |α| kT k.
[n4] If T,U ∈ B(N.N1 ), then
kT +Uk = sup{kT (x) +U(x)k : x ∈ N, kxk ≤ 1}
≤ sup{kT (x)k : x ∈ N, kxk ≤ 1} + sup{kU(x)k : x ∈ N, kxk ≤ 1}
= kT k + kUk
Hence B(N, N1 ) is normed linear space.
Now, to prove that B(N, N1 ) is complete if N1 is complete. Suppose that N1 is complete
and (Tn ) be a Cauchy sequence in B(N, N1 ). Then kTm − Tn k → 0. For each x ∈ N, we have
kTm (x) − Tn (x)k = k(Tm − Tn )(x)k ≤ kTm − Tn k kxk → 0. Hence (Tn ) is a Cauchy sequence in
N1 for each x ∈ N. Since N1 is complete so there exists a vector T (x) ∈ N1 such that Tn (x) →
T (x).
First we show that T is linear and bounded.
T (αx + β y) = lim Tn (αx + β y) = lim[αTn (x) + β Tn (y)] = αT (x) + β T (y).
Consider
kT (x)k = klim Tn (x)k = lim kTn (x)k ≤ lim kTn k kxk
≤ sup{kTn k kxk}
= sup{kTn k} kxk .
Also, we know that | kTm k − kTn k | ≤ kTm − Tn k → 0. Therefore (kTn k) is a Cauchy se-
quence of real numbers and hence convergent and bounded. So there exists k ≥ 0 such that
sup{kTn k} ≤ k. Hence kT (x)k ≤ k kxk. Thus T is bounded. Therefore T ∈ B(N, N1 ).
Finally, we have to show that Tn → T . Consider
kTn (x) − T (x)k = kTn (x) − Tm (x) + Tm (x) − T (x)k
≤ kTn (x) − Tm (x)k + kTm (x) − T (x)k
< ε/2 + ε/2 = ε.

94
0.30. BOUNDED LINEAR TRANSFORMATION-I

Hence the result.

Theorem 0.30.3. Let T be a linear transformation from a normed linear space N to normed
linear space N1 . Then T is continuous at everypoint of N.

Proof. Let x1 and x2 be any two points of N. Suppose T is continous at x1 . Then for each
ε > 0, there exists δ > 0 such that kx − x1 k < δ , we have kT (x) − T (x1 )k < ε.
Now, for kx − x2 k < δ , k(x + x1 − x2 ) − x1 k < δ we have

kT (x + x1 − x2 ) − T (x1 )k < ε.

By the linearity of T , we get

kT (x) + T (x1 ) − T (x2 ) − T (x1 )k < ε,

that is kT (x) − T (x2 )k < ε. This shows that T is continuous at x2 as well. Since x1 and x2
are arbitrary points, we have shown that if T is continuous at a particular point, then it is
continuous at all points.

95
0.31. BOUNDED AND CONTINUOUS LINEAR OPERATORS-2

0.31 Bounded and Continuous Linear Operators-2

Definition 0.31.1. Let X and Y be normed spaces and T : D(T ) → Y a linear operator, where
D(T ) ⊂ X. The operator T is said to be bounded if there is a real number c such that for all
x ∈ D(T ), kT xk ≤ c kxk.
What is the smallest possible c such that kT xk
kxk ≤ c still holds for all non zero x ∈ D(T ).
The answer to our question is that the smallest possible c is that supremum, that is
kT xk
kT k = sup .
x∈D(T ),x6=0 kxk

kT k is called the norm of the operator or induced operator norm. If D(T ) = {0}, we define
kT k = 0; in this (relatively uninteresting) case, T = 0 since T 0 = 0.
Note that in the preceding definition (with c = kT k) implies that kT xk ≤ kT k kxk.
Note: kT k = supx∈D(T ),x6=0 (kT xk / kxk) minimum value of c, Norm of bounded linear
operator.
If T = 0, then kT k = 0.
Integral operator T : C[0, 1] → C[0, 1], x 7→ y = T x where y(t) = 01 k(t, τ)x(τ) dτ, where
R

k is continuous on the closed region I × I in t − τ plane.

We claim T is linear and bounded. Linearity follows easily, T (αx+β y) = 01 k(t, τ)[αx(τ)+
R

β y(τ)] dτ = αT (x) + β Ty.

Let norm on C[0, 1] defined as kxk = maxt∈I |x(t)| T is bounded, kT xk = kyk = maxt∈I |y(t)| =
maxt∈I | 01 k(t, τ)x(τ) dτ| ≤ 01 |k(t, τ)||x(τ)| dτ. SInce K is continuous on closed region I ×I,
R R

so K is bounded, absK(t, τ) ≤ M. Hence kT xk ≤ M kxk. Hence T is a bounded linear oper-

ator.
Lemma 0.31.1. Let T be a bounded linear operator defined as kT xk ≤ kT k kxk for all x ∈ D(T ).
(1) Alternative formula for the norm of T is

kT k = sup kT xk.
x∈D(T ),x=1

(2) The function defined by

kT xk
kT k = sup ,
x∈D(T ),x6=0 kxk

satisfies all the properties of norm.

Proof. (1) Write kxk = a and set y = (l/a)x, where x 6= O. Then kyk = kxk /a = 1, and since
T is linear, we have
kT xk x
kT k = sup = sup T( ) = sup kTyk.
x∈D(T ),x6=0 a x∈D(T ),x6=0 a y∈D(T ),y=1

Writing x for y on the right, we have the result.

(2) By definition kT k ≥ 0 and k0k = 0. From kT k = 0, we have T x = 0 for all x ∈ D(T ),
so that T = 0. Hence first two properties of norm holds. Further,

sup kαT xk = sup |α| kT xk = |α| sup kT xk .

kxk=1 kxk=1 kxk=1

96
0.31. BOUNDED AND CONTINUOUS LINEAR OPERATORS-2

Finally,
sup k(T1 + T2 )xk = sup k(T1 x + T2 x)k ≤ sup kT1 xk + sup kT2 xk .
kxk=1 kxk=1 kxk=1 kxk=1

Example 0.31.2. 1. (Identity operator) The identity operator I : X → X on a normed space

X 6= {0} is bounded and has norm kIk = 1.
2. Zero operator. The zero operator 0 : X → Y on a normed space X is bounded and has
norm k0k = 0.
3. Differentiation operator is unbounded. Let X be the normed space of all polynomials on
I = [0, 1] with norm defined as kxk = max|x(t)|, t ∈ I, A differentiation operator T is
defined as T x(t) = x0 (t). This operator is linear but not bounded. Indeed, let xn (t) = t n ,
where n ∈ N. Then kxn k = 1 and T xn (t) = xn0 (t) = nt n−1 so that kT xn k = n and kT xn k
kxn k = n.
kT xn k
Since n ∈ N is arbitrary, this shows that there is no fixed nmnber c such that kxn k ≤ c.
Hence, we conclude that T is not bounded.
4. Integral operator. We can define an integral operator T : C[0, 1] → C[0, 1] by y = T x
where y(t) = 01 k(t, τ)x(τ) dτ.
R

Here k is a given function, which is called the kernel of T and is assumed to be continuous
on the closed square G = I × I in the tτ -plane, where I = [0, 1]. The operator is linear.
The continuity of k on the closed square implies that k is bounded, say, |k(t, τ)| ≤ k0 for
all (t, τ)(r, T ) ∈ G, where k0 is a real number. Furthermore, |x(t)| ≤ max|x(t)| = kxk .
Hence

kT xk = kyk
Z 1
= max| k(t, τ)x(τ) dτ|
t∈I 0
Z 1
≤ max| |k(t, τ)||x(τ)| dτ|
t∈I 0
≤ k0 kxk .
Hence T is bounded.
Theorem 0.31.3. If a normed space X is finite dimensional, then every linear operator on X is
bounded.
Proof. Let dim X = n and {e1 , ..., en } a basis for X. We take any x = ∑ni=1 αi ei and consider
any linear operator T on X. Since T is linear,
n n
kT xk = ∑ αiTei ≤ ∑|αi | kTei k ≤ max kTek k ∑|αi | = k ∑|αi |,
k 1 1

where k = maxk=1,...,n kTek k. Also using Lemma ??, we have

1 1
∑|αi| ≤ c kαieik = c kxk .
Hence kT xk ≤ kc kxk. Therefore T is bounded.

97
0.31. BOUNDED AND CONTINUOUS LINEAR OPERATORS-2

xi
Proposition 0.31.4. Define an operator `∞ → ∞ by T (xi ) = i for each (x) ∈ `∞ . Show that T
is linear and bounded.
Proof. Let (xi ), (yi ) ∈ `∞ . Then
αxi + β yi
T (αxi + β yi ) =
i
xi yi
= α +β
i i
= αT (xi ) + β T (yi ).
Hence T is linear.
Now, to show that T is bounded. Suppose that (x 6= 0) ∈ `∞ . Then
xi
kT (xi )k =
i
|xi |
= sup
i i
≤ sup|xi | = kxi k .
i

Hence T is bounded.
into
Lemma 0.31.5. Let T : D(T ) → Y be a bounded linear operator. Then
(a) (xn ) ⊂ D(T ), x ∈ D(T ) and xn → x then T xn → T x.
(b) The null space N(T ) is closed.

Proof. (a) Since kT xn − T xk = kT (xn − x)k ≤ kT k kxn − xk → 0.

(b) Let x ∈ N(T ). Then there exists (xn ) ∈ N(T ) such that xn → x. Since xn ∈ N(T ), so
T xn = 0 for all n Since xn → x and T is bounded linear opeartor, so T xn → T x but T xn = 0 so
T x = 0 hence x ∈ N(T ), that is N(T ) is closed.
Let T1 : Y → Z and T2 : X → Y be bounded linear opeartors and X,Y, Z are normed
spaces. T : X → X be bounded linear operator. T1 ◦ T2 : X → Z as D(T1 ) ⊃ R(T2 ). It can be
shown that T1 ◦ T2 is bounded linear operator and kT1 T2 k ≤ kT1 k kT2 k.
It is easy to prove that T1 ◦ T2 is a linear operator.
T1 ◦ T2 (αx + β y) = T1 (T2 (αx + β y)) = T1 (αT2 x + β T2 y) = αT1 T2 x + β T1 T2 y.
kT1 T2 k = sup kT1 T2 xk / kxk ≤ kT1 k kT2 k
x∈D(T ),x6=0
Particluar case: T : X → X is bounded linear operator on normed linear space X. Then
kT n k ≤ kT kn n ∈ N.

Theorem 0.31.6. Let T be a linear transformation from a normed linear space N to normed
linear space N1 . Then T is bounded iff it is continuous.

Proof. Let T be continuous and suppose that T is not bounded, that is, there exists no real
number k such that kT xk ≤ k kxk for every x ∈ N. Then for each positiveinteger n, we can find
kT xn k T xn xn
a vector xn such that kT xn k > n kxn k or nkxn k > 1 or nkx nk
> 1 or T nkx nk
> 1.
xn xn kxn k
Take yn = nkx nk
. kyn k = nkx nk
= nkx nk
= 1n → 0 and so yn → 0. Since kTyn k > 1, so
T (yn ) 6→ 0. Hence T is not continous at 0, which is contradiction. Hence T must be bounded.

98
0.31. BOUNDED AND CONTINUOUS LINEAR OPERATORS-2

Conversely, suppose that T is bounded such that there exists a real number k ≥ 0 such that
kT xk ≤ k kxk for all x ∈ N.
Now, we have to show that T is continuous. Let x ∈ N be arbitrary. For any ε > 0,
choose δ = ε/k. Then for all y ∈ N such that ky − xk < δ , we get kTy − T xk = kT (y − x)k ≤
k ky − xk < k(ε/k) = ε. Hence T is continuous at x. Since x is an arbitrary, so T is a continuous
mapping.

99
0.32. EXTENSION OF LINEAR OPERATOR

X Y
T

Figure 2: Extension\ Restriction of Mapping

0.32 Extension of Linear Operator

Definition 0.32.1. (Equality of Operators). Two operators T, S are said to be equal, if they have
same domain and we write T = S if D(T ) = D(S) where D(T ) and D(S) are the domains of T
and S respectively, that is, T x = Sx; for all x ∈ D(T ) = D(S).
Let X and Y be two normed linear spaces and B be a subspace of X. Suppose that
S : B → Y and T : X → Y are two linear mappings such that S(x) = T (x), for all x ∈ B. See
the following figure:
Then we say, T is called an extension of the linear map S and S is called restriction of
T . If moreover the extension T is such that, ||T || = ||S|| then T is called an extension of S
with the preservation of norm (or norm preserving extension).
into
Definition 0.32.2. (Restriction). Let T : D(T ) → Y and B ⊂ D(T ). We define the restriction
into
of T to B, denoted T |B : B → Y , as T |B x = T x, for all x ∈ B.
into into
Definition 0.32.3. Let T : D(T ) → Y and D(T ) ⊂ M. An extension of T is T̂ : M → Y such
that T̂ |D(T ) = T and T̂ x = T x, for all x ∈ D(T ).
If D(T ) is a proper subset of X, then a given operator T has many extensions. Of prac-
tical interest there are usually those extensions which preserve some basic property, for
instances linearity (if T happens to be linear) or boundedness (if D(T ) lies in a normed
linear space and T is bounded). The following important theorem is typical in that re-
spect. If concerns an extension of a bounded linear operator T to the closure D(T ) of the
domain such that the extended operator is again bounded and linear, and even has the
same norm. This includes the case of an extension from a dense set in a normed linear
space X to all of X. It also includes the case of an extension from a normed linear space
X to its completion.

Theorem 0.32.1. (Bounded linear Extension) Let T : D(T ) → Y be a bounded linear operator,
where D(T ) ⊂ X, X a normed space and Y is a Banach space. Then T has an extension
T̂ : D(T ) → Y where T̂ is bounded linear operator of norm T̂ = kT k .

Proof. Let x ∈ D(T ). Then there exists a (xn ) ∈ D(T ) such that xn → x. Since T is linear and
bounded, we obtain

kT xn − T xm k = kT (xn − xm )k ≤ kT k kxn − xm k .

Therefore (T xn ) is a Cauchy sequence because (xn ) is convergent. Since Y is a Banach space,

(T xn ) converges, T xn → y ∈ Y. Define T̂ by T̂ x = y. Due to uniqueness of limits, T̂ is uniquely
defined for all x ∈ D(T ).

100
0.32. EXTENSION OF LINEAR OPERATOR

Clearly, T̂ is linear and T̂ x = T x for every x ∈ D(T ), so that T̂ is an extension of T . We

now use kT xn k ≤ kT k kxn k. Then T xn → y = T̂ x. Since x 7→ kxk defines a continuous mapping,
we obtain T̂ x ≤ kT k kxk. Hence T̂ is bounded and T̂ ≤ kT k. Also T̂ ≥ kT k because
the norm, being defined by a supremum, cannot decrease in an extension. Therefore, T̂ =
kT k.

101
0.33. LINEAR FUNCTIONALS

0.33 Linear Functionals

A functional is an operator whose range lies on the real line R or in the complex plane C.
Functional analysis was initially the analysis of functionals.
Bounded linear functional is a bounded linear operator with domain in normed linear
space and range in the vector space or normed space of sclars field K of X.
(X, +, ., k.k) normed space f : X → K(RorC), X is normed space
Bounded linear functional f : X → K = RorC, if f is a linear functional, that is

f (αx + β y) = α f (x) + β f (y), x, y ∈ X, α, β ∈ K

and there exists c > 0 such that

| f (x)| ≤ c kxk
for all x ∈ D( f ).
| f (x)|
kfk = sup .
x∈D( f ),x6=0 kxk

kfk = sup | f (x)|

x∈D( f ),||x||≤1

kfk = sup | f (x)|

x∈D( f ),||x||=1

Hence | f (x)| ≤ k f k kxk x ∈ D( f ).

Theorem 0.33.1. Let, X be a normed linear space over a field K, f : X → K be a linear func-
tional on X. Then f is continuous if and only if ker( f ) is closed in X.

Proof. Suppose that f is continuous. ker f = {x ∈ X : f (x) = 0} = f −1 {0} = a closed set in

X.
Now, suppose that ker( f ) is closed in X. Assume that f is a zero functional on X, i.e.
f (x) = 0, for all x ∈ X. So, f is continuous in this case.
Assume that f is a non zero functional. But X\ ker f is non mepty and open in X. Let
a ∈ X\ ker f . So, f (a) 6= 0. Take b = a/ f (a). So f (b) = 1 and b ∈ X\ ker f .
As X\ ker f is open, there exists a r > 0 such that Br (b) ⊂ X\ ker f .
Take V = {x ∈ X : | f (x)| < 1}. Now, we have to prove that Br (θ ) ⊂ V , where θ is a zero
vector of X. Let x ∈ Br (θ ). If possible let, x ∈
/ V . So | f (x)| ≥ 1.
Take y = −x/| f (x)|. So f (y) = −1 and ||y|| = ||x||/| f (x)| ≤ ||x|| < r. So y ∈ Br (θ ).
Now, ||b + y − b|| = ||y|| < r, that is b + y ∈ Br (b). Again, f (b + y) = f (b) + f (y) =
1 + (−1) = 0. So, b + y ∈ ker f . So b + y ∈ (ker f ) ∩ Br (b), which is a contradiction to the fact
that Br (b) ⊂ X\ ker f . So Br (θ ) ⊂ V .
1 rx
Choose x 6= θ such that x ∈ Br (θ ). So ||x|| < r. Now, || 2||x|| rx|| = r/2 < r. So 2||x|| ∈
Br (θ ) ⊂ V . Hence

102
0.33. LINEAR FUNCTIONALS

rx
|f( )| < 1
2||x||
r
| f (x)| < 1
2||x||
r
| f (x)| < 1
2||x||
2||x||
| f (x)| < .
r
So, f is bounded, implying that f is continuous.
Example 0.33.2. dot product f : Rn → R, f (x) = x.a, where
x = (x1 , x2 , x3 , . . . , xn ) ∈ Rn and a = (a1 , a2 , a3 , . . . , an ) ∈ Rn . Then f (x) is a bounded linear
functional and k f k = kak.
We have | f (x)| = |x.a| ≤ kxk kak.
k f k = sup{| f (x)| : x ∈ X kxk ≤ 1} ≤ kak
But k f k ≥ | fkak
(a)|
= |a.a|/ kak = kak2 / kak = kak. Hence k f k = kak.

Example 0.33.3. Definite Integral:

Let X = C[a, b], space of all continuous real valued functions over [a, b] with the norm
kxk = supa≤t≤b |x(t)|. Define f : X → R by f (x) = ab x(t)dt, x(t) ∈ C[a, b] is a bounded linear
R

functional, with k f k = b − a.
It is easy to see that f is linear.
Also k f k = sup| f (x)| = sup| ab x(t)dt| ≤ sup |x(t)|(b−a) ≤ kxk (b−a). Hence f is bounded
R

and k f k ≤ b − a.
Let x = x0 such that kx0 k = 1, | f (x0 )| = | ab dt| = (b − a). So k f k ≥ | fkx(x0k)| = b − a. Hence
R
0
k f k = b − a, length of the interval.
Example 0.33.4. X = `2 , p f (x) = ∑∞
ipξi αi is a bounded linear functional on `2 because | f (x)| =
∞ ∞ ∞ 2
| ∑i ξi αi | ≤ ∑i |ξi ||αi | ≤ ∑i |ξi | ∑∞ 2
i |αi | ≤ c kxk [By Cauchy Schwarz inequality].

Example 0.33.5. Unbounded linear functional equation: Take X = `∞ with norm defined by
kxk = supi |ξi |.
Here x ∈ (ξ1 , . . . , ξn , . . .)`∞ and ξi are scalars. Define f : X → R by f (x) = ∑ ξi . Then f is
unbounded linear functional.
Clearly f is linear. Choose the sequence {xn } ∈ `∞ where xn = ∑ni=1 ei , where ei = (0, 0, . . . , 1 , 0, . . .), i =
ithplace
1, 2, . . . . For each natural number n, note that ||xn || = || ∑ni=1 ei || = 1 and f (xn ) = n. Thus we
see that for each natural number n, | f (xn )| = n||xn ||. So it follows that || f || = ∞ and hence f is
an unbounded linear functional on `∞ .
The following result is required for the next theorem
Lemma 0.33.6. Let f be a non zero linear functional on a linear space X and let x0 ∈ X\ ker( f ).
Then any x ∈ X can be expressed uniquely in the form x = y + αx0 , where y ∈ ker( f ) and α is
scalar.
f (x) f (x)
Proof. Since x0 ∈ X\ ker( f ), f (x0 ) 6= 0. Choose α = f (x0 ) , x ∈ X and y = x − f (x0 ) x0 . Then
f (x) f (x)
x = y + αx0 . So f (y) = f (x − f (x0 ) x0 ) = f (x) − f (x0 ) f (x0 ) = 0. Hence y ∈ ker f .

103
0.33. LINEAR FUNCTIONALS

For uniqueness, take x = y1 + α1 x0 and x = y2 + α2 x0 for some y1 , y2 ∈ ker f (y1 6= y2 ) and

α1 , α2 are scalars.
If α1 = α2 , then y1 = y2 .
If α1 6= α2 , then x0 = αy21 −y
−α2 ∈ ker f , contradiction to our assumption x0 ∈ X\ ker f . Hence
1

the result.
Theorem 0.33.7. Let X a normed linear space and f be a linear functional on X. Then the set
ker( f ) is either dense or closed in X.

Proof. Let us suppose that ker( f ) is not closed. Then there exists a point x0 ∈/ ker( f ) such that
x0 is a limit point of ker( f ). Since ker( f ) is a subspace of X, it follows that ker( f ) is also a
subspace of X. Therefore ker( f ) contains the linear span of x0 and ker( f ). Using previous
lemma, we get, X ⊂ ker( f ). Hence ker( f ) = X.
Using Theorems 0.33.1 and 0.33.7, we have the following result.

Corollary 0.33.8. Let, X be a normed linear space and f be a linear functional on X. Then f
is bounded (continuous) if and only if ker( f ) is not dense in X.

104
0.34. DUAL SPACE

0.34 Dual Space

Definition 0.34.1. (B(X;Y )). Let X;Y be vector spaces. The set of all linear operators from X
to Y is itself a linear space with operations (αS + β T )x = αSx + β T x. We define B(X,Y ) as
the set of all bounded linear operators from X to Y .
Definition 0.34.2. (Dual space X 0 ). Let X be a normed space. Then the set of all bounded linear
functionals on X constitutes a normed space with norm defined by k f k = supx∈X,x6=0 | fkxk(x)|
=
supx∈X,kxk=1 | f (x)|.
This is also called normed dual of X.
Let (X, +, .) is a vector space and f , g, h, ... are defined on X called linear functionals
denoted by X ∗ .
X be a vector space and X ∗ be set of all linear functionals defined on X. We claim that
X ∗ forms a vector space under the operation ( f + g)x = f (x) + g(x) and (α f )(x) = α f (x)
for all x ∈ X. It is easy to verify these conditions. So X ∗ is a vector space. Also called
algebraic dual of X.
So defined linear functionals on the vector space X ∗ and denote this vector space by
X ∗∗ . Also called 2nd algebraic dual of X.
What is the need of second algebraic dual of X? What is relation betweein X and
∗∗
X ??
One can find g ∈ X ∗∗ which is a linear functional on X ∗ by choosing a fixed x ∈ X.
(How??)
Definition 0.34.3. For fixed x ∈ X define g( f ) = gx ( f ) = f (x), where f ∈ X ∗ variable. We
claim that gx is linear functional.

g(α f1 + β f2 ) = gx (α f1 + β f2 )(x)
= (α f1 + β f2 )(x)
= α f1 (x) + β f2 (x)
= αgx ( f1 ) + β gx ( f2 ).

So gx is linear functional. So gx ∈ X ∗∗ .
To each x ∈ X there corresponds a gx ∈ X ∗∗ . So we define a mapping C : X → X ∗∗ by x 7→ gx
, x ∈ X Then C is called Canonical mapping of X into X ∗∗ . We claim C is also linear.
D(C) is a vector space and R(C) ⊂ X ∗∗ lies in Vector space and (C(αx + β y))( f ) =
gαx+β y ( f ) = f (αx + β y) = α f x + β f y = αgx ( f ) + β gy ( f ) = α(Cx) f + β (Cy)( f ). So C is
linear.
C is also called Canonical embedding of X into X ∗∗ .
L J
Definition 0.34.4. Isomorphism: Let (X, +, .) and (Y, , ) be two vector spaces. An isom-
porphism
J L J
T : X → Y is a bijective mapping which preserves the operation T (α.x + β .y) =
α T x β Ty.
Similarly if (X, d) and (Y, d1 ) be two metric spaces then an isomorphic T : X → Y is a
bijective mapping which preserves the distances, that is d(T x, Ty) = d(x, y), x, y ∈ X.
Definition 0.34.5. If X is isomorphic with a subspace of a vector space Y then we say that X is
embeddable in Y .
Since C : X → X ∗∗ x 7→ gx is a canonical embedding of X into X ∗∗ .
Definition 0.34.6. If R(C) = X ∗∗ , then X is said to be algebraically reflexive.

105
0.34. DUAL SPACE

Finite dimensional vector spaces are simpler than infinite dimensional ones, and it is
natural to ask what simplification this entails with respect to linear operators and func-
tionals defined on such a space. Linear operators on finite dimensional vector spaces can
be represented in terms of matrices, as explained below. In this way, matrices become the
most important tools for studying linear operators in the finite dimensional case.
Let X and Y be finite dimensional vector space over the same field. Let T : X → Y be
a linear operator. dim X = n and dimY = r Basis E = {e1 , ..., en } for X and B = {b1 , ..., br }
basis for Y arranged in definite order which we keep fixed.
Let x ∈ X, y = T x ∈ Y and Tei ∈ Y for all i has a unique representation

x = ∑ ξi ei (0.34.1)

Since T is linear

y = T x = T (∑ ξi ei ) = ∑ ξi T (ei ) (0.34.2)

We can say that T is uniquely determined if the images yi = Tei of the n-basis vectors
e1 , ..., en are prescribed. Since y, Tei ∈ Y , so
r
y = ∑ η jb j
j

or
r
Tei = ∑ τ ji b j
j
.
Consider y = ∑rj=1 η j b j = ∑ni=1 ξi T (ei ) = ∑ ξi ∑ τ j b j = ∑i (∑ j τ ji ξi )b j .
This implies that

η j = ∑ τ ji ξi (0.34.3)
i

So we conclude that the image y = T x = ∑ η j b j of x = ∑i ξi ei can be obtained from equation

(0.34.3). Further, equation (0.34.3) suggest a matrix TEB = (τ ji )r×n .
So operator T can be represented by means of a matrix. Infact y = TEB x, where y =
(η1 , ...ηr ) and x = (ξ1 , ..., ξn ).
Linear Functional on X, dim X = n and (e1 , ..., en ) basis, then f (x) = f (∑ ξi ei ) = ∑ ξi f (ei ) =
∑ ξi αi , where αi = f (ei ). So f is uniquely determined by its values αi at the n basis ele-
ments.
Conversely, if n-scalars are given (1, 0, 0, 0...), (0, 1, 0, 0, ..), . . .
(
1, i = j
fi (e j ) = δi j =
0, i 6= j

Then this functional f1 , f2 , ..., fn is called dual basis of e1 , ..., en .

Theorem 0.34.1 (Dimension of X ∗ ). Let X be an n-dimensional vector space and E =

{e1 , ..., en } a basis for X. (
1, i = j
Then F = { f1 , ..., fn } given by fi (e j ) = δi j = is a basis for the algebraic dual
0, i 6= j
X ∗ of X, and dim X ∗ = dim X = n.

106
0.35. EXAMPLES OF DUAL SPACE

Proof. F is a linearly independent set since ∑nk=1 βk fk (x) = 0 for all x ∈ X, with x = e j gives
∑nk=1 βk fk (e j ) = ∑nk=1 βk δ jk = β j = 0, so that all β j are zero. Now, we show that every f ∈ X ∗
can represented as a linear combination of the elements of F in a unique way. We write
f (e j ) = α j . Thus f (x) = ∑nj=1 ξ j α j for every x ∈ X. On the other hand, we obtain f j (x) =
f j (∑ni=1 ξi ei ) = ξ j . Hence, we get f (x) = ∑nj=1 α j f j (x). Hence the unique representation of the
arbitrary linear functional f on X terms of the functionals f1 , ... fn is f = α1 f1 + ... + αn fn .

Lemma 0.34.2. (Zero vector) Let X be a finite dlmensional vector space. If x0 ∈ X has the
property that f (x0 ) = 0 for all f ∈ X ∗ , then x0 = O.

Proof. Let {e1 ; ...; en } be a basis for X and let x0 = ∑ni=1 ξ0i ei . Then f (x0 ) = ∑ni=1 ξ0i αi . By
assumption f (x0 ) = 0, for all f ∈ X ∗ . Hence ξ0i = 0 for each i. Hence the result.

Theorem 0.34.3. (Algebraic reflexivity). A finite dimensional vector space is algebraically

reflexive.

Proof. We know that canonical mapping C : X → X ∗∗ is a linear. Then Cx0 =⇒ (Cx0 )( f ) =

gx0 ( f ) = f (x0 ) = 0 for all f ∈ X ∗ . Using previous lemma (zero vector), we have x0 = 0. There-
fore, C has an inverse C−1 : R(C) → X such that dim R(C) = X. Therefore C is an isomorphism
and X is algebraically reflexive.
Example 0.34.4. Let { f1 ; f2 ; f3 } be the dual basis of {e1 ; e2 ; e3 } for R3 ; where e1 = (1; 1; 1);
e2 = (1; 1; −1); e3 = (1; −1; −1). Find f1 (x); f2 (x); f3 (x); where x = (1; 0; 0).
Solution. Write x = (1; 0; 0) in terms of e1 ; e2 ; e3 as (1, 0, 0) = a(1; 1; 1) + b(1; 1; −1) +
c(1; −1; −1), we obtain

a+b+c = 1
a+b−c = 0
a−b−c = 0

has the solution a = c = 1/2 and b = 0. So x = 21 e1 + 12 e3 . Now, f1 (x) = f1 ( 12 e1 + 12 e3 ) =

1 1 1 1 1
2 f 1 e1 + 2 f 1 e3 = 2 (1) + 2 (0) = 2 .
Similarly, we obtain f2 (x) = f2 ( 12 e1 + 12 e3 ) = 12 f2 e1 + 12 f2 e3 = 0 + 0 = 0 and f3 (x) = 12 .

0.35 Examples of Dual Space

Some examples of dual/conjugate space:
Example 0.35.1. The dual space of Rn is Rn .
Solution. If a normed space X is finite dimensional then every linear operator is bounded
on X. Thus we have Rn∗ = Rn0 . Now every f ∈ Rn∗ has a representation
n
f (x) = ∑ ξi αi , αi = f (ei ).
i=1

So by the Cauchy-Schwarz inequality, we get

s s s
n n n n
| f (x)| ≤ ∑ |ξi αi | ≤ ∑ |ξi|2 ∑ |αi|2 = kxk ∑ |αi|2.
i=1 i=1 i=1 i=1

107
0.35. EXAMPLES OF DUAL SPACE

p
Taking the supremum over all x of norm 1, we get k f k ≤ ∑ni=1 |αi |2 . Thus k f k ≤ kCk , ehere
C = (α1 , . . . , αn ) ∈ Rn . Thus for f ∈ X 0 , we get C = (α1 , . . . , αn ) ∈ Rn .
Conversely, C = (α1 , . . . , αn ) ∈ Rn . Define f : X → Rn by f (x) = f (x) = ∑ni=1 ξi αi , x =
∑ ξi ei , αi = f (ei ). So f is linear.
Now,

| f (x)| = | ∑ ξi αi |
≤ ∑ |ξi αi |
q q
≤ ∑ |ξi |2 ∑ |αi |2 = ||x||||C||.

Thus for C = (α1 , . . . , αn ) ∈ Rn , we get f ∈ X 0 .

If we choose x = (αi )ni=1 = C equality is achieved in the Cauchy-Schwarz inequality, that
is, ξi = αi , so we have

f (x) = ∑ αi2
k f (x)k
kfk ≥
kxk
s
n
= ∑ |αi|2
i=1
p
k f k = ∑ni=1 |αi |2 .
This proves that the norm of f is the Euclidean norm, and k f k = kCk, where C = (αi ) ∈ Rn .
Hence the mapping of Rn0 onto Rn defined by f 7→ C = (αi ), αi = f (ei ) is norm preserving,
linear and bijective, it is an isomorphism. Hence the result.
Example 0.35.2. The dual/conjugate space of ` p , (1 < p < ∞) is `q ; here, 1 < p < ∞ and q is
the conjugate of p, that is, l/p + l/q = 1.
Solution. A Schauder basis for ` p is (ek ), where ek = (δki ) has 1 in the kth place and
zeros otherwIse. Then every x = (ξ1 , . . .) ∈ ` p has a unique representation x = ∑∞
k=1 ξk ek and
1
∑ |ξi | p < ∞, kxk = (∑ |ξi | p ) p .
We consider any f ∈ `0p , where `0p is the dual space of ` p .
Since f is linear and bounded, f (x) = ∑∞ k=1 ξk γk , where γk = f (ek ) are uniquely determined
by f .
(n)
Let q be(the conjugate of p and consider xn = (ξk ) with
(n) |γk |q /γk , i f k ≤ n, γk 6= 0,
ξk =
0, otherwise.
(n)n q
Hence f (xn ) = ∑∞
k=1 ξk γk = ∑k=1 |γk | .
Also, we know (q − 1)p = q, we have,

f (xn ) ≤ k f k kxn k
(n)
= k f k (∑ |ξk | p )1/p
= k f k (∑|γk |(q−1)p )1/p
= k f k (∑|γk |q )1/p .

Thus we get f (xn ) = ∑nk=1 |γk |q ≤ k f k (∑|γk |q )1/p .

108
0.35. EXAMPLES OF DUAL SPACE

Therefore, we get (∑n1 |γk |q )1−1/p = (∑n1 |γk |q )1/q ≤ k f k .

Taking n → ∞, we get (∑∞ q 1/q ≤ k f k. Hence (γ ) ∈ ` .
1 |γk | ) k q
Conversely, for any b = (βk ) ∈ `q we can get a corresponding bounded linear functional g
on ` p . In fact, we may define g on ` p by setting g(x) = ∑∞ 1 ξk βk , where x = (ξk ) ∈ ` p . Then g
is linear, and boundedness follows from the Holder inequality. Hence g ∈ `0p .
Now, finally we have to prove that the norm of f is the norm on the space `q . From the
Holder inequality we have

| f (x)| = |∑ ξk γk |
≤ (∑ |ξk | p )1/p (∑ |γk |q )1/q
= kxk (∑ |γk |q )1/q .

Taking the supremum over all x of norm 1, we get k f k ≤ (∑ |γk |q )1/q .

Therefore, we get k f k = (∑∞ q 1/q .
1 |γk | )
Hence this formula can be written k f k = kckq , where c = (γ j ) ∈ `q . It shows that the
bijective linear mapping of `0p , onto `q defined by f 7→ c = (γ j ) is linear and bijective, and
norm preserving, so that it is an an isomorphism.
Example 0.35.3. The dual space of `1 is `∞ .
Solution. A Schauder basis for `1 is (ek ), where ek = (δki ) has 1 in the kth place and zeros
otherwIse. Then every x ∈ `1 has a unique representation x = ∑∞ k=1 ξk ek .
0 0
We consider any f ∈ `1 , where `1 is the dual space of `1 . Since f is linear and bounded,
f (x) = ∑∞ k=1 ξk γk , where γk = f (ek ) are uniquely determined by f . Also kek k = 1 and |γk | =
| f (ek )| ≤ k f k kek k = k f k, supk |γk | ≤ k f k. Hence (γk ) ∈ `∞ .
On the other hand, for every b = (βk ) ∈ `∞ we can obtain a corresponding bounded linear
functional g on `1 . In fact, we may define g on `1 by g(x) = ∑∞ k=1 ξk βk , where x = (ξk ) ∈ `1 .
Then g is linear and and boundedness follows from |g(x)| ≤ ∑∞ k=1 |ξk βk | ≤ supi |βi | ∑|ξk | =
supi |βi | kxk. Hence g ∈ `1 . 0

Finally, we have to show that norm of f is the norm on the space `∞ . We have | f (x)| =
|∑ ξk γk | ≤ supi |γi | ∑ |ξk | = supi |γi | kxk .
Taking the supremum over all x of norm 1, we get k f k ≤ supi |γi |.
Hence we get k f k = supi |γi |, which is the norm on `∞ . Hence this formula can be writ-
ten k f k = kck∞ , where c = (γ j ) ∈ `∞ . It shows that the bijective linear mapping of `01 , onto
`∞ defined by f 7→ c = (γ j ) is linear and bijective, and norm preserving, so that it is an an
isomorphism.

109
0.36. LINEAR FUNCTIONALS AND HAHN-BANACH THEOREM

0.36 Linear Functionals and Hahn-Banach Theorem

Now, our aim is study some special type of continuous linear transformations. R and C
are the simplest of all normed linear spaces. B(N, R) or B(N, C) denote respectively the set
of all continous linear transformations from N into R or C. We denote either of these sets
by N ∗ and call N ∗ by the conjugate space or adjoint space or dual space of N. Members
of N ∗ are called continous linear functionals or simply functionals. Thus a functional on
a normed linaer space N is a continuous linear transformations from N into R or C.
We know that R and C are Banach spaces with respect to norm k.k. Hence we may
consider linear functional f ∈ N ∗ as a linear transformation of the normed linear N into
the Banach space R or C. Thus if T is a linear transformation from a normed linear space
into R or C, then
(i) T is continuous iff it is bounded.
(ii) If T is continuous at x0 ∈ N, then it is continuous at every point of N.
Further, if the functionals on N are added and multiplied by scalars pointwise and the
norm of the functional f ∈ N ∗ defined by k f k = {sup | f (x) : x ∈ N, kxk ≤ 1} = inf{K : K ≥
0, k f (x)k ≤ K kxk , ∀x ∈ N}, then N ∗ is a Banach space. The conjugate space (N ∗ )∗ of N ∗ is
called the second conjugate space of N and denoted by N ∗∗ . Here N ∗∗ is also complete.
Two elements a and b are called comparable elements if they satisfy a ≤ b or b ≤ a
(or both). A totally ordered set or chain is a partially ordered set such that every two
elements of the set are comparable. In other words, a cham is a partially ordered set that
has no incomparable elements
Definition 0.36.1. Partially ordered set: A POS is a set M on which there is defined a partial
ordering, that is a binary relation denoted by ≤ which satisfy
(PO1) Reflexive a ≤ a for all a ∈ M
(PO2) Antisymmetery if a ≤ b and b ≤ a then a = b
(PO3) Transitive if a ≤ b, b ≤ c, then a ≤ c
Note: If M contains elements for which niether a ≤ b nor b ≤ a then a and b are called
incomparable elements.
Two elements a and b in M are said to be comparable if either a ≤ b or b ≤ a or both.
Definition 0.36.2. Totally oredred set or chain. It is a POS such that every two elements of the
set are comparable.
For example: M= set of Natural numbers defined as a ≤ (R)b if a < b, this is PO relation.
Also TOS.
Definition 0.36.3. Upper bound or maximal: An upper bound of a subset W of a POS M is an
element u ∈ M such that x ≤ u for every x ∈ W .
u may or may not exist on M and W .
A maximal element of M is m ∈ M such m ≤ x =⇒ m = x.
For example (1) M = set of all real number and x ≤ (R)yif x ≤ y Then M is TOS but M has
no maximal element.
(2) Let P(X) be the power set of given set X and A ≤ B if A ⊆ B. Then P(X)isPOS and
maximal element is X.
Lemma 0.36.1. Zorn’s Lemma. Let M 6= 0/ be a partially ordered set. Suppose that every chain
C ⊂ M has an upper bound. Then M has atleast one maximal element.

Theorem 0.36.2. Every vector space X 6= {0} has a Hamel basis.

110
0.36. LINEAR FUNCTIONALS AND HAHN-BANACH THEOREM

Proof. Let M be the set of all linearly independent subsets of X. Since X 6= {0} so there exists
x 6= 0 and {x} ∈ M, then M 6= {0}.
Define inclusion relation on M, that is A ≤ B if A ≤ B. So M is Partially Oordered set.
We claim every chain C ⊂ M has an upper bound because union of all subset s of C are
elements of C. By Zorn’s lemma, M has a maximal element say B. We claim B is a Hamel
basis for X. Suppose it is not true, that is span B ⊂ X so there is a z ∈ X − spanB such that
B ∪ {z} would be Linearly independent set containing B as proper subset, which contradicts
thatB is maximal element. So spanB = X. Therefore B is a Hamel basis for X.

Theorem 0.36.3. In every Hilbert space H 6= {0} there exists a total orthonormal set.

Proof. Let M be set of all orthonormal subsets of H. Since H 6= {0} so there exists x 6= 0 and
x
an orthonormal subset of H is {y}, where y = kxk , then M 6= {0}.
On M define set inclusion relation, this gives M a POS. Pick a chain (TOS) C ⊂ M. This
will have an upper bound because union of all subsets in C.
By Zorn’s lemma, M will have a maximal element say F. We shall prove F is total in
H. Suppose F is not total in H, so there exists a nonzero z ∈ H such that z ⊥ F. Hence
z
F1 = F ∪ {e}, where e = kzk , is a orthonormal set which contains F as a proper subset. SO F
can not be maximal element. Therefore a contradiction, and shows that F is total in H.
Remarks 0.36.1. Hahn Banach Theorem is an extension theorem for linear functional on vec-
tor spaces.
This guarantes the normed space dual.
Uniform Boundedness Theorem (Banach-Steinhaus Theorem): This theorem gives condi-
tion sufficient for sequence (kTn k) to be bounded where Tn : X → Y , where X is Banach space
Y is normed space and Tn is bounded linear operator.
It has lot of applications like Summability theory, Analysis of Fourier series, weak conver-
gence
Open Mapping Theorem: This theorem says that a bounded linear operator T : X → Y X,Y
are Banach space is an open mapping, that is T sends open set to open set. Hence T is bijective
and continuous. [T is continuous iff T −1 G is open, G is open]
Closed graph theorem: It gives a condition under which a closed linear operator is bounded.
The operator T : X → Y (X,Y is normed space) is closed linear operator if the graph of T
(G(T ) = D(T ) ⊂ X → Y ) is closed in the normed space X × X.

Theorem 0.36.4. Hahn Banach Theorem Let M be a linear subspace of a normed linear space
N and f be a functional defined on M. Then f can be extended to a functional F defined on
the whole space N such that kFk = k f k.

HBT is discovered by Hahn in 1927 and in modified form by S. Banach in 1929 and
later extended to complex vector space by H. F. Bhonenhust et al in 1938. HBT is an
extension of linear functional. In HBT, the object to be extended is a linear functional f
which define on subspace Z of a vector space X and has a certain boundedness property
in terms of sublinear functional.
Definition 0.36.4. A Sublinear functional p(x) is a real valued functional on a vector space X
which is subadditive, that is, p(x + y) ≤ p(x) + p(y) for all x, y ∈ X and positive homogeneous,
that is, p(αx) = α p(x) for all α ≥ 0 in R and x ∈ X.

111
0.36. LINEAR FUNCTIONALS AND HAHN-BANACH THEOREM

Theorem 0.36.5. Hahn-Banach Theorem (real vector space): Let X be a real vector space and
p a sublinear functional on X. Further, let f be a linear functional defined on subspace Z of X
and satisfies f (x) ≤ p(x), for all x ∈ Z.
Then f has a linear extension fˆ : Z → X satisfying fˆ(x) ≤ p(x) for all x ∈ X, that is fˆ is a
linear functional on X, satisfies fˆ(x) ≤ p(x) on X and fˆ(x) = f (x) for any x ∈ Z.

Proof. Step 1. The set E of all linear extensions g of f satisfying the condition g(x) ≤ p(x) on
their domain D(g) can be partially ordered and Zorn’s lemma yields a maximal element fˆ of
E.
Step 2. fˆ is defined on the entire space X.
Step 3. An auxiliary relation which is used is step 2.
Proof of Step 1. Let E be the set of all linear extensions g of f which satisfies the condition
g(x) ≤ p(x) for all x ∈ D(g).
Clearly E 6= 0/ because f ∈ E ( f is linear and f (x) ≤ p(x), x ∈ D( f )). On E, we define a
partial order as follows g ≤ h if h is an extension of g, where g, h ∈ E, that is D(h) ⊃ D(g) and
h(x) = g(x) for all x ∈ D(g).
For any chain(TOS) C ⊂ E and define ĝ ∈ C by ĝ(x) = g(x) whenever x ∈ D(g).
Clearly ĝ is a linear functional [ĝ(αx + β y) = g(αx + β y) = αg(x) + β g(y), if αx + β y ∈
D(g) and g is linear] and domain D(ĝ) = ∪g∈C D(g). Obviously D(ĝ) is a vector space, as C is
a chain.
We claim that ĝ is an upper bound of C. Because, x ∈ D(g1 ) ∩ D(g2 ) with g1 , g2 ∈ C,
g1 (x) = g2 (x). Since C is chain so g1 ≤ g2 or g2 ≤ g1 . Hence g ≤ ĝ for all g ∈ C. So ĝ is an
upper bound of C.
Since C is a chain which has an upper bound and C is arbitrary chain of E. So by Zorn’s
lemma E has a maximal element fˆ, say. By the definition of E, fˆ is a linear extension of f
which satisfies fˆ(x) ≤ p(x), x ∈ D( fˆ).
Step 2. Now, we have to show that D( fˆ) = X. Suppose D( fˆ) 6= X, so we can choose a
y1 ∈ X − D( fˆ). Cleraly, y1 6= 0 as 0 ∈ D( fˆ).
Consider a subspace Y1 of X spanned by D( fˆ) and y1 . So any x ∈ Y1 can be written as
x = y + αy1 , y ∈ D( fˆ). We claim that this representation is unique, That is x = y + αy1 =
y + β y1 , y, y ∈ D( fˆ), so y − y = (β − α)y1 . But LHS y − y ∈ D( fˆ) and RHS y1 ∈/ D( fˆ). So only
possibility that y − y = 0 and β − α = 0, that is y = y.
Introduce a functional g1 on Y1 as follows g1 (x) = g1 (y + αy1 ) = fˆ(y) + αc, y ∈ D( fˆ) and
c is a real number. Then g1 is linear [y 7→ ay0 + by00 , fˆ is linear].
Further, for α = 0, g1 (y) = fˆ(y), so g1 is a proper extension of fˆ, that is an extension such
that D( fˆ) is a proper subset of D(g1 ).
If we prove that g1 (x) ≤ p(x) for all x ∈ D(g1 ), then g1 ∈ E and hence so that D( fˆ) = X
will contradict the maximality of fˆ, and our assumption D( fˆ) 6= X is wrong.
Step 3. Consider y, z ∈ D( fˆ). Then as fˆ is linear, we have

fˆ(y) − fˆ(z) = fˆ(y − z)

≤ p(y − z)
= p(y + y1 − y1 − z)
≤ p(y + y1 ) + p(−y1 − z).

It implies that −p(−y1 −z)− fˆ(z) ≤ p(y+y1 )− fˆ(y), where y1 is fixed, y, z ∈ D( fˆ). Taking the
supremum of LHS and infimum of RHS, we get m0 = supz {−p(−y, −z) − fˆ(z)} ≤ infy {p(y +
y1 ) − fˆ(y)} = m1 , there exists c such that m0 ≤ c ≤ m1 , so −p(−y, −z) − fˆ(z) ≤ c z ∈ D( fˆ)
and c ≤ p(y + y1 ) − fˆ(y) for all y ∈ D( fˆ).

112
0.36. LINEAR FUNCTIONALS AND HAHN-BANACH THEOREM

Now, we have to prove that g1 (x) ≤ p(x) for all x ∈ D(g1 ), that is g(y + αy1 ) ≤ p(x). We
prove for cases when α > 0, α < 0 or α = 0.
Take α < 0. Replace z by α −1 y in −p(−y1 − z) − fˆ(z) ≤ c, we get −p(−y1 − α −1 y) −
ˆf (α −1 y) ≤ c. Multiply −α > 0, we get α[p(−y1 − α −1 y) + fˆ(α −1 y)] ≤ −αc, so −p(−αy1 −
y) ≥ fˆ(y) + αc Therefore p(x) = p(αy1 + y) ≥ g1 (x).
Take α = 0, result is obvious.
For α > 0, in a similar way, one can prove that g1 (x) ≤ fˆ(y) + αc ≤ p(x). Hence the result.

113
0.36. LINEAR FUNCTIONALS AND HAHN-BANACH THEOREM

Theorem 0.36.6. Generalized Form [Complex vector Space] Let X be a real or complex vector
space and p a real valued functional on X which is sub-additive, that is p(x + y) ≤ p(x) + p(y)
for all x, y ∈ X and for scalar α satisfies p(αx) = |α|p(x), x ∈ X.
Further, let f be a linear functional which is defined on a subspace Z of X and satisfying
| f (x)| ≤ p(x) for all x ∈ Z. Then f has a linear extension fˆ : Z → X satisfying | fˆ(x)| ≤ p(x)
for all x ∈ X.

Proof. (a) For real vector space: If X is real, then | f (x)| ≤ p(x) =⇒ f (x) ≤ p(x) for all x ∈ Z.
So by Hahn-Banach theorem (for real vector space), there is a linear extension fˆ from Z to X
such that f (x) ≤ p(x) for all x ∈ X.
Now to establish, | fˆ(x)| ≤ p(x). Consider − fˆ(x) ≤ fˆ(−x) ≤ p(−x) = | − 1|p(x) = p(x).
Thus | fˆ(x)| ≤ p(x)
(b) For complex vector space: Let X be a complex vector space. Then Z ⊂ X is also a
complex vector space. Hence f is complex valued linear functional defined on Z. So f (z) =
f1 (x) + i f2 (x), x ∈ Z and f1 , f2 are real valued.
Take X and Z as real vector spaces and denoted by Xr and Zr respectively, that is we restrict
multiplication by scalars to real numbers instead of complex numbers. Since f is linear on
Z, f1 , f2 are real valued and linear functionals defined on Zr . Further f1 (x) ≤ | f (x)|. So
f1 (x) ≤ p(x) for all x ∈ Zr . By HBT, there is a linear extension fˆ1 of f1 from Zr to Xr such that

fˆ1 (x) ≤ p(x), (0.36.1)

for all x ∈ Xr .
Returning Z and X, and f = f1 + i f2 is linear on Z. For all x ∈ Z, i[ f1 (x)+ i f2 (x)] = i f (x) =
f1 (ix) + i f2 (ix). Equating real and imaginary parts, we have

f2 (x) = − f1 (ix). (0.36.2)

Hence if for all x ∈ X, we set fˆ(x) = fˆ1 (x) − i fˆ1 (ix).

Clearly X is restricted to Z, that is, x ∈ Z then from (0.36.2), we have fˆ(x) = f (x) =
f1 (x) − i f1 (ix) on x ∈ Z. This shows that fˆ is an extension of f from Z to X.
It remained to show that (i) fˆ is linear functional on the complex vector space X and (ii) fˆ
satisfies | fˆ(x)| ≤ p(x) for all x ∈ X.
In order to prove fˆ is linear functional on the complex vector space X, consider

fˆ((a + ib)(x)) = fˆ1 (ax + ibx) − i f1 (iaxˆ− bx)

= a fˆ1 (x) + b fˆ1 (ix) − i[a fˆ1 (ix) − b fˆ1 (x)]
= (a + ib) f (x).

So fˆ is linear.
Now, we’ll prove that fˆ satisfies | fˆ(x)| ≤ p(x) for all x ∈ X.
ˆ = 0 for any x then | fˆ(x)| ≤ p(x) as p(x) ≥ 0 [p(x + y) ≤ p(x) + p(y), p(αx) =
If f (x)
|α|p(x) =⇒ p(x) ≥ 0].
Let x be such that f (x)ˆ 6= 0. So using polar form f (x) ˆ = | f (x)|e
ˆ iθ =⇒ | f (x)| ˆ = f (x)e
ˆ −iθ =
f (xeˆ−iθ ). Since | f (x)|
ˆ is real so f (xeˆ−iθ ) is also real. Hence f (xeˆ−iθ ) = f1 (xeˆ −iθ ) ≤ p(xe−iθ ) =
|e−iθ |p(x) = p(x) [|e−iθ | = 1]. Hence | fˆ(x)| ≤ p(x) for all x ∈ X. Hence the result.

114
0.36. LINEAR FUNCTIONALS AND HAHN-BANACH THEOREM

Another form of Hahn Banach Theorem (In normed space)

Theorem 0.36.7. Let f be a bounded linear functional on subspace Z of a normed space X,
then there exists a bounded linear functional fˆ on X which ia n extension of f to X and has the
same norm, that is

fˆ X
= k f kZ , (0.36.3)

where fˆ = sup ˆ and k f k =

| f (x)| sup | f (x)|.
X Z
x∈X,kxk=1 x∈Z,kxk=1

Proof. If Z = {0}, then f = 0 and its extension fˆ = 0.

If Z 6= {0}. For all x ∈ Z, we have | f (x)| ≤ k f kZ kxk. Choose p(x) = k f kZ kxk. Clearly p is
sublinear functional over (X, k.k) because p(x + y) = k f kZ kx + yk ≤ k f kZ kxk + k f kZ kyk =
p(x) + p(y)(by triangular inequality) and p(αx) = k f kZ kαxk = |α|p(x), for all x ∈ X.
So use Hahn Banach Theorem for real vector space, there exists a linear functional fˆ on X
ˆ ≤ p(x) = k f k kxk. This implies that
which is an extension of f and satisfies | f (x)| Z

ˆ
| f (x)|
sup ≤ k f kZ .
x∈X,kxk=1 kxk

So f (x)ˆ ≤ k f kZ . But Z ⊂ X so ˆ
f (x) ≥ k f kZ . Hence ˆ
f (x) = k f kZ . Hence the
X X X
result.
Remark 0.36.1. In case of Hilbert space.
If Z is closed subspace of a Hilbert space X = H, then every bounded linear functional f has
a Riesz Representation, that is, f (x) = hx, Zi, Z is uniquely determined by f and kZk = k f k.
Since the inner product is defined on all of H so this gives at once a linear extension fˆ of f
from Z to H and fˆ has the same norm as f because fˆ = kZk = k f k.

115
0.37. APPLICATION OF HAHN BANACH THEOREM

0.37 Application of Hahn Banach Theorem

In this section, we discuss some applications of Hahn Banach theorem.
Theorem 0.37.1. Bounded linear Functional: Let X be a normed linear space and x0 6= 0 be
any element of X. Then there exists a bounded linear functional fˆ on X such that fˆ = 1,
f (xˆ 0 ) = kx0 k.
Proof. Let x0 ∈ X. Consider a subspace Z of X consisting of all elements x = αx0 , α is scalar,
that is Z = [x0 ]= the subspace spanned by {x0 }.
On Z define a linear functional f by f (x) = f (αx0 ) = α kx0 k. f is bounded and has a norm
1 because
| f (x)| = | f (αx0 )| = |α| kx0 k = kαx0 k = kxk .
So k f k = sup | fkxk
(x)|
≤ 1. But k f k ≥ | fkxk
(x)|
= 1. So k f k = 1. So by the Hahn Banach Theorem in
normed space, we have f has a bounded linear extension fˆ from Z to X of norm fˆ = k f k = 1.
Therefore, fˆ(x0 ) = f (x0 ) = kx0 k . Hence the result.
Theorem 0.37.2. Let X be a normed linear space over a field K and x0 6= 0 be an arbitrary
member of X and M0 be an arbitrary positive real. Then there exists a bounded linear functional
f on X such that k f k = M0 and f (x0 ) = k f k kx0 k.
Proof. Let Z = [x0 ] = {tx0 : t ∈ R}. Clearly M is a subspace of X. Define φ on Z by φ (tx0 ) =
tM0 kx0 k. It is easy to see that φ is linear.
Now, for x ∈ Z, x = tx0 for some t. So we have
|φ (x)| = |φ (tx0 )| = |tM0 | kx0 k = M0 |t| kx0 k = M0 ktx0 k = M0 kxk .
So |φ (x)| = M0 kxk. Clearly kφ k = M0 . So φ is a bounded linear functional on Z. So there
exists an extension f of φ over X such that k f k = kφ k = M0 . Hence k f k = M0 . But φ (x0 ) =
M0 kx0 k. Now, for x0 ∈ Z, f (x0 ) = φ (x0 ) = M0 kx0 k = k f k kx0 k. Hence the result.
Theorem 0.37.3. Let X be a non trivial real normed linear space, that is X 6= {0}. Then its first
conjugate space X 0 is also non trivial.
Proof. Let x0 6= 0 ∈ X. So there exists a bounded linear functional say F on X such that
F(x0 ) = kx0 k 6= 0 and kFk = 1, that is F ∈ X 0 is a non zero bounded linear functional on
X.

Theorem 0.37.4. For every x ∈ X, kxk = sup | kf (x)|

f k , [Θ is a zero functional on X.]
f 6=Θ

Proof. We can find a non zero bounded functional f0 ∈ X 0 such that f0 (x) = kxk and k f0 k = 1.
Now,
| f (x)| | f0 (x)|
sup ≥ = kxk .
f 6=Θ k f k k f0 k
Again for all f 6= Θ ∈ X 0 , we have | f (x)| ≤ k f k kxk. Therefore,
| f (x)|
sup ≤ kxk .
f 6=Θ∈X 0 k f k

| f (x)|
Hence kxk = sup kfk
f 6=Θ∈X 0

116
0.37. APPLICATION OF HAHN BANACH THEOREM

Corollary 0.37.5. If f (x) = 0 for every f ∈ X 0 where f 6= Θ ∈ X 0 , then x = 0.

Proof. As kxk = sup f 6=Θ∈X 0 | kf (x)|

f k = 0, so x = 0.

Corollary 0.37.6. If for every f 6= Θ ∈ X 0 , f (x) = f (y), for x, y ∈ X then x = y.

Proof. Here f (x) = f (y), so f (x−y) = 0, for every f 6= Θ ∈ X 0 . Hence x−y = 0 and x = y.

Theorem 0.37.7. Let M be a closed subspace of a normed linear space X such that M 6= X.
Assume that u ∈ X\M and d = dist(u, M) = infm∈M ku − mk. Then there is a bounded linear
functional f ∈ X 0 such that
(i) f (x) = 0 for all x ∈ M.
(ii) f (u) = 1.
(iii) k f k = d1 .

Proof. Clearly, d > 0. Let N = [M ∪ {u}]. Clearly N is a subspace of X spanned by M and

{u}. So every member of N is of the form m + tu, where m ∈ M, t ∈ R.
Define g on N by g(m +tu) = t. Here g is linear. Now, g(m) = 0 for some m ∈ M, g(u) = 1.
For t 6= 0, (−m/t) ∈ M we have
|t| km + tuk km + tuk 1
|g(m + tu)| = |t| = = ≤ km + tuk .
km + tuk ku − (−m/t)k d

Let x = m + tu ∈ N. So |g(x)| ≤ d1 kxk implying that g is bounded and kgk ≤ 1/d. So

g ∈ N 0 . Again d = dist(u, M) = infm∈M ku − mk. So there exists a sequence (mn ) ∈ M such
that ku − mn k → d as n → ∞. Now, since g(u − mn ) = g(u) − g(mn ) = 1 − 0 = 1, we have
1 = |g(u − mn )| ≤ kgk ku − mn k. Taking n → ∞, 1 ≤ kgk d or d1 ≤ kgk. Hence we get kgk = d1 .
So there exists a bounded linear functional f ∈ X 0 which is an extension of g on N such that
f (x) = g(x) for all x ∈ N and k f k = kgk. Thus we have
(i) f (x) = 0 for all x ∈ N.
(ii) f (u) = 1
(iii) k f k = d1

Theorem 0.37.8. Let M be a subspace of a normed linear space X such that M 6= X. Assume
that u ∈ X\M such d = dist(u, M) = infm∈M ku − mk > 0. Then there is a bounded linear
functional f ∈ X 0 such that
(i) f (x) = 0 for all x ∈ M.
(ii) f (u) = d.
(iii) k f k = 1.

Proof. Clearly, d > 0. Let N = [M ∪ {u}]. Clearly N is a subspace of X spanned by M and

{u}. So every member of N is of the form m + tu, where m ∈ M, t ∈ R.
Define g on N by g(x = m + tu) = td. Here g is linear. Now, g(m) = 0 for some m ∈ M,
g(u) = d.
For t 6= 0, −m/t ∈ M we have

km + tuk = k−t((−m/t) − u)k = |t| k(−m/t) − uk ≥ |t|d.

So |g(m + tu)| = |t|d =≤ km + tuk .

117
0.37. APPLICATION OF HAHN BANACH THEOREM

Let x = m + tu ∈ N. So |g(x)| ≤ kxk implying that g is bounded and kgk ≤ 1. So g ∈ N 0 .

Again d = dist(u, M) = infm∈M ku − mk.
Let ε > 0. By infimum property there exists an element m ∈ N such that ku − mk < d + ε.
m−u m u 1 m
Choose v = km−uk = km−uk − km−uk ∈ N where t = km−uk and km−uk ∈ M. So kvk = 1.
d d
But |g(v)| = km−uk ≥ d+ε kvk , as kvk = 1.
As ε > 0 is arbitrary, |g(v)| ≥ kvk, so kgk ≥ 1. Hence we get kgk = 1.
So there exists a bounded linear functional f ∈ X 0 which is an extension of g on N such that
f (x) = g(x) for all x ∈ N and k f k = kgk. Thus we have
(i) f (x) = 0 for all x ∈ N.
(ii) f (u) = d
(iii) k f k = 1.
Example 0.37.9. Let X = R2 . Take (x, y) ∈ X. Define k(x, y)k = |x| + |y|.
Clearly X is a normed linear space with respect to norm defined.
Take M = {(x, 0) ∈ R2 , x ∈ R}, is a subspace of X.
Define f : M → R by f (x, 0) = x. Here f is a linear. Now, | f (x, 0)| = |x| = k(x, 0)k. So
k f k = 1. Clearly, f is bounded. So f is a bounded linear functional on M.
Define F1 : X → R and F2 : X → R by F1 (x, y) = x + y and F2 (x, y) = x − y for all x, y ∈ X.
Here F1 and F2 are linear.
Now, |F1 (x, y)| ≤ |x + y| ≤ |x| + |y| = k(x, y)k and |F2 (x, y)| ≤ |x − y| ≤ |x| + |y| = k(x, y)k.
SO F1 and F2 are bounded and kF1 k ≤ 1 and kF2 k ≤ 1, that is kF1 k ≤ k f k and kF2 k ≤ k f k.
Now, for all (x, 0) ∈ M, F1 (x, 0) = x = f (x, 0) and F2 (x, 0) = x = f (x, 0).
So F1 and F2 are extensions of f .
Further, for all (x, 0) ∈ M, F1 (x, 0) = f (x, 0). So | f (x, 0)| = |F1 (x, 0)| ≤ kF1 k k(x, 0)k. So
k f k ≤ kF1 k. Similarly, we get k f k ≤ kF2 k . So kF1 k = k f k = kF2 k.

118
0.38. UNIFORM BOUNDEDNESS THEOREM

0.38 Uniform Boundedness Theorem

Uniform Boundedness Theorem given by S. Banach and H. Steinhaus (1927), so it is also
called Banach-Steinhaus Theorem.
This theorem gives sufficient condition for sequence (kTn k) to be bounded where Tn :
X → Y , where X is Banach space Y is normed space and Tn is bounded linear operator.
It has lot of applications like Summability theory, Analysis of Fourier series, weak
convergence. We require Baire-Category theorem to prove uniform boundedness theo-
rem.
Definition 0.38.1. Category: A subset M of a metric space (X, d) is said to be
(a) rare (or nowhere dense) in (X, d) if its closure, M has no interior point.
(b) meager (or of first category) in X if M is the union of countably many sets each of which
is rare in (X, d).
For example: Set of rational numbers in R or in itself are of first category, because let
M =set of all rational numbers and Q is countable so M is countable. Each xi ∈ Q, M = ∪∞ i {xi },
each xi is rare in Q or M.
(c) Non-meager (or of second category) in X if M is not meager in X. for example ` p , `1
are nonmeager.
Theorem 0.38.1. Baire-category theorem(complete metric space). If a metric space X 6= 0/ is
complete, then it is nonmeager in itself.
Hence if X 6= 0/ is complete and X = ∪∞i=1 Ai (each Ai is closed), then at least one Ai contains
a nonempty open subset.
Proof. Suppoe that metric space X 6= 0/ is complete, and of first category in itself. Then
X = ∪∞
i=1 Mi ,

where each Mi is rare in X.

Now, we have to construct a Cauchy sequence (pk ) whose limit p (which exits by com-
pleteness) is not in Mk , and this contradicts X = ∪∞
i=1 Mi .
Since by assumption M1 is rare in X, so by definition, M1 does not contain a nonempty
open set. But (X, d) is complete, (so every Cauchy sequence is convergent) so it will contain
c
a nonempty set. Hence M1 6= X. So M1 = X − M1 which is nonempty and open. So, we may
c
choose a point p1 ∈ M1 , and a open ball about it such that
c
B1 = B(p1 , ε1 ) ⊂ M1 ,
where ε1 < 1/2.
Further, by assumption M2 is rare in X, so M2 does not contain a nonempty open set. Hence
c
it does not contain the open ball B(p1 , ε1 /2). This implies that B(p1 , ε1 /2) ∩ M2 is not empty
and open, so that we may choose an open ball in this set, say,
c
B2 = B(p2 , ε2 ) ⊂ M2 ∩ B(p1 , ε1 /2),
where ε2 < ε/2.
By induction we thus obtain a sequence of balls Bk = B(pk , εk ), where εk < 2−k such that
Bk ∩ Mk = 0/ and Bk+1 ⊂ B(pk , εk /2) ⊂ Bk , k = 1, 2, ....
Since εk < 2−k , the sequence (pk ) of the centers is Cauchy and converges, say, pk → p ∈ X
because X is complete by assumption. Also, for every m and n > m we have Bn ⊂ B(pm ; εm /2),
so that
d(pm , p) ≤ d(pm , pn ) + d(pn , p) < εm /2 + d(pn , p) → εm /2,

119
0.38. UNIFORM BOUNDEDNESS THEOREM

c
as n → ∞. Hence p ∈ Bm for every m. Since Bm ⊂ M m , we now see that p ∈ / Mm for every m,
so that p ∈
/ dMm = X. This contradicts that p ∈ X. Hence Baire’s theorem is proved.
Remark 0.38.1. The converse of Baire category theorem is not true. Infact there are incomplete
metric space which is nonmeager and shown by Bourbaki incomplete normed space which is
of second category.
From Baire’s theorem we shall now readily obtain the desired uniform boundedness
theorem. This theorem states that if X is a Banach space and a sequence of operators
Tn ∈ B(X,Y ) is bounded at every point x ∈ X, then the sequence is uniformly bounded. In
other words, pointwise boundedness implies boundedness in some stronger sense, namely,
uniform boundedness.

Theorem 0.38.2. Uniform Boundedness Theorem. Let (Tn ) be a sequence of bounded linear
operators Tn : X → Y , X is Banach space and Y is normed space such that (kTn xk) is bounded for
every x ∈ X, that is kTn xk ≤ cx , where cx is a real number depends on x. Then the sequence of
the norms kTn k is bounded, that is, there is a c(independent of x) such that kTn k ≤ c, n = 1, 2, ...
[pointwise boundedness of operator Tn : X → Y is uniformly bounded.]

Proof. For every k ∈ N, let Ak ⊂ X be the set of all x such that kTn xk ≤ k for all n.
Ak is closed. Indeed, for any x ∈ Ak there is a sequence (x j ) in Ak converging to x. This
means that for every fixed n we have Tn x j ≤ k and obtain kTn xk ≤ k because Tn is continuous
and so is the norm. Hence x ∈ Ak , and Ak is closed.
As (kTn xk) is bounded, so each x ∈ X belongs to some Ak . Hence X = ∪∞ k=1 Ak .
Since X is complete, Baire’s theorem implies that some Ak contains an open ball, say B0 =
B(x0 , r) ⊂ Ak0 .
r
Let x ∈ X be arbitrary not zero. We set z = x0 + γx, where γ = 2kxk . Then

kz − x0 k ≤ r,

so that z ∈ B0 . Hence from the definition of Ak and B0 = B(x0 , r) ⊂ Ak0 , we obtain kTn zk ≤ k0
for all n. Also kTn x0 k ≤ k0 since x0 ∈ B0 . Therefore we obtain x = z−x0
γ . For all n, we get

kTn (z − x0 )k 1 4
kTn xk = ≤ (kTn zk + kTn x0 k) ≤ kxk k0 .
γ γ r
Hence for all n,
4
kTn k = sup kTn xk ≤ k0
kxk=1 r

or kTn k ≤ c for c = 4r k0 .

Proposition 0.38.3. Space of Polynomials. The normed space X of all polynomials wlth norm
defined by
kxk = max|αi |,
i
where α0 , α1 , ... are coefficients of x, is not complete.

Proof. We construct a sequence of bounded linear operators on X which satisfies kTn xk ≤ cx

but not kTn xk ≤ c, so that X cannot be complete.

120
0.38. UNIFORM BOUNDEDNESS THEOREM

We may write a polynomial x 6= 0 of degree Nx in the form x(t) = ∑∞j=0 α j t j , where α j = 0

for j > Nx .
As a sequence of operators on X we take the sequence of functionals Tn = fn defined by

Tn 0 = fn (0) = 0, Tn x = fn (x) = α0 + α1 + ... + αn−1 .

Here fn is linear and bounded since |α j | ≤ kxk, so | fn (x)| ≤ n kxk.

Furthermore, for each fixed x ∈ X the sequence (| fn (x)|) satisfies kTn xk ≤ cx because a
polynomial x of degree Nx has Nx + 1 coefficients, we have | fn (x)| ≤ (Nx + 1) max j |α j | = cx .
Now, we have to show that the sequence (| fn (x)|) does not satisfy kTn xk ≤ c, that is there
is no c such that the sequence (| fn (x)|) satisfies kTn k = k fn k ≤ c for all n. For fn we choose x
defined by x(t) = 1 + t + ... + t n . Then kxk = maxi |αi | = 1 and fn (x) = 1 + 1 + ... + 1 = n =
n kxk. Hence k fn k ≥ | fn (x)|/ kxk = n, so sequence (k fn k) is unbounded.

121
0.39. OPEN MAPPING THEOREM

0.39 Open Mapping Theorem

It will be concerned with open mappings. These are mappings such that the image of
every open set is an open set. More specifically, the open mapping theorem states con-
ditions under which a bounded linear operator is an open mapping, As in the uniform
boundedness theorem we again need completeness, and the present theorem exhibits an-
other reason why Banach spaces are more satisfactory than incomplete normed spaces.
The theorem also gives conditions under which the inverse of a bounded hnear operator
is bounded. The proof of the open mapping theorem will be based on Baire’s category
theorem.
Definition 0.39.1. (Open mapping). Let X and Y be metric spaces. Then T : D(T ) → Y with
domain D(T ) ⊂ X is called an open mapping if for every open set in D(T ) the image is an open
set in Y .
Note that if a mapping is not surjective, one must take care to distinguish between the
assertions that the mapping is open as a mapping from its domain
(a) into Y ,
(b) onto its range.
(b) is weaker than (a). For instance, (a) if X ⊂ Y , the mapping x 7→ x of X into Y is open
if and only if X is an open subset of Y , whereas (b) the mapping, x 7→ x of X onto its range
(which is X) is open in any case.
Remark: f : X → Y is continuous iff f −1 (A) is open in X for each open set A in Y . It
does not give that f is an open mapping, that is every continuous function may not be an
open mapping, for example, f : R → R as f (t) = sint t ∈ R, Here t 7→ sint is continuous
but maps (0, 2π) onto [−1, 1]. So we require some additional condition for this given by
open mapping theorem.

Theorem 0.39.1. Open Mapping Theorem, Bounded Inverse Theorem. A bounded linear op-
erator T from a Banach space X onto a Banach space Y is an open mapping. Hence if T is
bijective, T −1 is continuous and thus bounded.

We need the following result to prove the open mapping theorem.

Lemma 0.39.2. (Open unit ball). A bounded linear operator T from a Banach space X onto a
Banach space Y has the property that the image T (B0 ) of the open unit ball B0 = B(0; 1) ⊂ X
contains an open ball about 0 ∈ Y .

Proof. Proceeding stepwise, we prove:

(a) The closure of the image of the open ball B1 = B(0; 1/2) contains an open ball B∗ .
(b) T (Bn ) contains an open ball Vn about 0 ∈ Y , where Bn = B(0; 2−n ) ⊂ X.
(c) T (B0 ) contains an open ball about 0 ∈ Y .
First (a) In connection with subsets A ⊂ X we shall write αA (α a scalar) and A + w (w ∈ X)
to mean
(1) αA = {x ∈ X : x = αa, a ∈ A}
(2) A + w = {x ∈ X : x = a + w, a ∈ A}
and similarly for subsets of Y .
We consider the open ball B1 = B(0; 1/2) ⊂ X. Any fixed x ∈ X is in kB1 with real k
sufficiently large (k > 2 kxk). Hence X = ∪∞ k=1 kB1 .

122
0.39. OPEN MAPPING THEOREM

Since T is surjective and linear,

Y = T (X) = T (∪∞ ∞ ∞
k=1 kB1 ) = ∪k=1 kT (B1 ) = ∪k=1 kT (B1 ).
Here, note that by taking closures we did not add further points to the union since that
union was already the whole space Y . Since Y is complete, it is nonmeager in itself, by Baire’s
category theorem. Hence we conclude that a kT (B1 ) must contain some open ball. This implies
that T (B1 ) also contains an open ball, say, B∗ = B(y0 ; ε) ⊂ T (B1 ). It follows that B∗ − y0 =
B(0; ε) ⊂ T (B1 ) − y0 .
(b) We now prove that B∗ − y0 ⊂ T (B0 ), where B0 = B(0; 1) ⊂ X. We’ll show show this by
T (B1 ) − y0 ⊂ T (B0 ).
Let y ∈ T (B1 ) − y0 . Then y + y0 ∈ T (B1 ) and y0 ∈ T (B1 ). Hence there are
un = Twn ∈ T (B1 ) such that un → y + y0 ,
and
vn = T zn ∈ T (B1 ) such that vn → y0 .
Since wn , zn ∈ B1 and B1 has radius 1/2, it follows that
kwn − zn k ≤ kwn k + kzn k < 1/2 + 1/2 = 1,
so that wn − zn ∈ B0 . Further, we get
T (wn − zn ) = Twn − T zn = un − vn → y,
so y ∈ T (B0 ). Since y ∈ T (B1 ) − y0 , we get T (B1 ) − y0 ⊂ T (B0 ). Therefore, we have B∗ − y0 =
B(0, ε) ⊂ T (B0 ).
Let Bn = B(0, 2−n ) ⊂ X. Since T is linear, T (Bn ) = 2−n T (B0 ) Thus we obtain Vn =
B(0, ε/2n ) ⊂ T (Bn ).
(c) Finally, we prove that V1 = B(0, ε/2) ⊂ T (B0 ) by showing that every y ∈ V1 is in T (B0 ).
Let y ∈ V1 . For n = 1, we have V1 = B(0, ε/21 ) ⊂ T (B1 ). Hence y ∈ T (B1 ). So there
must be a v ∈ T (B1 ) close to y, say, ky − vk < ε/4. Now v ∈ T (B1 ) implies v = T x1 for some
x1 ∈ B1 . Hence ky − T x1 k < ε/4.
For n = 2, we see that y − T x1 ∈ V2 ⊂ T (B2 ). As before we conclude that there is an x2 ∈ B2
such that k(y − T x1 ) − T x2 k < ε/8. Hence y − T x1 − T x2 ∈ V3 ⊂ T (B3 ) and so. In the nth step,
we can choose xn ∈ Bn such that y − ∑nk=1 T xk < ε/2n+1 , n = 1, 2, ...
Let zn = x1 +...+xn . Since xk ∈ Bk , we have kxk k < 1/2k . So for n > m, we get kzn − zm k ≤
n k
∑k=m+1 kxk k < ∑∞ k=m+1 1/2 → 0 as m → ∞. Hence (zn ) is Cauchy. (zn ) converges, say,
zn → x because X is complete. Also x ∈ B0 since B0 has radius 1 and ∑∞ ∞ i
i=1 kx1 k ≤ ∑i=2 1/2 <
n n+1
1/2 + 1/2 = 1. Since T is continuous, T zn → T x, taking n → ∞ in y − ∑k=1 T xk < ε/2 ,
we get T x = y. Hence y ∈ T (B0 ).
Proof. Open Mapping Theorem. We have to prove that for every open set A ⊂ X the image
T (A) is open in Y . This we do by showing that for every y = T x ∈ T (A) the set T (A) contains
an open ball about y = T x.
Let y = T x ∈ T (A). Since A is open, it contains an open ball with center x. Hence A − x
contains an open ball with center 0; let the radius of the ball be r and set k = 1/r, so that r = 1/k.
Then k(A − x) contains the open unit ball B(0; 1). Lemma now implies that T (k(A − x)) =
k[T (A) − T x] contains an open ball about 0, and so does T (A) − T x. Hence T (A) contains an
open ball about T x = y, Since y ∈ T (A) was arbitrary, T (A) is open.
Finally, if T −1 : Y → X exists, it is continuous because T is open. Since T −1 is linear, it is
bounded.

123
0.40. CLOSED GRAPH THEOREM

0.40 Closed Graph Theorem

Not all linear operators of practical importance are bounded For instance, the differen-
tial operator is unbounded, and in quantum mechanics and other applications one needs
unbounded operators quite frequently. However, practically all of the linear operators
which the analyst is likely to use are so called closed linear operators. This makes it
worthwhile to give an introduction to these operators. Now, we define closed linear oper-
ators on normed spaces and consider some of their properties, in particular in connection
with the important closed graph theorem which states sufficient conditions under which
a closed linear operator on a Banach space is bounded.
Definition 0.40.1. Let X and Y be normed spaces and T : D(T ) → Y a linear operator with
domain D(T ) ⊂ X. Then T is called a closed linear operator if its graph G(T ) = {(x, y) : x ∈
D(T ), y = T x} is closed in the normed space X × Y , where the two algebraic operations of a
vector space in X ×Y are defined as usual, that is
(x1 , y1 ) + (x2 , y2 ) = (x1 + x2 , y1 + y2 )
and
α(x, y) = (αx, αy),
α is scalar and the norm on X ×Y is defined by
k(x, y)k = kxk + kyk .
Under what conditions will a closed linear operator be bounded? It’s answer is given
by the following result.
Theorem 0.40.1. Closed Graph Theorem. Let X and Y be Banach spaces and T : D(T ) → Y a
linear operator with domain D(T ) ⊂ X. If D(T ) is closed in X, then the operator T is bounded.
Proof. We first show that the norm X ×Y defined by
k(x, y)k = kxk + kyk
is complete.
Let (zn ) be Cauchy in X ×Y , where zn = (xn , yn ). Then for every ε > 0 there is an N such
that
kzn − zm k = k(xn , yn ) − (xm , ym )k = kxn − xm k + kyn − ym k < ε,
for m, n > N. Hence (xn ) and (yn ) are Cauchy in X and Y , respectively, and converge, say,
xn → x and yn → y, because X and Y are complete. This implies that zn → z = (x, y), since with
m → ∞ we have kzn − zk ≤ ε for n > N. Since the Cauchy sequence (zn ) was arbitrary, X ×Y
is complete.
By assumption, G(T ) is closed in X × Y and D(T ) is closed in X. Hence G(T ) and D(T )
are complete.
We now consider the mapping P : G(T ) → D(T ) as (x, T x) 7→ x. Here it is easy to check
that P is linear and one-one. Also, P is bounded because
kP(x, T x)k = kxk ≤ kxk + kT xk = k(x, T x)k .
P is bijective, inverse mapping is P−1 : D(T ) → G(T ) as x 7→ (x, T x). Since G(T ) and D(T ) are
complete, so by the bounded inverse theorem, we have P−1 is bounded, say k(x, T x)k ≤ b kxk
for some b and x ∈ D(T ). Hence T is bounded because
kT xk ≤ kT xk + kxk = k(x, T x)k ≤ b kxk ,
for all x ∈ D(T ).

124
0.40. CLOSED GRAPH THEOREM

By definition, G(T ) is closed if and only if z = (x, y) ∈ G(T ) implies z ∈ G(T ). Also we
see that z ∈ G(T ) if and only if there are zn = (xn , T xn ) ∈ G(T ) such that zn → z, hence
xn → x, T xn → y,
and z = (x, y) ∈ G(T ) iff x ∈ D(T ) and y = T x. This proves the following useful criterion
which expresses a property that is often taken as a definition of closedness of a linear
operator.

Theorem 0.40.2. (Closed linear operator). Let T : D(T ) → Y be a linear operator, where
D(T ) ⊂ X and X and Y are normed spaces. Then T is closed if and only if it has the fol-
lowing property. If xn → x, where xn ∈ D(T ), and T xn → y, then x ∈ D(T ) and T x = y.

Note well that this property is different from the following property of a bounded
linear operator. If a linear operator T is bounded and thus continuous, and if (xn ) is a
sequence in D(T ) which conveIges in D(T ), then (T xn ) also converges.
This need not hold for a closed linear operator. However, if T is closed and two se-
quences (xn ) and (yn ) in the domain of T converge with the same limit and if the corre-
sponding sequences (T xn ) and (Tyn ) both converge, then the latter have the same limit.

Proposition 0.40.3. Let X = C[0, 1] and T : D(T ) → X x 7→ x0 , where 0 denotes differentiation

and D(T ) is the subspace of functions x ∈ X which have a continuous derivative. Then T is not
bounded, but is closed.

Proof. Here T is not bounded as xn (t) = t n ,kxn k = 1 T xn (t) = xn0 = nt n−1 , so kT xn k = n and
kT xn k / kxn k = n, so there is no fixed c such that kT xn k / kxn k ≤ c.
Now, we’ll prove that T is closed by Closed linear operator theorem.
Let (xn ) in D(T ) be such that both (xn ) and (T xn ) converge, say, xn → x and T xn = xn0 → y.
Since convergence in the norm of C[O, 1] is uniform convergence on [0, 1], from xn0 → y,
we have Z t Z t Z t
0
y(τ)dτ = lim xn (τ)dτ = lim xn0 (τ)dτ = x(t) − x(0),
0 0 n→∞ n→∞ 0
that is, Z t
x(t) = x(0) + y(τ)dτ.
0
This shows that x ∈ D(T ) and x0 = y, because D(T ) ⊂ C[0, 1], set of all functions x which
has continuous derivatives. Hence T is closed.
It is worth noting that in above proposition, D(T ) is not closed in X since T would then
be bounded by the closed graph theorem.
Closedness does not imply boundedness of a linear operator. Conversely, boundedness
does not imply closedness.
For example: Let T : D(T ) → D(T ) ⊂ X be the identity operator on D(T ), where D(T )
is a proper dense subspace of a normed space X. Then it is trivial that T is linear and
bounded. However, T is not closed. This follows from closed linear operator theorem if
we take x ∈ X − D(T ) and a sequence (xn ) in D(T ) which converges to x.

Lemma 0.40.4. Closed Operator. Let T : D(T ) → Y be a bounded linear operator with domain
D(T ) ⊂ X, where X and Y are normed spaces. Then
(a) If D(T ) is a closed subset of X, then T is closed.
(b) If T is closed and Y is complete, then D(T ) is a closed subset of X.

125
0.40. CLOSED GRAPH THEOREM

Proof. (a) If (xn ) is in D(T ) and converges, say, xn → x, and is such that (T xn ) also converges,
then x ∈ D(T ) = D(T ) since D(T ) is closed, and T xn → T x since T is continuous. Hence T is
closed.
(b) For x ∈ D(T ), there is a sequence (xn ) in D(T ) such that xn → x. Since T is bounded,

kT xn − T xm k = kT (xn − xm )k ≤ kT k kxn − xm k .

This shows that (T xn ) is Cauchy. (T xn ) converges, say, T xn → y ∈ Y because Y is complete.

Since T is closed, x ∈ D(T ) (and T x = y). Hence D(T ) is closed because x ∈ D(T ) was
arbitrary.

126
0.41. PROJECTION OPERATORS

0.41 Projection Operators

Let H be a Hilbert space and Y be a closed subspace of H. Then Projection Theorem
says that each member x ∈ H has a unique L representation as x = y + z, where y ∈ Y and
z ∈ Y ⊥ ⊂ H and in this case we write H = Y Y ⊥ as a direct sum decomposition of H.
Now consider a mapping P : H → H sends x ∈ H to y ∈ Y , where x = y + z, y ∈ Y ,
z ∈ Y ⊥ . It is well defined because it sends x to a unique element y ∈ H.
Definition 0.41.1. The mapping P : H → H defined as Px = y where x = y + z, x ∈ H, y ∈ Y is
called a Projection operator or orthogonal Projection operator on H.
Note that Y is the corresponding range subspace of P.
It is readily seen that a Projection operator P : H → H is a linear operator where Y is
given a closed linear subspace of H and P(x ∈ H) ∈ Y .
If y ∈ Y ⊂ H, then y = y + 0, 0 ∈ Y (by representation formula) and Py = y ∈ Y . There-
fore P restricted to Y becomes the identity operator on Y .
On the other hand if z ∈ Y ⊥ ⊂ H, we have z = 0 + z, where 0 ∈ Y and z ∈ Y ⊥ , we see
that P(z) = 0 and conversely if P(w) = 0, we have w = 0 + w showing w ∈ Y ⊥ .
Hence null space of P = Y ⊥ .
Also we can now write x = y + z = P(x) + z. Thus z = x − P(x) = I(x) − P(x) = (I − P)(x),
where I : H → H is the identity operator. So we say that the operator I −P as the Projection
operator of H onto Y ⊥ .
Remark 0.41.1. If P is the projection operator over H with it’s range, a closed subspace Y of
H, then I − P is a projection operator over H with its range Y ⊥ and vice versa.
Theorem 0.41.1. A bounded linear operator P : H → H, H being a Hilbert space is a projection
operator if and only if P is idempotent (i.e. P2 = P) and self adjoint.
Proof. Let P : H → H beL a projection operator and write P(H) = Y , where Y is a closed
subspace of H, and H = Y Y ⊥ .
If x ∈ H, then P(x) = y ∈ Y . So P2 (x) = P(P(x)) = P(y) = y = P(x), this is true for all
x ∈ H. Hence P2 = P, implying that P is idempotent.
Again suppose that x1 = y1 + z1 and x2 = y2 + z2 where x1 ; x2 ∈ H, y1; y2 ∈ Y , z1 ; z2 ∈ Y ⊥ .
Then hy1 , z2 i = 0 = hz1 , y2 i, since Y ⊥ Y ⊥ . So
hP(x1 ), x2 i = hy1 , x2 i
= hy1 , y2 + z2 i
= hy1 , y2 i + hy1 , z2 i
= hy1 , y2 i
= hy1 , y2 i + hz1 , y2 i
= hy1 + z1 , y2 i
= hx1 , y2 i = hx1 , P(x2 )i ,
since hy1 , z2 i = 0 = hz1 , y2 i . So P is self adjoint.
Conversely, suppose that P is self adjoint (i.e. P∗ = P) and idempotent (i.e. P2 = P). Let
P(H) = Y . Thus for every x ∈ H, we have x = P(x) + (I − P)(x). Now as (I − P)x ∈ Y ⊥ , P∗ = P
and P2 = P, we have
hP(x), (I − P)xi = hx, P∗ (I − P)(x)i
= hx, P(I − P)(x)i
= x, P(x) − P2 (x) = hx, 0i .

127
0.41. PROJECTION OPERATORS

Now (I − P)(P(x)) = P(x) − P2 (x) = 0, shows that null space of (I − P), that is, N(I − P)
satisfies Y ⊂ N(I − P). Again (I − P)(x) = 0 =⇒ x = P(x) ∈ Y . So N(I − P) ⊂ Y .
Consequently, N(I − P) = Y . So Y is always a closed subspace of H, as N(I − P) is a closed
subspace of H.
Finally P restricted to Y is the identity operator because y = P(x) gives P(y) = P(P(x)) =
P2 (x) = P(x) = y. This implies that P is a Projection operator on H.

Theorem 0.41.2. The projection operator P : H → H, H being a Hilbert space, is bounded and
self adjoint such that ||P|| = 1.

Proof. Suppose that Px = y where x = y + z; y ∈ Y = P(H); z ∈ Y ⊥ . Now ||Px||2 = ||y||2 ≤

||y + z||2 = ||x||2 , implies that P is bounded with ||P|| ≤ 1. Now if x 6= 0 such that x ∈ Y . Then
Px = x and ||Px|| = ||x|| =⇒ ||P|| = 1.
We shall now show that P is self adjoint. Let x1 , x2 ∈ H. Thus we have x1 = y1 + z1 ,
x2 = y2 + z2 , where y1 , y2 ∈ Y and z1 , z2 ∈ Y ⊥ . So Px1 = y1 , Px2 = y2 . Therefore,

hPx1 , x2 i = hy1 , x2 i = hy1 , y2 + z2 i = hy1 , y2 i + hy1 , z2 i = hy1 , y2 i ,

since hy1 , z2 i = 0 and

hx1 , Px2 i = hx1, y2 i = hy1 + z1 , y2 i = hy1 , y2 i + hz1 , y2 i = hy1 , y2 i ,

since hz1 , y2 i = 0. So hPx1 , x2 i = hx1 , Px2 i. Hence it follows that P is self adjoint.

128
0.42. COMPACT LINEAR OPERATORS

0.42 Compact linear operators

Definition 0.42.1. A linear operator T : X → Y between normed spaces X and Y is called a
compact linear operator if for every bounded sequence (xn ) in X, the sequence (T xn ) has a
convergent subsequence.
We note that every compact operator T is bounded. Indeed, if ||T || = ∞, then there
exists a sequence (xn) such that ||xn || ≤ 1 and ||T xn || → ∞. Then (T xn ) cannot have a
convergent subsequence. Hence, ||T || < ∞.
Example 0.42.1. Consider the linear operator TN : `2 → `2 defined by TN x = (x1, x2, . . . , xN , 0, 0, . . .).
We claim that TN is compact. Given any bounded sequence (x(n) ) ∈ `2 , the sequence (TN x(n) )
is bounded in FN ⊂ `2 . Since every bounded sequence in RN or CN has a convergent subse-
quence, it follows that TN is compact.
Example 0.42.2. The identity operator I : `2 → `2 is not compact. To prove this, consider the
√ (en ) where en = (0, . . . , 0, 1, 0, . . .). Then for any n 6= m, ||Ien − Iem ||2 =
bounded sequence
||en − em ||2 = 2. This implies that any subsequence of (Ien ) cannot be Cauchy and, hence,
cannot converge. So that I is not compact.

Theorem 0.42.3. Let (T n) be a sequence of compact linear operators from a normed space X
into a Banach spaces Y . If Tn → T (that is, ||Tn − T || → 0), then the limit operator T is compact.

Proof. Since T1 is a compact operator, we know that the sequence (T1 (xn )) has a convergent
(hence Cauchy) subsequence (T1 (x1,m )), where (x1,m ) is a subsequence of the original sequence
(xn ). The subsequence (x1,m ) is bounded, so we can repeat the argument with T2 to produce a
subsequence (x2,m ) of (x1,m ) with the property that (T2 (x2,m )) converges. We continue in the
same way, and then define a sequence (ym ) = (xm,m ). Notice that (ym ) is a subsequence of (xn ),
so it is bounded, say by ||yn || ≤ c, and it has the property that for every fixed n, the sequence
(Tn (ym )) is convergent, and hence Cauchy.
We claim that (T (ym )) is a Cauchy sequence in Y . Let ε > 0. Since ||Tn − T || → 0, there
ε
is some p ∈ N such that ||Tp − T || < 3c . Also, since (Tp (ym )) is Cauchy, there is some N > 0
such that ||Tp (y j ) − Tp (yk )|| < ε3 , whenever j, k > N. Therefore, for j, k > N, we have

Ty j − Tyk ≤ Ty j − Tp y j + Tp y j − Tp yk + Tp yk − Tyk
≤ Ty j − Tp y j + Tp y j − Tp yk + Tp yk − Tyk
ε
< T − Tp y j + + T − Tp kyk k
3
ε ε ε
< c + + c = ε.
3c 3 3c
Hence (T (ym )) is a Cauchy sequence in Y . Since Y is a Banach space, it is by definition
complete, so (T (ym )) converges. We have thus produced, for an arbitrary bounded sequence
(xn ) ⊂ X, a convergent subsequence of its image under T . Therefore, T is compact.
Example 0.42.4. Let T : `2 → `2 be an operator defined by T x = (αn xn ) for a sequence αn → 0.
We claim that T is compact. To show this, we approximate T by compact operators TN such
that TN x = (α1 x1 , . . . , αN xN , 0, . . .). As in the previous example we observe that TN is a compact

129
0.42. COMPACT LINEAR OPERATORS

operator. For x ∈ `2 , we have

!1
2
2
kTN x − T xk2 = ∑ |αnxn|
n>N
!1
  2
2
≤ sup |αn | ∑ |xn|
n>N n>N
 
≤ sup |αn | kxk2 .
n>N

This shows that ||TN − T || ≤ supn>N |αn |. Since αn → 0, it follows that ||TN − T || → 0. Hence
T is compact.

130
0.43. STRONG CONVERGENCE AND WEAK CONVERGENCE

0.43 Strong Convergence and Weak Convergence

Definition 0.43.1. A sequence {xn } in a normed linear space X over a field K is said to be
weakly convergent to an element x ∈ X if for every f ∈ X ∗ , f (xn ) → f (x) as n → ∞ and written
w
as xn −
→ x.
Definition 0.43.2. A sequence {xn } in a normed linear space X over a field K is said to be
s
strongly convergent to an element x ∈ X if ||xn − x|| → 0 as n → ∞ and written as xn →
− x.
w
Theorem 0.43.1. Suppose that a sequence {xn } in a normed linear space such that xn −
→ x ∈ X.
Then
(i) The weak limit x of {xn } is unique.
(ii) Every subsequence of {xn } in X is convergent weakly to x.
(iii) The sequence {||xn ||} is norm bounded.
w w
Proof. Take xn − → x ∈ X and xn − → y ∈ X, where x 6= y. Using the application of Hahn Banach
Theorem [Given a normed linear space X over a field K and a non zero member x0 ∈ X, there
is a bounded linear functional F over X such that F(x0 ) = ||x0 || and ||F|| = 1.], there exists a
w w
f ∈ X ∗ such that f (x − y) = ||x − y|| with || f || = 1. Since xn −
→ x ∈ X and xn −
→ y ∈ X,, so for
every f ∈ X ∗ f (xn ) → f (x) and f (xn ) → f (y) as n → ∞. So f (x) = f (y).
So f (x − y) = 0, implies that ||x − y|| = 0, so x = y, a contradiction to our assumption.
Hence the weak limit x of {xn } is unique.
w
(ii) As xn −→ x ∈ X then for every f ∈ X ∗ f (xn ) → f (x). So, every subsequence of { f (xn )}
is convergent and converges to f (x). Let { f (xnk )} be a subsequence of { f (xn )}. So f (xnk ) →
w
f (x), that is, xnk −
→ x.
w
(iii) As xn −→ x ∈ X for every f ∈ X ∗ , f (xn ) → f (x). So | f (xn )| ≤ M f for some constant
M f > 0. Let C : X → X ∗∗ be a canonical mapping defined by C(x) = Fx where Fx ∈ X ∗∗ satis-
fying ||Fx || = ||x||. So for every f ∈ X ∗ , we have Fx ( f ) = f (x). Take a sequence of functionals
{gn } ∈ X ∗∗ such that gn ( f ) = f (xn ). So |gn ( f )| ≤ M f for every f ∈ X ∗ . As X ∗ is complete
and {gn ( f )} over X ∗ is a bounded sequence of scalars, so by Uniform Boundedness Principle,
||gn || is bounded. But ||gn || = ||xn ||. So {||xn ||} is bounded.
Theorem 0.43.2. Let X be a normed linear space. Then Strong convergence implies weak
convergence of a sequence in X.
s
Proof. Let {xn } be a sequence in X and let xn →− x ∈ X.
∗
Let f ∈ X . Now, | f (xn ) − f (x)| = | f (xn − x)| ≤ || f ||||xn − x|| → 0. This is true for all
w
f ∈ X ∗ . Hence xn −
→ x.
The converse is not true. For example
Example 0.43.3. Let X = `2 . So, X ∗ = `2 . Let {ek }k=1,2,... be a Schauder basis of X where ek =
(0, 0, . . . , 1 , 0, 0, . . .). Let x ∈ `2 where x = (ξ1 , ξ2 , . . . , ξn , . . .) with ||x|| = (∑∞ 2 1/2 <
i=1 |ξi | )
kthplace
∞. Take x ∈ `2 . Then x = ∑∞ i=1 αk ek , where αk ’s are scalars.
∗
Let f ∈ X . So f (x) = f (∑∞ ∞ ∞
k=1 αk ek ) = ∑k=1 αk f (ek ) = ∑k=1 αk γk where γk = f (ek ). Note
that γk ∈ `2 , i.e. (γ1 , γ2 , . . . , γn , . . .) ∈ `2 where ∑∞ 2
k=1 |γk | < ∞. So, γk → 0. So f (ek ) → 0,
w
f (ek ) → f (0). Hence ek − →√0 for any two positive integers n and m with n 6= m. Now ||en −
em || = 2 =⇒ ||en − em || = 2 > 0. So {en } is not a Cauchy sequence in X and hence en 9s 0.

131
0.43. STRONG CONVERGENCE AND WEAK CONVERGENCE

In an infinite dimensional normed linear space X weak convergence of a sequence in

X does not necessarily imply its strong convergence.
Theorem 0.43.4. Let X be a normed linear space. If dim(X) < ∞ then weak convergence and
strong convergence of a sequence in X coincide.
Proof. Note that strong convergence of a sequence in a normed linear space implies its weak
convergence. Since, dim(X) < ∞ and say dim(X) = k. Take, e1 , e2 , . . . , ek be a basis vectors of
(n) (n) (n)
X. Let {xn } ∈ X and x ∈ X such that xn = α1 e1 + α2 e2 + . . . + αk ek and x = α1 e1 + α2 e2 +
. . . + αk ek . (
0i f j 6= n (n)
Let f1 , f2 , . . . , fn ∈ X ∗ such that f j (en ) = So, f j (xn ) = α j and f j (x) = α j .
1i f j = n
w (n)
→ x. So, f j (xn ) → f j (x), for j = 1, 2, . . . , k, α j → α j .
Let xn −
Let M = max ||e j ||. Let ε > 0 be arbitrary. So there exists a positive integer N ∈ N such
1≤ j≤n
(n) ε
that |α j − α j | < Mk for all n ≥ N for each j = 1, 2, . . . , k. Now

k n
(n)
||xn − x|| = ∑ α j e j − ∑ α je j
j=1 k=1
k
(n)
= ∑ (α j − α j )e j
j=1
k
(n)
≤ ∑ (α j − α j )e j
j=1
k
(n)
= ∑ |α j − α j| e j
j=1
ε
<k M = ε∀n ≥ N.
Mk
s
Thus xn →
− x.
Let X and Y be two normed linear space over the same field of scalars K.
Definition 0.43.3. Let {Tn } ∈ BLd (X,Y ) and let T ∈ BLd (X,Y ) then Tn → T in norm if for
every ε > 0 there exists a positive integer N such that ||Tn − T || < ε∀n ≥ N.
Definition 0.43.4. Let {Tn } ∈ BLd (X,Y ) and let T ∈ BLd (X,Y ) then Tn → T weakly written as
w w
Tn − → T x i.e. for every y0 ∈ Y 0 , y0 (T n(x) → y0 (T x).
→ T if for all x ∈ X, Tn x −
Definition 0.43.5. The sequence {Tn } ∈ BLd (X,Y ) is said to be strongly convergent if for every
x ∈ X, {Tn (x)} converges in Y .
Definition 0.43.6. Let {Tn } ∈ Bd L(X,Y ) and T ∈ Bd L(X,Y ). Then Tn is called strongly con-
s
vergent to T if for x ∈ X the sequence {Tn x} converges to T x and we write Tn →
− T.
Note that norm converges of a sequence of operators {Tn } ∈ BLd (X,Y ) to T ∈ BLd (X,Y )
implies its strong converges, because ||Tn (x) − T (x)|| = ||(Tn − T )x|| ≤ ||Tn − T ||||x|| → 0.

132
0.43. STRONG CONVERGENCE AND WEAK CONVERGENCE

Definition 0.43.7. Let {Tn } ∈ BLd (X,Y ) and let T ∈ BLd (X,Y ). Then {Tn x} converges to T (x)
uniformly for all x ∈ X such that ||x|| ≤ 1, if for every ε > 0, there exists a positive integer N
u
such that ||Tn x − T x|| < ε whenever n ≥ N written as Tn x →
− T x.
Alternatively we say uniform convergence of a sequence {Tn } ∈ BLd (X, y) in ||x|| ≤ 1
u
written as Tn →− T.
Theorem 0.43.5. If {Tn } ∈ BLd (X,Y ) be such that {Tn } ∈ BLd (X,Y ) uniformly for all x ∈ X
satisfying ||x|| ≤ 1 then Tn (x) → T (x) uniformly for all x ∈ X satisfying ||x|| ≤ 1.

Proof. Now, ||Tn x − T x|| = ||(Tn − T )(x)|| ≤ ||Tn − T ||||x||. Let x ∈ X such that ||x|| ≤ 1. So
u
||Tn x − T x|| → 0, implying Tn x →
− T x uniformly for all x ∈ X satisfying ||x|| ≤ 1.
u
Theorem 0.43.6. If {Tn } ∈ BLd (X,Y ) and T ∈ BLd (X,Y ) be such that Tn x →
− T x for all x ∈ X
such that ||x|| ≤ 1 then Tn → T in norm.
Proof. Let x ∈ X such that ||x|| ≤ 1. For any ε > 0 there exists N ∈ N such that ||Tn x − T x|| < ε
x
whenever n ≥ N. Let y = ||x|| , x 6= 0.
u
So, ||y|| = 1. So, Tn y →
− Ty, Therefore, we have
||Tn y − Ty|| < ε∀n ≥ N
||Tn x/||x|| − T x/||x|| || < ε∀n ≥ N
||Tn x − T x||/||x|| < ε∀n ≥ N
||(Tn − T )x||/||x|| < ε∀n ≥ N
||(Tn − T )x|| < ε||x||, ∀n ≥ N
So, sup ||x|| ≤ 1||(T n − T )x|| ≤ ε, ∀n ≥ N. So, ||(Tn − T )|| ≤ ε, ∀n ≥ N, implying that Tn → T
in norm.
s w
Theorem 0.43.7. Let {Tn } ∈ BLd (X,Y ) and let T ∈ BLd (X,Y ). If Tn →
− T then Tn −
→ T.
s
Proof. Let Tn → − T . Take ε > 0 be an arbitrary. So there exists a positive integer N such that
||Tn − T || < ε for all n ≥ N. Let y0 ∈ Y 0 . Take x ∈ X. Now,
|y0 (T nx) − y0 (T x)| = |y0 (Tn x − T x)|
≤ ||y0 ||||Tn x − T x||
≤ ||y0 ||||Tn − T ||||x|| → 0.
w w
Hence y0 (Tn x) → y0 (T x), Tn x −
→ T x, Tn −
→ T. Thus we have uniform convergence in ||x|| ≤
1.
So strong convergence implies weak convergence but weak convergence doesnot imply
its norm convergence, for example:
Example 0.43.8. Let X = `2 . Define Tn : X → X by Tn (x) = (0, 0, . . . , ξn+1 , ξn+2 , . . .), where
x = (ξ1 , ξ2 , . . . , ξn , . . .) ∈ X. Clearly for each positive integer n, Tn is linear. Now ||Tn (x)|| =
(∑∞ 2 1/2 ≤ ( ∞ |ξ |2 )1/2 = ||x||, implying that T ∈ BL (X,Y ). Since ∞ |ξ |2 < ∞
i=n+1 |ξi | ) ∑i=n i n d ∑i=1 i
w
so ξn → 0. So Tn (x) = (0, 0, . . . , 0, ξn+1 , ξn+2 , . . .) → (0, 0, . . . , 0, . . . , 0, . . .) = 0. Hence Tn −
→ Θ,
the zero operator on X. Take x = (0, 0, . . . , 0, ξn+1 , ξn+2 , . . .) ∈ X such that ||x|| ≤ 1. Then
Tn (x) = x. So ||Tn || = 1 for all n. But ||Tn − Θ|| = ||Tn || = 1, implying that Tn 9 Θ in norm.
So by {Tn } does not converge uniform to the zero operator.

133
0.44. REFERENCES

0.44 References
1. Kreyszig, E., Introductory Functional Analysis with Applications, Wiley (1989).

2. Rudin, W., Functional Analysis, McGraw Hill (1991).

3. Simmons, G.F., Introduction to Topology and Modern Analysis, Tata McGraw Hill (1963).

4. Keswan, S. Functional Analysis, Hindustan Book Agency (2009).

5. Limaye, B.V., Functional Anaylsis, Wiley Eastern Ltd (2007).

6. Conway J.B., A course in Functional Analysis, Springer Verlag (1990).

7. https://nptel.ac.in/courses/111/105/111105037 and https://epgp.inflibnet.ac.in.

134