Metric and Banach Spaces

Alexandre Daoud
Kings College London
alex.daoud@mac.com
April 28, 2016

Chapter 1
Sequence Spaces
1.1

Finite Dimensional Case

Definition 1.1. Let K be a field and V a vector space over K. A function || || : V R

is called a norm if, given x, y V and K, we have
1. Positivity - ||x|| 0 with equality if and only if x = 0
2. Homogeneity - ||x|| = ||x
3. Triangle inequality - ||x + y|| ||x|| + y||
Definition 1.2. Let p [1, ] and n a natural number. We define the p-norm on Cn to
be
! p1
n
X
||x||p =
,
for p <
|xi |p
i=1

||x|| = max |xi |,

i=1,...,n

for p =

The positivity and homogeneity norm axioms are easy to check. The triangle inequality
is more complicated. We first require the following two results:
Lemma 1.3 (Youngs Inequality). Let p, q R be such that p, q > 0 and 1/p + 1/q = 1.
Then for all a, b R such that a, b 0 we have
ab

ap b q
+
p
q

with equality if and only if a = b1 .

Proof. We have





1
1
1
1
p
q
ab = exp[ln(ab)] = exp[ln a + ln b] = exp p ln a + q ln b = exp
ln a + ln b
p
q
p
q
Now the exponential function is strictly increasing and strictly convex so, combined with
the hypothesis 1/p + 1/q = 1, we have
1
1
ap b q
p
q
ab exp[ln a ] + exp[ln b ] =
+
p
q
p
q

CHAPTER 1. SEQUENCE SPACES

Lemma 1.4 (Holders Inequality). Let p, q R be such that p > 1 and q > 1 and 1/p+1/q =
1. Then, given any x, y Cn , we have
||xy||1 ||x||p ||y||q
Proof. Let u = x/||x||p and v = y/||y||q . Then, clearly, ||u||p = ||v||q = 1. By Youngs
Inequality we have
|ui |p |vi |q
+
|ui vi |
p
q
Passing to the sum on both sides we see that
1
1
||uv||1 ||u||p + ||v||q = 1
p
q
Hence
||xy||1 = ||x||p ||y||q ||uv||1 ||x||p ||y||q

Remark. In the case that p = q = 2, Holders inequality reduces to the Cauchy-Schwarz

inequality.
Proposition 1.5 (Minkowskis inequality). Given x, y Cn , we have that
||x + y||p ||x||p + ||y||p
Proof. We shall first prove this for the case that p = 1. We have that
||x + y||1 =

n
X

|xi + yi |

i=1
n
X
i=1
n
X

|xi | + |yi |
|xi | +

i=1

n
X

|yi |

i=1

= ||x||1 + ||y||1
We now prove the proposition for the case where p = . We have that
||x + y|| = max |xi + yi |
i=1,...,n

max (|xi | + |yi |)

i=1,...,n

max |xi | + max |yi |

i=1,...,n

i=1,...,n

= ||x|| + ||y||
Now fix p (1, ) and choose q (1, ) such that p1 + 1q = 1. We first consider the triangle
inequality for the modulus of complex numbers. Let x, y Cn . Then
|xi + yi | |xi | + |yi |

CHAPTER 1. SEQUENCE SPACES

multiplying through by |xi + yi |p1 we get

|xi + yi |p |xi ||xi + yi |p1 + |yi ||xi + yi |p1
We now sum over i on both sides of the inequality:
n
X

n
X

|xi + yi |p

i=1

|xi ||xi + yi |p1 +

i=1

n
X

|yi ||xi + yi |p1

i=1

Now applying Holders inequality to both terms of the right hand side of the above inequality,
it follows that
! 1q
! 1q
! p1
! p1
n
n
n
n
n
X
X
X
X
X
|xi + yi |p
|xi + yi |(p1)q
+
|xi + yi |(p1)q
|xi |p
|yi |p
i=1

i=1

i=1

i=1

i=1

Now, q = (1 p1 )1 whence (p 1)q = p so this becomes

n
X

n
X

|xi + yi |p

i=1

! p1
|xi |p

i=1

n
X

! 1q
|xi + yi |p

i=1

n
X

! p1

n
X

|yi |p

i=1

! 1q
|xi + yi |p

i=1

Dividing through by the common factor in the terms of the right hand side yields
n
X

!1 1q
|xi + yi |p

n
X

i=1

! p1
|xi |p

i=1

n
X

! p1
|yi |p

i=1

We now note that 1 1/q = 1/p and Minkowskis Inequality is proven.

Definition 1.6. Let K be a field and V be a vector space over K. Suppose that || ||1
and || ||2 are two norms on V . We say that || ||1 and || ||2 are equivalent if there exist
c, C R such that for all x X we have
c||x||1 ||x||2 C||x||1
Proposition 1.7. Let p, q [1, ]. Then || ||p and || ||q are equivalent as norms on Cn .
1

Proof. Fix x Cn . It suffices to prove that ||x|| ||x||p n p ||x|| . We have that
p

max |xi |

i=1,...,n

n
X

|xi |p n max |xi |p

i=1...n

i=1

Now taking the pth root across these inequalities yields the desired result.

1.2

`p spaces

Proposition 1.8. Let p [1, ]. Consider the set

! p1
 X


`p = (x1 , x2 , . . . ) C 
|xi |p
<


i=1

CHAPTER 1. SEQUENCE SPACES

when p < and


=

(x1 , x2 , . . . ) C




 sup |xi | <
 iN

when p = . Then `p is a normed vector space over C with norm given by

! p1

X
||x||p =
|xi |p
i=1

when p < and

||x|| = sup |xi |
iN

when p = .
Proof. We shall prove the case where p [1, ). It is clear that `p contains an additive
identity, contains additive inverses, and contains scalar multiples. All additive and scalar
multiplicative properties of vector spaces are also satisfied trivially. We must check that if
C and x, y `p then x + y `p . In the finite dimensional case, Minkowskis Inequality
implies that
!1/p
!1/p
!1/p
n
n
n
X
X
X
|xi + yi |p

|xi |
+ ||
|yi |
i=1

i=1

i=1

Now, passing to the limit n , the right hand side is bounded by some constant and
thus x + y `p . Hence `p is a (infinite dimensional) vector space over C.
The positivity and homogeneity properties of || ||p follow immediately. The triangle
inequality also follows by a similar argument to the above where we pass to the limit n
in the finite dimensional Minkowski Inequality.
Recall that a vector space is complete if every Cauchy sequence in the vector space
converges to a limit in the vector space.
Definition 1.9. Let K be a field and V a vector space over K. We say that V is a Banach
space if V is a complete normed linear space.
Theorem 1.10. `p is a Banach space for any 1 p .
Proof. By Proposition 1.8, `p is a normed linear space. We must show that `p is complete.
(k)
Let { x(k) }kN be a Cauchy sequence in `p and fix i N. It is clear that |xi | ||x(k) ||p for
(k)
all k N whence { xi }kN is a Cauchy sequence of complex numbers. Since C is complete,
(k)
we have that xi xi as k .
We shall prove the theorem in the case 1 p < . Since x(k) is Cauchy, given any
> 0, there exists N N such that
!1/p
n
X
(k)
(m) p
|xi xi |

i=1

for all k, m N and for all n N. Letting m , we have

!1/p
n
X
(k)
|xi xi |p

i=1

CHAPTER 1. SEQUENCE SPACES

for all k N and for all n N. Since n is an arbitrary natural number, it follows that

!1/p
(k)

|xi xi |p

(1.1)

i=1

for all k N . By the definition of `p , it follows that x(k) x is an element of `p for any
k. Since x(k) is also an element of `p , by linearity, we must have that x `p . We can now
rewrite (1.1) as follows:
||x(k) x||p
for all k N . Since is arbitrary, we then have that ||x(k) x||p 0 as k whence `p
is complete.
Proposition 1.11. Let p, q [1, ]. Then `p and `q are not equivalent as normed spaces.
Proof. Suppose, without loss of generality, that p > q. Consider the linear subspace f of
`p and `q consisting of sequences with only finitely many non-zero terms. Let { g (n) }nN be
the sequence in f whose ith term is the sequence in f whose first i terms are 1 and the rest
zero. In other words,
g (1) = (1, 0, 0, 0, . . . )
g (2) = (1, 1, 0, 0, . . . )
g (3) = (1, 1, 1, 0, . . . )
..
.
Define the sequence
f (n) = n1/p g (n)
Then
||f (n) ||p = n1/p ||g (n) ||p = n1/p

!1/p
(n)

|gi |p

= n1/p n1/p = 1

i=1

So f (n) does not converge to 0 in `p . Now,

||f (n) ||q = n1/p ||g (n) ||q = n1/p n1/q = n1/q1/p
But p > q whence 1/q > 1/p so f (n) 0 as n and is thus Cauchy. Hence `p and `q
cannot be equivalent as normed spaces.
Definition 1.12. Let K be a field and X, Y vector spaces over K. Suppose that X is
equipped with the norm || ||X and Y is equipped with the norm || ||Y . If X is a subspace
of Y and there exists a constant c R such that ||x||Y c||x||X for all x X then we say
that X embeds in Y and we denote it as X , Y .
Proposition 1.13. Let p, r [1, ] such that p < r. Then `p , `r .

CHAPTER 1. SEQUENCE SPACES

Proof. We first prove that `p , ` . To see that `p is a subspace of ` , let x `p . Then

there exists some constant R such that
!1/p

X
|xi |p
=
i=1

Taking the pth power of both sides yields the following comparison:
p

||x|| = sup |xi | sup |xi |

iN

iN

|xi |p = p <

i=1

and so x ` . Furthermore, for any i N we have

p

|xi |

|xi |p

i=1

and so `p , ` . We now prove the general case. We may write r = p + s so that for all
iN
||x||rr = |xi |r = |xi |p |xi |s |xi |p sup |xi |s |xi |p ||x||s ||x||pp ||x||sp = ||x||rp
iN

Taking the 1/rth power across this inequality yields ||x||r ||x||p and so `p , `r .

1.3

Separability of `p spaces

Definition 1.14. Let A and B be two sets. We say that A and B have the same cardinality
if there exists a bijection f : A B. We say that A is finite if there exists some n N
such that A has the same cardinality as { 1, . . . , n }. We say that A is countable if it is
either finite or has the same cardinality as N.
Theorem 1.15. Let A be a set and X A. If A is countable then X is countable.
Proof. It suffices to show that any subset A N is countable. Let
a1 = min A, A1 = A\ { a1 }
a2 = min A1 , A2 = A1 \ { a2 }
..
.
If at some stage, An is empty then A is finite. If not then the map n 7 an is a bijection
between N and A.
Theorem 1.16. Let A be a set. If there exsists a surjection f : N A then A is countable.
Proof. Consider a A and let g(a) = min f 1 ({ a }) and set B = { g(a) | a A }. Clearly
B N and g is a bijection between B and A. The theorem then follows from Theorem
1.15.
Theorem 1.17. A countable union of countable sets is countable.

CHAPTER 1. SEQUENCE SPACES

Proof. Let An be countable sets where n N and N is itself countable. First assume that
(n) (n)
N, A1 , . . . are all infinite. Then we may assume
that N = N. Suppose An = { a1 , a2 , . . . }.
S
We can enumerate the elements of A = n=1 An as follows (in a diagonal fashion):
(1)

(1)

(2)

(1)

(2)

(3)

(1)

a1 , a2 , a1 , a3 , a2 , a1 , a4 , . . .
This establishes a surjection f : N A. Now by Theorem 1.16, A is countable. In the case
that any of the sets N, A1 , . . . are finite, we can add extra elements to these sets to make
them infinite.
Corollary 1.18. If A and B are countable then the Cartesian product A B is countable.
Proof. Note that we can write
AB =

A {b}

bB

This is a countable union of countable sets so by 1.17, it is countable.

Example 1.19. Z and Q are countable.
Theorem 1.20. Let X be non-empty. Then the power set 2X of X cannot have the same
cardinality as X.
Proof. Assume there exists a bijection f : X 2X . Consider the set
S = {x X | x
/ f (x) } X
Now let s X be such that f (s) = S. First suppose that s S. then s
/ f (s). But this
contradicts the definition of s. Now suppose that s
/ S. We have that s f (s). Again,
this contradicts the definition of s. Hence there can exist no such bijection.
Corollary 1.21. R and N have different cardinalities.
Proof. Consider the interval [0, 1]. We claim that 2N has the same cardinality as [0, 1]. Let
x [0, 1]. Then x can be written uniquely in the form of a binary expansion 0.a1 a2 a3 . . .
where aj { 0, 1 }. Such a binary expansion uniquely defines a subset of N. This establishes
a bijection between [0, 1] and 2N .
Definition 1.22. Let X be a Banach space. We say that X is separable if it contains a
countable dense subset.
Example 1.23. Since Q is countable and dense in R, the space R is separable.
Theorem 1.24. Let p < . Then `p is separable. ` is not separable.
Proof. First suppose p < . Let An `p be the set of all sequences x `p such that
all coordinates of x are rational and xk = 0 for all k > n. Furthermore, let A =
n=1 An .
Clearly each An is countable and hence A is itself countable. We claim that A is dense in
`p .
To this end, let y `p \A. We need to exhibit an x A such that x is arbitrarily close
to y. Let n N and denote Pn : `p `p to be the mapping
Pn (y) = (y1 , y2 , . . . , yn , 0, 0, . . . )

CHAPTER 1. SEQUENCE SPACES

Given any > 0, we can clearly find an n N such that ||y Pn (y)||p 2 .
Now since Q is dense in R, it follows that the set of all complex rationals is dense in C. It is
thus clear that we can find an x An such that ||Pn (y) x||p 2 . Now using the triangle
inequality, we have that there exists an n N such that ||y x||p < . Since is arbitrary,
we can always find an x arbitrarily close to y and we are done.
We now show that ` is not separable. Let X N and define eX ` as follows: eX
n = 1
p
=
0
if
n

/
X.
Then
the
number
of
such
elements
of
`
is
uncountable
and,
if n X and eX
n
X
Y
for any X 6= Y , we have ||e e || = 1.
Now suppose that D ` is a countable dense subset. Consider the balls B 1 (eX ) over
2
all X N . Since 2N is uncountable, this collection of balls is uncountable. Now, these
balls are disjoint and each contain an element of D since D is dense. Hence there exists a
bijection between a subset of D (which is countable) and the set of all such balls (which is
uncountable). This is a contradiction and hence there can exist no such set D.

Chapter 2
Lebesgue integration and Lp spaces
2.1

Riemann Integral

Definition 2.1. Fix an interval [a, b] R. We define a step function to be a finite linear
combination of characteristic functions of bounded intervals:
X
cn n (x)
f (x) =
n

Definition 2.2. Let f be a step function. We define the integral of f as follows:

Z b
X
cn (n )
f (x) dx =
a

where () is understood to be the length of the interval .

There are multiple issues with the Reimann integral:
1. If fn f pointwise then we cannot conclude that

fn

f.

2. The class of Riemann integrable functions R[a, b] contains many functions which, intuitively, should be integrable
3. C[a, b] with the norm
Z
||f ||1 =

|f (x)|dx
a

is not complete. Its completion is much larger than R[a, b] so we would like to be able
to better describe it.

2.2

Lebesgue measure

Definition 2.3. Let X be a set and a collection of subsets of X such that

1. X
2. is closed under complements
3. is closed under countable union
9

CHAPTER 2. LEBESGUE INTEGRATION AND LP SPACES

10

Then we say that is a -algebra over X.

Definition 2.4. We define the Borel -algebra of R, denoted B(R) to be the -algebra of
R containing every open intervals. Any set in B(R) is called a Borel set.
Example 2.5. Let S be the set of all irrational numbers whose continued fractions are of
the form
a0 +

1
a1 +

1
a2 +

..

such that there exists a sequence 0 < i1 < i2 < . . . where each aik divides aik+1 . Then
S 6 B(R).
Definition 2.6. We define the Lebesgue measure to be the function : B(R) R given
by
X
(B) =
(bi ai )
i

when B is a countable union of some open intervals (ai , bi ). Otherwise it is defined as

(B) = inf (O)
BO

where the infimum is taken over open sets O containing B.

Theorem 2.7. Let An be a countable collection of mutually disjoint Borel sets. Then
X
(An )
(n An ) =
n

In other words, the Lebesgue measure is countably additive.

Definition 2.8. A measure space is a triple (, A, ) where is a set, A is a -algebra
over and is a countably additive function on A.

2.3

Borel functions

Definition 2.9. Let f : R R be a function. We say that f is a Borel function if

f 1 [(a, b)] is a Borel set for any interval (a, b) R.
Remark. Clearly, if f is continuous then f is Borel.
Proposition 2.10. A function f is Borel if and only if for all B B(R), the set f 1 [B] is
Borel.
Proof. First suppose that f is Borel. Let X be the set
X = { B B(R) | f 1 (B) B(R) }
Clearly, X B(R). We claim that X is a -algebra that contains the open intervals. By
the minimality of B(R), it would then follow that X = B(R). Since f is Borel, X necessarily
contains the open intervals. We now check the axioms of a -algebra one by one.

CHAPTER 2. LEBESGUE INTEGRATION AND LP SPACES

11

We have B(R) and f 1 () = and so X.

Let B X. Then B B(R) and f 1 (B) B(R). Now, B(R) is a -algebra and so
B c B(R). We have f 1 (B c ) = f 1 (B)c B(R) whence B c X.
Let { An } be a countable collection S
of elements of X. Then each An B and f 1 (An )
B. Since B(R) is a -algebra, we have
i=1 Ai B. Hence
!

[
[
f 1
Ai =
f 1 (Ai ) B(R)
i=1

i=1

Thus X is a -algebra. Furthermore, since f is Borel, X necessarily contains all open

intervals whence X = B(R).
Conversely, suppose that for all B B(R) we have f 1 (B) B(R). Then clearly for all
open intervals (a, b) R we have f 1 ((a, b)) B(R) and so f is Borel.
Corollary 2.11. Let f and g be Borel functions. Then the following functions are Borel:
1. f g
2. |f |
3. f = max{f, 0}
Proof. Let (a, b) R. Since f is Borel, we have g 1 ((a, b)) = B for some Borel set B. Since
f is Borel, we have f 1 (B) = B 0 for some Borel set B 0 . Hence (f g)1 ((a, b)) = B 0 whence
f g is Borel.
Now, the absolute value function is Borel since it is continuous. Hence, by the first part
of the corollary, |f | is Borel.
To show that f is Borel, consider the function g sending x to max{x, 0}. x is clearly
a continuous function as is the zero function. Since the maximum of any two continuous
functions is continuous, it follows that g is continuous. Therefore, g is Borel. Appealing to
the first part of the corollary, we see that f = f g is Borel.
Theorem 2.12. Let fn be a sequence of Borel functions. If fn (x) f (x) for all x as
n then f is Borel.
Remark. The previous theorem is not necessarily true for C[a, b] and R[a, b].
Definition 2.13. Let B be a set and A B a subset. We define the characteristic
function of A to be:

1 if x A
A (x) =
0 if otherwise
Definition 2.14. Let f : R R be a function. We say that f is a simple function if it is
a linear combination of characteristic functions of Borel sets. The vector space of all simple
functions over R is denoted Simp(R).
Proposition 2.15. Let f be a simple function. Then f is Borel.
Proof. Given c R and B be a Borel set, it suffices to show that
f (x) = cB (x)
is Borel. Let (a, b) R be an interval. First suppose that 0, c B. Then f 1 ((a, b)) =
R B(R). Next suppose that 0 B and c 6 (a, b). Then f 1 ((a, b)) = R\B B(R). Now
suppose that 0 6 (a, b) but c (a, b). Then f 1 ((a, b)) = B B(R). Finally, suppose that
neither 0 (a, b) nor c (a, b). Then f 1 ((a, b)) = B(R). Hence in all cases, we see
that f 1 ((a, b)) is a Borel set whence f is Borel.

CHAPTER 2. LEBESGUE INTEGRATION AND LP SPACES

12

Proposition 2.16. Let f be a positive Borel function. Then there exists a sequence of
simple functions fn such that fn+1 fn and fn f for all n and fn (x) f (x) for all x as
n . Furthermore, if f is bounded then fn can be chosen so as to converge uniformly to
f.
Corollary 2.17. A linear combination of Borel functions is a Borel function.
Proof. Let af +bg be a linear combination of Borel functions for some a, b R. Since b and g
are Borel, Proposition 2.16 implies that there exists monotone sequences of simple functions
{ fn } and { gn } such that fn f and gn g for all n and fn (x) f (x), gn (x) g(x) for
all x as n . Then afn + bgn is a sequence of simple functions and are thus also Borel.
It clearly converges to af + bg. By Theorem 2.12, af + bg is Borel.

2.4

Lebesgue Integral

Definition 2.18. We define the Lebesgue integral to be the functional given by

Z
f (x) d =
R

when f =

Pn

i=1 ci Bi

n
X

ci (Bi )

i=1

is a simple function. If f 0 is a Borel function then it is defined by

Z
f (x) d = sup g(x) d
R

gSimp(R)
gf

R
Proposition
R
R 2.19. Let f be a Borel function. Then |f | d is finite if and only if both
f+ d and f d are finite.
R
R
Proof. First suppose that |fR | d is finite. Then it is clear that f d are finite.
Conversely, suppose that f d is finite. Then
Z
Z
|f | d = f+ f d
Z
Z
Z
Z
Z
= sup
g d sup
g d sup
h d f+ d f d <
gSimp(R)
gf+ f

gSimp(R)
gf+

hSimp(R)
hf+

Definition 2.20.
Denote by L1 (a, b) the class of all Borel functions f on (a, b) such that
R
the integrals f dx are both finite. For f L1 we define the Lebesgue integral of f by
Z
Z
Z
f d = f+ d f d
Theorem 2.21 (Montonone Convergence Theorem). Let fn 0 be a sequence of Borel
Rfunctions on R such that fn+1 fn1 for all n and fn (x) f (x) for all x as n . If
fn (x) dx C for all n then f L (R) and
Z
Z
fn (x) d f d

CHAPTER 2. LEBESGUE INTEGRATION AND LP SPACES

13

Theorem 2.22 (Dominated Convergence Theorem). Let fn be a sequence of Borel functions

on R such that fn (x) f (x) for all x as n and fn (x) F (x) for all n and x and
some F L1 (R) then
Z
|fn (x) f (x)| 0
as n and thus
Z

Z
|fn (x)| d

|f | d

as n .
Theorem 2.23. Let f, g L1 (a, b). Then
R
R
1. f g whenever f g
R
R
2. | f | |f |
R
3. | f | (b a)||f ||
R
R
R
4. (af + bg) = a f + b g
where ||f || = supx(a,b) |f (x)|.
Proof.
Part 1: Suppose that f g. Then
Z
Z
f d = sup
h1 d
h1 Simp(R)
h1 f

Z
sup

Z
h2 d =

g d

h2 Simp(R)
h2 g

Part 2: We have that




Z
Z
 

Z
Z
Z



 


 f d =  sup


g d sup  g d sup
|g| d = |f | d

 
 gSimp(R)
gSimp(R)
gSimp(R)
 gf

gf
gf
for simple functions g f .
Part 3: We have that
Z
 Z


 f d |f | d =



Z
sup

|g| d

gSimp(R)
gf

P (g)
(g)
(g)
Now if g = n cn Bn(g) for some cn R and Borel sets Bn , we have
Z

Z
X
X


(g)
(g)
 f d sup
|g|
dx
=
sup
|c
(B
)|

sup
|c(g)
n
n
n (b a)|


gSimp(R)
gf

gSimp(R)
gf

gSimp(R)
gf

= (b a)

sup

gSimp(R)
gf

(b a) sup |f |
x(a,b)

= (b a)||f ||

|c(g)
n |

CHAPTER 2. LEBESGUE INTEGRATION AND LP SPACES

14

Part P
4: We first prove that
P the integral is a linear functional for simple functions. Let
= ni=1 cn Cn and = m
j=1 dn Dn (for collections of pairwise disjoint sets Ci and Di )
be simple functions and a, b R. We have that
Z
n X
m
X
(a + b) d =
(aci + bdj )Ci Dj
i=1 j=1
n X
m
X

=a
=a
=a

ci Ci Dj + b

i=1 j=1
n
m
X
X

ci

i=1
n
X

Ci Dj + b

j=1

ci Ci + b

i=1

m
X

dj Ci Dj

i=1 j=1
m
n
X
X

dj

j=1

Ci Dj

i=1

dj Di

j=1

Z
=a

n X
m
X

Z
d + b

Now let f and g be Borel functions. By Lemma 2.16 we can always find a sequence of
increasing simple functions fn and gn such that fn f and gn g. Then by the Monotone
Convergence Theorem, we have that
Z
Z
Z
Z
a f d + b g d = lim
afn d + lim
bgn d
n
n
Z

Z
= lim
afn d + bgn d
n

Now fn and gn are all simple functions which are linear by the previous claim whence

Z
Z
Z
afn + bgn d
af d + bg d = lim
n

Now again by the monotone convergence theorem, the limit on the right is equal to
bg d.

af +

Theorem 2.24 (Fatous Lemma). Let fn be a sequence of non-negative Borel functions.

Then
Z
Z
lim inf fn (x) d = lim inf fn (x) d
Proof. Let hm (x) = inf nm { fn (x) }. Clearly, hm (x) fm (x) and hm (x) hm+1 (x) for all
x and m. We first observe that
Z
Z
hm (x) dx fn (x)
for all n m. This is equivalent to
Z
Z
hm (x) dx inf
fn (x)
nm

CHAPTER 2. LEBESGUE INTEGRATION AND LP SPACES

15

Now, by the Monotone Convergence Theorem

Z
Z
lim inf fn (x) dx =
lim hm (x) dx
m
Z
= lim
hm (x) dx
m
Z
lim inf fn (x) dx

Definition 2.25. Let f : C C be a complex valued function. We define the Lebesgue

integral of f to be
Z
Z
Z
f d = Re(f ) d + i Im(f ) d

2.5

The Lp spaces

Definition 2.26. Let p > 1 and (, A, ) a measure space. We denote by Lp () the

collection of measurable functions f which satisfy
Z
||f ||p =

|f (x)|p d(x)

 p1
<

Example 2.27. Lp (R) where A = B(R) and is the Lebesgue measure.

Proposition 2.28. Let (, A, ) be a measure space and f Lp () integrable. Then
||f ||p = 0 if and only if f = 0 almost everywhere.
Proof. First suppose that ||f ||p = 0. Denote
A = { x | f (x) 6= 0 }
S
Let An = { x | |f (x)| > 1/n }. Then, clearly, A = n An . Now suppose, for a contradiction, that A has strictly positive measure. Then at least one of the An has strictly positive
measure, say Ak . Then |f (x)| dominates 1/kAk on Ak . We then have that
Z
Z
1
1
p
Ak d = p (An ) > 0
0=
|f | d >
p
k
Ak
Ak k
which is a contradiction.
Now suppose that f = 0 almost everywhere. Then for any simple function g satisfying
0 g |f |p , we must have that g is 0 almost everywhere. Let such a g have representative
P (g) (g)
(g)
(g)
n cn Bn for some cn R and Borel sets Bn . Then
Z
(g)
|f |p d = sup c(g)
n Bn = 0
gSimp(R)
g|f |p
(g)

since, given any n, either cn = 0 or (Bn ) = 0.

Proposition 2.29. Let (, A, ) be a measure space. Then Lp () is a vector space over R
(or C).

CHAPTER 2. LEBESGUE INTEGRATION AND LP SPACES

16

Proof. Let f, g Lp () and R. Then f + g is a measurable function. Hence

Z
Z
p
|f + g| d (|f | + |||g|)p d
Z
(2 max{|f |, |||g|})p d
Z
p
=2
max{|f |p , |||g|p } d
Z
p
2
|f |p + |||g|p d
Z

Z
p
p
p
|f | d + |||g| d <
=2
and so f + g Lp (). The rest of the vector space axioms are clear from the basic
properties of functions and thus Lp () is a vector space.
Proposition 2.30 (Holders Inequality for Lp spaces). Let p, q R be such that p, q > 1 and
1/p + 1/q = 1. If (, A, ) is a measure space and f Lp (), g Lq () then f g L1 ()
and
||f g||1 ||f ||p ||g||q
Proof. First assume that either ||f ||p = 0 or ||g||q = 0. It then follows that f g is equal to 0
almost everywhere. Holders Inequality then follows immediately in this case.
Hence, we may assume that neither ||f ||p and ||g||q are zero. Let u = f /||f ||p and
v = g/||g||q . We claim first that ||uv||1 ||u||p ||v||q = 1. By Youngs Inequality we have,
for all x R,
|u(x)|p |v(x)|q
+
|u(x)v(x)|
p
q
Passing to the Lebesgue integral we have
Z
Z
Z
1
1
p
|u| d +
|v|q d = 1
|uv| d
p
q
We then have that




fg


 ||f ||p ||g||q  1
1

whence
||f g||1 ||f ||p ||f ||q

Proposition 2.31 (Minkowskis Inequality for Lp spaces). Let (, A, ) be a measure space

and p > 1. Suppose that f, g Lp (). Then
||f + g||p ||f ||p + ||g||p

CHAPTER 2. LEBESGUE INTEGRATION AND LP SPACES

17

Proof. If p = 1 then Minkowskis Inequality follows directly by the properties of the Lebesgue
integral hence suppose p > 1. Let q R be such that 1/p + 1/q = 1. We have that
|f + g| |f | + |g|
Multiplying this inequality through by |f + g|p1 we have
|f + g|p |f ||f + g|p1 + |g||f + g|p1
Now, (p 1)q = p whence (f + g)p1 Lp () so we may pass to the Lebesgue integral to
get
||f + g||pp ||(|f ||f + g|p1 )||1 + ||(|g||f + g|p1 )||1
Applying Holders Inequality yields
||f + g||pp ||f ||p ||(f + g)p1 ||q + ||g||p ||(f + g)p1 ||q
Now note that
p1

||(f + g)

Z
||q =

q(p1)

|f + g|

1/q
1/q Z
p
|f + g| d
= ||f + g||p/q
d
=
p

and so
p/q
||f + g||pp ||f ||p ||f + g||p/q
p + ||g||p ||f + g||p

Since p p/q = 1/p and we may assume that ||f + g||p > 0 it follows that
||f + g||p ||f ||p + ||g||q

Corollary 2.32. Let (, A, ) be a measure space. Then Lp () is a semi-normed1 vector

space over R (or C).
Definition 2.33. Let (, A, ) be a measure space and p 1. Define an equivalence relation
on Lp () by f g if and only if f = g almost everywhere. We define the Lp () space to
be the collection of equivalence classes of . By an abuse of notation, for any equivalence
class [f ] Lp () we shall write f Lp () to be one of its representatives (if a continuous
represntative exists then we shall usually chose that one).
Proposition 2.34. Let (, A, ) be a measure space and p 1. Then Lp () is a normed
vector space over R (or C).
Proof. Lp () is clearly a semi-normed vector space over R - this follows directly from the
results for Lp (). To see that || ||p is a norm on Lp (), we just need to prove that ||f || = 0
if and only if f = 0. Now, f = 0 means that f is the equivalence class containing 0. This
equivalence class contains all functions that are 0 almost everywhere. We may thus choose
such a representative, such as the 0 function, to see that ||f ||p = 0.
1

recall that a semi-normed space is a set equipped with a so-called semi-norm which satisfies all properties
of a norm except ||x|| = 0 x = 0.

CHAPTER 2. LEBESGUE INTEGRATION AND LP SPACES

18

Proposition 2.35. Let (, A, ) be a measure space and p 1. Then Lp () is a Banach

space.
Proof. We need to show that Lp () is complete. Let { fn }nN be a Cauchy sequnce in
Lp (). Let { fnk }kN be a subsequence of { fn } such that
||fnk+1 fnk ||p 2k
for all k 1. Now consider the functions

X
f (x) = fn1 (x) +
(fnk+1 (x) fnk (x))
i=1

g(x) = |fn1 (x)| +

|fnk+1 (x) fnk (x)|

i=1

and denote their partial sums by SK (f ), SK (g). We have

||SK (g)||p ||fn1 (x)||p +

K
X

||fnk+1 fnk ||p ||fn1 (x)||p + 2k

i=1

Clearly, there exists a constant C R such that ||SK (g)||p < C for all K 1. Furthermore,
SK+1 (g) SK (g) for all K 1 and SK (g) g pointwise. Appealing to the Montonone
Convergence Theorem, we have that g Lp (). It then follows that f Lp ().
We now claim that f is the limit of { fn }. Observe that SK1 (g) = fnK so fnK f
pointwise as K . We also show that fnK f in Lp (). We have that
|f (x) SK (f )(x)|p (2 max{|f (x)|, |SK (f )(x)|})p
2p |f (x)|p + 2p |SK (f )(x)|p
2p+1 |g(x)|p
Appealing to the Dominated Convergence Theorem, we see that ||f fnK ||p 0 as k .
Finally, since { fn } is Cauchy, for all > 0, there exists an N N such that for
all n, m > N we have ||fn fm ||p /2. Now choose nK such that nK > N . Then
||fnK f ||p /2. By the triangle inequality, we then have that
||fn f ||p ||fn fnK ||p + ||fnK f ||p
and we are done.
Proposition 2.36. Let (, A, ) be a measure space and p 1. Then the collection of all
simple functions in Lp () is dense in Lp ().
Proof. Let f Lp (). For all > 0, it suffices to exhibit a simple function g such that
||f g||p < .
Recall that every measurable function f can be approximated pointwise by a sequence
of simple functions { fn } such that fn+1 fn . We clearly have that fn (x) f (x) for all
x and n. By the Dominated Convergence Theorem, we thus have that ||fn f ||p 0 as
n . Hence for all > 0, there exists an n such that ||fn f ||p < . Therefore, the
simple functions are dense in Lp ().
Proposition 2.37. The collection of all continuous functions in Lp (a, b) is dense in Lp (a, b)
for all intervals (a, b) R (including infinite intervals).

CHAPTER 2. LEBESGUE INTEGRATION AND LP SPACES

19

Proof. By 2.36, it suffices to show that the characteristic function of any Borel set B can
be approximated arbitrarly well by a continuous function. In other words, for all > 0, we
need to show that there exists a continuous function g Lp (R) such that ||B g||p < .
To this end, fix > 0, choose an open set G B such that (G\B) < . Let f (x)
denote the function
f (x) =

d(x, G)
d(x, G) + d(x, B)

where d(x, C) := inf yC |x y|. Since d is continuous, so is f (x). Now if x B then

d(x, B) = 0 and, consequently, f (x) = 1. Now if x
/ G then f (x) = 0 by definition of
d(x, C). Hence f (x) B (x) = 0 for all x B and x G and |f (x) B (x)| < 1 for all
x G\B. Thus
Z
|f (x) B (x)|p d < (G\B) <

Corollary 2.38. Lp (a, b) is separable for any (possibly unbounded) interval (a, b) R.
Proof. We prove the corollary for Lp (R). By 2.5, the continuous functions are dense in Lp (R).
By the Stone-Weierstrass Theorem, we can approximate continuous functions arbitrarily well
with polynomial functions. Now, we can approximate any polynomial arbitrarily well with
polynomials with rational coefficients. The latter collection is clearly countable and dense
in Lp (R) so Lp (R) is separable.
Definition 2.39. Let f be a Borel function. We define the essential supremum of f to
be
ess sup f = sup { t | ({ x | f (x) > t }) > 0 }
Remark. Let f be a Borel function and L = ess sup f . Clearly, if t < L then ({ x | f (x) > t }) >
0. If t > L then ({ x | f (x) > t }) = 0.
Proposition 2.40. Let f be a Borel function. Then ({ x | f (x) > ess sup f }) = 0.
Proof. Let L = ess sup f . If L = then the proposition is clear so assume that L is finite.
Note that { x | f (x) > ess sup } = f 1 ((L, ]). We have that
f

((L, ]) =

f 1 ((L + 1/k, ])

k=1

By the definition of ess sup, each set in this union has measure zero whence f 1 ((L, ]) has
measure zero.
Definition 2.41. Denote by L (a, b) the collection of all Borel functions f on (a, b) such
that there exists M R with |f (x)| M almost everywhere. Define the function || || :
L (a, b) R by
||f || = ess sup |f (x)|
Define an equivalence relation on L (a, b) where f g if and only if ||f g|| = 0 almost
everywhere. We denote by L (a, b) the collection of all equivalence classes of .

CHAPTER 2. LEBESGUE INTEGRATION AND LP SPACES

20

Proposition 2.42. L (a, b) is a normed vector space over R (or C) for all (possibly unbounded) intervals (a, b) R.
Proof. We first show that L (R) is a vector space over R. To this end, let R and
f, g L (R). We need to show that f + g L (a, b). That is to say, we need to show
that there exists M R such that
|f + g| M
almost everywhere. By hypothesis, there exists M1 , M2 R such that |f | M1 and |g| < M2
almost everywhere. Hence
|f + g| |f | + |||g| M1 + ||M2
almost everywhere whence f + g L (R). The rest of the vector space axioms follow
directly from the basic properties of functions and so L (R) is a vector space.
For the norm axioms, we first prove homogeneity. Let c R and f L (R). We have
that
||cf || = ess sup |cf | = sup { t | ({ x | |cf (x)| > t } > 0) } = |c|||f ||
We next prove the triangle inequality. Let f, g L (R). Then there exist sets of
measure zero X, Y R such that |f (x)| ||f (x)|| for all x R\X and |g(x)| ||g(x)||
for all x R\Y . Then X Y is again a set of measure zero. Then for all x R\(X Y )
we have
||f + g|| = ess sup |f (x) + g(x)| |f (x) + g(x)| |f (x)| + |g(x)| ||f || + ||g||
Now if f L (R) then it is clear that ||f || 0. Now suppose that ||f || = 0. By
Proposition 2.40, we have that { x | |f (x)| > 0 } = { x | f (x) 6= 0 } has measure zero. Hence
f is zero almost everywhere whence f is the equivalence class of the 0 function and we are
done.
Proposition 2.43. L (a, b) is a Banach space for any (possibly unbounded) interval (a, b)
R.
Proof. We shall prove the proposition for L (R). It suffices to show that L (a, b) is complete. To this end, let { fn } be a Cauchy sequence in L (R). That is to say, for all > 0,
there exists N N such that for all n, m > N we have ||fm fn || < . This is equivalent to
there existingSsets of measure zero Ym,n such that for all x R\Ym,n we have |fm fn | < .
Denote Y = m,n Ym,n . Then Y has measure zero and for all x R\Y we have |fm fn | <
for all m, n > N .
Hence for all x R\Y , { fn (x) } is a Cauchy sequence in R. Since R is complete, this
sequence converges pointwise to some limit f (x). This limit f is defined outside of the
measure zero set Y . For x Y , we may take f (x) = 0. In other words, f = limn R\Y fn .
Note that this function is measurable. In the Cauchy sequence condition, we may let n
so that for all m N we have
|fm (x) f | <
But then
||fm (x) f || = ess sup |fm (x) f | |fm (x) f | <

CHAPTER 2. LEBESGUE INTEGRATION AND LP SPACES

21

so that fm (x) f as m with respect to the L norm.

Finally, we show that f L (R). We have that
||f || = ||f + fm fm || ||f fm || + ||fm || ||fm || + <
and so f L (R).
Proposition 2.44. L (a, b) is not separable for any (possibly unbounded) interval (a, b)
R.
Proof. Consider the collection of functions (c,d) over the intervals (c, d) (a, b). This
collection is clearly uncountable. Furthermore, if c 6= c0 and d 6= d0 then ||(c,d)) (c0 ,d0 ) || =
1.
Now suppose there exists a dense countable subset D L (a, b). Consider the balls
B 1 ((c,d) ) around each interval. These balls are clearly disjoint and uncountable. Since D
2
is dense, there must be an element of D in each such ball. But this is a contradiction as it
implies the existence of a surjection from a countable set onto an uncountable set.
Proposition 2.45. Let p, q 1 be such that p > q. Then Lp (a, b) , Lq (a, b) for any
bounded interval (a, b).
Proof. Assume that f Lp (a, b). Since p > q, we have q/p < 1. Observe that q/p + (1
q/p) = 1. Let u = p/q and v = 1/(1 q/p) so that 1/u + 1/v = 1. By Holders Inequality,
we have that
q/p
Z
q(p/q)
q
q
q
|f |
d
(b a)1q/p
||f ||q = || |f | ||1 = || |f | 1 ||1 || |f | ||u ||1||v =
Taking the q th root across this inequality we have that
||f ||q (b a)1/q1/p ||f ||p
as desired.
Proposition 2.46. Let r 1 and f Lr (a, b) for some bounded interval (a, b). Then
lim ||f ||p = ||f ||

Proof. Let t [0, ||f || ). By definition,

A = { x | |f (x)| t }
is a set of positive measure. We then have the following inequality:
Z
1/p
p
||f ||p
|f |
A
p

(t (A))1/p
= t(A)1/p
If (A) is finite then (A)1/p 1 as p . If (A) is infinite then (A)1/p is infinite for
all p. In either case, we have
lim inf ||f ||p t
p

CHAPTER 2. LEBESGUE INTEGRATION AND LP SPACES

22

Now t is arbitrary, we have

lim inf ||f ||p ||f ||
p

For the reverse inequality, note that |f (x)| ||f || for almost all x. Then for all p r
we have
Z
1/p Z
1/p Z
1/p
p
r
pr
r
pr
1r/p
||f ||p =
|f | d
=
|f | |f |
d

|f | ||f || d
= ||f ||
||f ||r/p
r
r/p

Now, ||f ||r

< for all p and so

1r/p
lim sup ||f ||p lim sup ||f ||
||f ||r/p
= ||f ||
r
p

Chapter 3
Hilbert spaces
3.1

The geometry of Hilbert spaces

Definition 3.1. Let V be a vector space over C. We say that V is an inner product
space if there is a complex valued function (, ) on V V such that, given any x, y, z V
and C,:
1. (x, x) 0 and (x, x) = 0 if and only if x = 0
2. (x, y + z) = (x, y) + (x, z)
3. (x, y) = (x, y)
4. (x, y) = (y, x)
(, ) is referred to as an inner product.
Example 3.2. Cn is an inner product space. Given x = hx1 , . . . , xn i and y = hy1 , . . . , yn i,
we define their inner product as
(x, y) =

n
X

xj y j

j=1

Example 3.3. Let C[a, b] denote the complex-valued continuous functions on the interval
[a, b]. Given f (x), g(x) C[a, b], we define their inner product as
Z b
(f, g) =
f (x)g(x) dx
a

Definition 3.4. Let V be an inner product space and x, y V . We say that x and y are
orthogonal if (x, y) = 0. A collection of elements { xi } V is said to be an orthonormal
set if (xi , xi ) = 1 for all i and (xi , xj ) = 0 for all i 6= j.
Remark. Let
pV be an inner product space with inner product given by (, ). If x V we
write ||x|| = (x, x).
Theorem 3.5 (Pythagorean theorem). Let V be an inner product space and x1 , . . . , xN V
be an orthonormal set. Then, given any x V , we have that

2
N
N


X
X



||x||2 =
|(x, xn )|2 + x
(xn , x)xn 


n=1

n=1

23

CHAPTER 3. HILBERT SPACES

24

Proof. We first claim that

N
X

(xn , x)xn ,

N
X
(xn , x)xn

n=1

n=1

are orthogonal. We have that

N
N
X
X
(xn , x)xn , x
(xn , x)xn
n=1

!
!
N
N
N
X
X
X
(xn , x)xn , x
(xn , x)xn ,
(xn , x)xn

n=1

n=1

n=1

{z
A

{zn=1
B

Firstly, we have
A=

N
X

(xn , x)(xn , x)

n=1

Secondly, we have
B=

N X
N
X

(xi , x)(xj , x)(xi , xj )

i=1 j=1

N
X

(xi , x)(xi , x)

i=1

where we have used the fact that (xi , xj ) = 0 for all i 6= j. Hence A B = 0 as desired. It
then follows that
2

2  
N
N

X


X




(xn , x)xn 
(x, x) =  (xn , x)xn  + x




n=1
n=1

 2
N
N


X
X


2
=
|(xn , x)| + x
(xn , x)xn 



n=1
n=1
Corollary 3.6 (Bessels inequality). Let V be an inner product space and x1 , . . . , xN V
an orthonormal set. Then, given any x V , we have
2

||x||

N
X

|(x, xn )|2

n=1

Corollary 3.7 (Cauchy-Schwarz inequality). Let V be an inner product space and x, y V .

Then
|(x, y)| ||x|| ||y||
y
. This
Proof. This is clearly true for y = 0 hence assume y 6= 0. Consider the vector ||y||
vector by itself clearly forms an orthonormal set. Hence we may apply Bessels inequality
to any x V so that



y 
2

||x||  x,
||y|| 

|(x, y)|2
||y||2

whence the result follows by muptiplying through by ||y||2 and taking the square root across
the inequality.

CHAPTER 3. HILBERT SPACES

25

Theorem 3.8. Let Vpbe an inner product space. Then V is a normed linear space with
norm given by ||x|| = (x, x).
Proof. The first two axioms of a norm are satisfied directly from the first four properties of
the inner product. It remains to show that the triangle inequality holds. We have that
||x + y||2 = (x, x) + (x, y) + (y, x) + (y, y)
= (x, x) + 2Re(x, y) + (y, y)
(x, x) + 2|(x, y)| + (y, y)
p
p
(x, x) + 2 (x, x) (y, y) + (y, y)
= (||x|| + ||y||)2
where in the second to last inequality we have used the Cauchy-Schwarz inequality. The
result then follows by taking the square root across the inequality.
Remark. This norm naturally induces the metric
p
d(x, y) = (x y, x y)
from which we can introduce convergence, completeness and density to the inner product
space.
Proposition 3.9 (Parallelogram Law). Let V be an inner product space with inner product
given by (, ). Then for all x, y V we have
||x + y||2 + ||x y||2 = 2(||x||2 + ||y||2 )
Proof. We have that
||x + y||2 + ||x y||2 = (x + y, x + y) + (x y, x y)
= (x + y, x) + (x + y, y) + (x y, x) (x y, y)
= (x, x) + (x, y) + (y, x) + (y, y) + (x, x) (x, y) (y, x) + (y, y)
= 2(x, x) + 2(y, y)
= 2(||x||2 + ||y||2 )
Proposition 3.10 (Polarisation Identity). Let V be an inner product space with inner
product given by (, ). Let x, y V . If V is an R-vector space then
1
(x, y) = (||x + y||2 ||x y||2 )
4
If V is a C-vector space then
1
(x, y) = (||x + y||2 ||x y||2 + i||x + iy||2 i||x iy||2 )
4
Proof. We shall verify only the real case, the complex case follows by a similar argumentation. Expanding the right hand side of the claimed identity gives
1
1
[(x + y, x + y) (x y, x y)] = [(x + y, x) + (x + y, y) (x y, x) + (x y, y)]
4
4
1
= [(x, x) + (x, y) + (y, x) + (y, y) (x, x) + (x, y)
4
+ (y, x) (y, y)]
1
= [2(x, y) + 2(y, x)]
4
= (x, y)

CHAPTER 3. HILBERT SPACES

26

Theorem 3.11. Let V be a normed vector space over R (or C) with norm given by || ||.
Then || || is induced by an inner product if and only if the Parallelogram Law holds in V .
Proof. First suppose that || || is induced by an inner product. Then V is clearly an inner
product space and Proposition 3.9 implies that the Parallelogram Law holds.
Now suppose that the Parallelogram Law holds. We shall only prove the case where V
is a R-vector space. Define (x, y) by the Polarisation Identity
1
(x, y) = (||x + y||2 ||x y||2 )
4
We claim that (, ) is an inner product that induces || ||. First note that
1
1
(x, x) = (||x + x||2 ) = (4||x||2 ) = ||x||2
4
4
and so (, ) induces || ||.
We now check each axiom of inner products. We clearly have that (x, x) 0 with
equality if and only if x = 0. It is also clearly symmetric. To prove linearity in the second
argument, let x, y, z V . By the Parallelogram Identity, we have
1
(x, y + z) = (||x + y + z||2 ||x (y + z)||2 )
4
1
= (2||x + y||2 + 2||z||2 ||x + y z||2 2||y||2 2||x z||2 + ||x + y z||2 )
4
1
= (2||z||2 2||y||2 + 2||x + y||2 2||x z||2 )
4
Now note that these expressions should be symmetric in y, z. In other words, we have
||x + y + z||2 ||x (y + z)||2 = 2||z||2 2||y||2 + 2||x + y||2 2||x z||2
= 2||y||2 2||z||2 + 2||x + z||2 2||x y||2
Adding these two expressions together and dividing by two, we have
||x + y z||2 ||x (y + z)||2 = ||x + y||2 + ||x + z||2 ||x z||2 ||x y||2
and so
1
(x, y + z) = (||x + y||2 + ||x + z||2 ||x z||2 ||x y||2 )
4
1
1
= (||x + y||2 ||x y||2 ) + (||x + z||2 ||x z||2 )
4
4
= (x, y) + (x, z)
Finally, we must check that (x, y) = (x, y) for all R. First suppose that N. We
shall prove the claim by induction on . If = 1 then there is nothing to prove so assume
that the claim holds for arbitrary natural . We have
(x, ( + 1)y) = (x, y + y) = (x, y) + (x, y) = (x, y) + (x, y) = ( + 1)(x, y)
and so the claim holds for all N. Observe that
1
1
(x, y) = (||x y||2 ||x + y||2 ) = (||x + y||2 ||x y||2 ) = (x, y)
4
4

CHAPTER 3. HILBERT SPACES

27

It then follows that the claim holds true for all integers. Now suppose that Q. We may
assume, without loss of generality, that = 1/b where b is a stricly positive integer. Then
2 
2 !





1
1 
1
1 
x y 
x, y =
y

x
+

b
4 
b 
b 
11
(||bx + y||2 ||bx y||2 )
b2 4
1
1
1
= 2 (bx, y) = 2 (y, bx) = (x, y)
b
b
b
=

Now, given any x V , (x, ) is clearly a continuous function in the second argument. Indeed,
this function is expressed in terms of basic arithmetic operations of || || which is continuous.
Since Q is dense in R, the claim must hold for all R and we are done.
Corollary 3.12. Let p [1, ]. Then `p is an inner product space if and only if p = 2.
Proof. By Theorem 3.11, `p is an inner product space if and only if the Parallelogram Law
holds in `p . Let x = e1 and y = e2 where the ei are the standard basis vectors for `p . We
have
||x + y||2p + ||x y||2p = 2||x||2p + 2||y||2p 22/p + 22/p = 4 22/p = 2 p = 2
If p is infinite then the condition reduces to 2 = 4 which is clearly absurd. Hence the
Parallelogram Law does not hold if p 6= 2.
It remains to show that `2 is indeed an inner product space. Define
(, ) : `2 `2 R

X
xi y i
(x, y) 7
i=1

Clearly, (x, x) 0 for all x `2 with equality holding if and only if x = 0. Next, conjugate
symmetry indeed holds:
(x, y) =

xi yi =

i=1

xi yi = (y, x)

i=1

Finally, let x, y, z `2 and C. Then

(x, y + z) =

xi (yi + zi ) =

i=1

xi y i +

i=1

xi zi

i=1

and so (, ) is an inner product on `2 .

Corollary 3.13. Let p [1, ]. Then Lp (R) is an inner product space if and only if p = 2.
Proof. First suppose that p is finite. Let X, Y R be disjoint sets of finite measure. Define

f (x) =

1
(X)

1/p


X (x),

g(x) =

1
(Y )

1/p
Y (x)

CHAPTER 3. HILBERT SPACES

28

We have that
!1/p
1/p

1/p p
Z 
1
1


||f + g||p =
X +
Y  d

 (X)

(Y )
1/p
Z
1
1
X +
Y d
=
(X)
(Y )
= 21/p
where we have used the fact that X and Y to see that X Y = 0. A similar argument also
shows that ||f g||p = 21/p . Note that ||f ||p = 1 and ||g||p = 1. Now, in order for Lp (R) to
be an inner product space, the Parallelogram Law must hold. We have that
||f + g||2p + ||f g||2p = 2||f ||2p + 2||g||2p 21/p + 21/p = 4 21/p = 2

p = 2

Now suppose that p = . Let f = X and g = Y where X and Y are disjoint and have
non-zero measure. Then
||f || = sup { t | ({ x | |f (x)| > t }) > 0 } = 1
Similarly, we see that ||g|| = ||f + g|| = ||f g|| = 1. Now, checking the Parallelogram
Law, we have
||f + g||2 + ||f g|| = 2||f ||2 + 2||g||2 2 = 4
which is absurd. Hence Lp (R) is not an inner product space except possibly at p = 2.
To show that L2 (R) is an inner product space, define
(, ) : L2 (R) L2 (R)
Z
(f, g) 7 f g d
We must first check that this function is well-defined. We have
Z
Z
Z
1
1
f g d |f g| d
|f (x)|2 + |g(x)|2 d <
2
2
Clearly, (f, f ) 0 with equality if and only if f = 0. Next, conjugate symmetry indeed
holds:
Z
Z
(f, g) = f g d = f g = (g, f )
Finally, let f, g, h L2 (R) and C. Then
Z
Z
Z
(f, g + h) = f (g + h) d = f g d + f h d = (f, g) + (f, h)

Definition 3.14. Let V be an inner product space. We say that V is a Hilbert space if
it is complete.
Example 3.15. `2 is a Hilbert space.

CHAPTER 3. HILBERT SPACES

29

Example 3.16. L2 (R) is a Hilbert space.

Definition 3.17. Let V1 and V2 be inner product spaces with inner products given by (, )1
and (, )2 respectively. If U : V1 V2 is a linear operator such that
(U x, U y)2 = (x, y)1
for all x, y V1 then we say that U is unitary
Definition 3.18. Let H1 and H2 be Hilbert spaces. We say that H1 and H2 are isomorphic
if there exists a unitary linear operator from H1 to H2 .
Definition 3.19. Let H1 and H2 be Hilbert spaces with inner products given by (, )1 and
(, )2 respectively. We define their direct sum, denoted H1 H2 , to be the Hilbert space
consisting of the collection of pairs hx, yi together with the inner product
(hx1 , y1 i , hx2 , y2 i) = (x1 , x2 )1 + (y1 , y2 )2
Definition 3.20. Let { Hi }iN be a countable collection of Hilbert spaces where Hi is
equipped with the inner product (, )1 . We define their countable direct sum, denoted

Hi

i=1

to be the Hilbert space consisting of infinite sequences { xi }iN with each xi Hi satisfying

||xi ||21 <

i=1

equipped with the natural inner product.

3.2

The Riesz Lemma

Proposition 3.21. Let H be a Hilbert space and M H a closed (with respect to the
natural topology induced by the norm) subspace. Then M is a Hilbert space whose inner
product is inherited from H.
Proof. Let (, ) be the inner product on H. M is clearly an inner product space with
(, ) restricted to the subspace. Now, any closed subspace of a complete normed space is
necessarily complete whence M is a Hilbert space.
Definition 3.22. Let H be a Hilbert space with inner product given by (, ) and M H
a subspace. We define the orthogonal complement, denoted M to be the following set
M = { v H | (v, m) = 0 m M }
Proposition 3.23. Let H be a Hilbert space with inner product given by (, ) and M H
a subspace. Then M is a closed subspace of H.

CHAPTER 3. HILBERT SPACES

30

Proof. It follows directly from the linearity property of the inner product that M is a
subspace of H.
We now show that M is closed. Indeed, consider a sequence { xn } M such that
xn x as n for some x H. Fix some m M. Then for all n N we have
(xn , m) = 0. By the continuity of the inner product, we have
lim (xn , m) = 0
n

lim xn , m = 0
n

(x, m) = 0
whence x M . The orthogonal complement thus contains all its limit points and is
therefore closed.
Lemma 3.24. Let H be a Hilbert space and M H a closed subspace. For any x H,
there exists a unique m M which is closest to x.
Proof. Fix x H and set d = inf yM ||x y||. Choose a sequence { yn } M such that
||x yn || d. By the Parallelogram Law we have
||yn ym ||2 = ||(yn x) (ym x)||2
= 2||yn x||2 + 2||ym x||2 || 2x + yn + ym ||2
2



1
2
2
= 2||yn x|| + 2||ym x|| 4 x (yn + ym )
2
Now note that 21 (yn ym ) M and thus
2



1


||yn ym || = 2||yn x|| + 2||ym x|| 4 x (yn + ym )
2
2
2
2
2||yn x|| + 2||ym x|| 4d
2

Passing to the limit n, m on both sides yields limn,m ||yn ym ||2 2d2 +2d2 4d2 =
0. Therefore, { yn } is a Cauchy sequence of elements in M. But M is closed whence
{ yn } z for some z M. Hence ||x z|| = d.
It remains to show that such a z is unique. Let z1 , z2 M be such that ||x z1 || =
||x z2 || = d. Then, by the Parallelogram Law, we have
||z1 z2 || = ||(z1 x) (z2 x)||2
= 2||z1 x||2 + 2||z2 x||2 || 2x + z1 + z2 ||
2



1
2
2
= 2||z1 x|| + 2||z2 x|| 4 x (z1 + z2 )
2

2


1
= 4d2 4 x (z1 + z2 )
2
Note that 12 (z1 + z2 ) M and thus

2


1
||z1 z2 || = 4d 4 x (z1 + z2 )
2
2
2
4d 4d
=0
2

Now by the properties of the norm, we can see that z1 = z2 .

CHAPTER 3. HILBERT SPACES

31

Theorem 3.25 (Projection Theorem). Let H be a Hilbert space with inner product given
by (, ) and M H a closed subspace. Then for all x H there exists unique z M and
w M such that x = z + w.
Proof. Fix x H. By Lemma 3.24, there exist a z M closest to x. Define w = x z. We
claim that w M . First, set d = ||x z||. For all y M and t R we have
d2 ||x (z + ty)||2
= ||w ty||2
= ||w||2 + t2 ||y||2 2tRe(w, y)
= d2 + t2 ||y||2 2tRe(w, y)
which implies that
0 t2 ||y||2 2tRe(w, y)
for all t R. Hence Re(w, y) = 0. A similar argument using ti instead of t shows that
Im(w, y) = 0. Hence w M . It remains to show that such a w M is unique. Suppose
that x = z + w1 and x = z + w2 for some w2 M . Then clearly, w1 = w2 .
Remark. The projection theorem implies that that there is a natural isomorphism between
M M and H given by
hz, wi 7 z + w
We will often surpress the isomorphosm and just write H = M M
Definition 3.26. Let V be a normed vector space over C with norm given by || ||X . We
say that f : V C is a linear continuous functional on V if
1. f (x + y) = f (x) + f (y) for all x, y V and C
2. f is bounded. In other words, there exists a C R such that |f (x)| C||x||X for all
x
The vector space of all linear continuous functionals on V is called the dual space of V
and is denoted V . We can endow X with the norm
||f ||X = sup |f (x)|
||x||X =1

Theorem 3.27 (Riesz Representation Theorem). Let H be a Hilbert space with inner product given by (, ). Let T H . Then there exists a unique yT H such that T (x) = (yT , x)
for all x H. Furthermore, ||yT ||H = ||T ||H
Proof. Let N be the subset of H consisting of elements x such that T (x) = 0. By linearity
and continuity, N is a closed subspace. Suppose first that N = H, Then T (x) = 0 = (0, x)
for all x H.
Now suppose that N 6= H. By the Projection Theorem, H = N N and there exists a
non-zero vector x0 N . We define yT = T (x0 )||x0 ||2 x0 . We claim that yT is the desired
element of H. If x N then T (x) = 0 = (yT , x). Now set x = x0 for some scalar . Then
T (x) = T (x0 ) = T (x0 ) = T (x0 )

||x0 ||2
= (T (x0 )||x0 ||2 x0 , x0 ) = (yT , x0 ) = (yT , x)
||x0 ||2

CHAPTER 3. HILBERT SPACES

32

Now, T () and (yT , ) are both linear and they agree on N and at x0 . Hence they must also
agree on the space spanned by N { x0 }. We claim that H = span { N { x0 } }. Indeed,
we can always write y H in the form


T (y)
T (y)
y= y
x0 +
x0
T (x0 )
T (x0 )
which is clearly in span { N { x0 } }. Hence T (x) = (yT , x) for all x H.
It remains to show that ||T ||H = ||yT ||H . We have that
||T ||H = sup |T (x)| = sup |(yT , x)| sup ||yT || ||x||H = ||yT ||H
||x||H =1

||x||H =1

||x||H =1

where we have used the Cauchy-Schwarz inequality. Conversely,

 
 



y
y
T
T
 = yT ,
= ||yT ||H
||T ||H = sup |T (x)| T
||yT || 
||yT ||
||x||H =1
whence ||T ||H = ||yT ||H .

3.3

Orthonormal Bases

Definition 3.28. Let H be a Hilbert space with inner product given by (, ) and { en } H
a subset. We say that { en } is a complete system if (x, en ) = 0 for all x H and for
allen { en } implies that x = 0.
Remark. Let u1 , u2 , . . . be linearly independent vectors. We can construct a set of orthonormal vectors from these vectors using the following:
n1
X
w n = un
(vk , un )vk ,

vn =

k=1

wn
||wn ||

The family of vectors vn is then an orthonormal set. For example,

w 1 = u1
w2 = u2 (v1 , u2 )v1
..
.
This is referred to as the Gram-Schmidt process.
Theorem 3.29. Let H be a separable Hilbert space. Then H contains a complete countable
orthonormal system.
Proof. Suppose H is separable and fix a countable dense subset { xn }. By removing certain
elements from { xn }, we can obtain a linearly independent set whose span is equal to the span
of { xn }. Applying the Gram-Schmidt process to the subcollection, we obtain a complete
countable orthonormal system.
Theorem 3.30. Let H be a separable Hilbert space and S = { xn }nN a complete countable
orthonormal system. Then for all y H we have
X
y=
(xn , y)xn
(3.1)
nN

||y||2 =

X
nN

|(xn , y)|2

(3.2)

CHAPTER 3. HILBERT SPACES

33

Conversely, given a countable subset { cn } C such that

X
|cn |2 <
nN

Then
X

cn x n

nN

converges to an element of H.
Proof. By Bessels Inequality, we have for any finite set N 0 N
X
||y||2
|(xn , y)|2
nN 0

Hence for at most countably many n N we have (xn , y) 6= 0. Order these N1 , N2 , . . . .

Now,
M
X

|(xNi , y)|2

i=1

is a bounded, monotone increasing sequence and hence converges as M . Let { yn } be

the sequence given by
yn =

n
X

(xNj , y)xNj

j=1

Then for any n > m we have


2
n
n
 X

X



2
|(xNj , y)|2
||yn ym || = 
(xNj , y)xNj  =


j=m+1

j=m+1

Therefore { yn } is a Cauchy sequence by the previous result and must converge to some
limit y 0 in H. Now, given any Nl we have
!
n
X
0
(y y , xNl ) = lim y
(xNj , y)xNj , xNl
n

j=1

"
= lim (y, xNl )
n

n
X

!#
(xNj , y)xNj , xNl

j=1

"
= lim (y, xNl )
n

n
X

#
(xNj , y)(xNj , xNl )

j=1

= lim [(y, xNl ) (xNl , y)]

n

=0
Moreover, if we have N 6= Nl for any l then
(y y 0 , xN ) = 0

CHAPTER 3. HILBERT SPACES

34

Hence for all xN { xn } we have (y y 0 , xN ). Since { xn } is a complete orthonormal system,

it follows that y y 0 = 0 and, in particular, y = y 0 . Hence

X
y=
(xNj , y)xNj
j=1

Furthermore,

2
n


X


0 = lim y
(xNj , y)xNj 
n 

j=1

= lim

n
X

n
X
(xNj , y)xNj , y
(xNj , y)xNj

j=1

"
y

= lim

j=1

n
X

!
(xNj , y)xNj , y

j=1

= lim (y, y)

y,

N
X

n
X

(xNj , y)xNj

j=1
2

= lim ||y||
n

n
X

= lim ||y||2 2
n

(xNj , y)(y, xNj )

n
X

= lim ||y||2
n

n
X

!#
(xNj , y)xNj

j=1

!
(xNj , y)xNj , y)

n
X

(xNj , y)(xNj , y) +

(xNj , y)(xNj , y) +

n
X

n
X

(xNj , y)xNj ,

j=1
n
X

j=1

j=1

"

(xNj , y)xNj ,

n
X

j=1

j=1

"

n
X
j=1

"

"

i=1

(xNi , y)

n
X

!#
(xNj , y)xNj

j=1
n
X

(xNj , y)(xNi , xNj )

j=1

#
(xNi , y)(xNi , y)

i=1

#
(xNj , y)(xNj , y)

j=1

"
= lim ||y||2
n

n
X

#
|(xNj , y)|2

j=1

and so ||y||2 = nN |(xn , y)|2 .

P
Conversely, suppose that { cn } C is such that nN |cn |2 < . We need to show that
X
cn x n
P

nN

converges to an element of H. Since H is a Hilbert space, it is complete so we just need to

show that the sequence given by
n
X
yn =
ci x i
i=1

is Cauchy. Note that

M
X

|ci |2

i=1

is a bounded, monotone increasing sequence which converges to some limit as M . We

have that
n
n
X
X
2
2
||yn ym || = ||
ci xi || =
|ci |2
i=m+1

and thus { yn } is a Cauchy sequence.

i=m+1

Chapter 4
Banach Spaces
4.1

Dual Spaces

Proposition 4.1. Let X be a normed vector space over R (or C) with norm given by || ||.
Let f X be a continuous linear functional. Then the following are equivalent definitions
for ||f ||X :
1. I = inf { c | |f (x)| c||x|| x X }
2. S1 = sup||x||1 |f (x)|
3. S2 = sup||x||=1 |f (x)|
4. S3 = supx6=0

|f (x)|
||x||

Proof. We clearly have S2 S1 . Furthermore, observe that

 

|f (x)| 
x 
= f
||x||
||x|| 
and so S3 S2 . Now suppose that ||x|| 1. Then
|f (x)|

|f (x)|
||x||

and so S3 S1 . It then follows that S1 = S2 = S3 . Now note that

|f (x)| S3 ||x||
for all x X and thus I 3. Conversely, by the definition of the supremum we have
I

|f (x)|
S3 x
||x||

for some x 0. In particular, we can find a sequence xn such that xn =

the limit n yields I S3 and we are done.

1
n

and passing to

Theorem 4.2. Let X be a normed vector space over C with norm given by || ||X . Then
X is a Banach space.

35

CHAPTER 4. BANACH SPACES

36

Proof. We must show that X is complete with respect to its norm. Let { fn } X be a
Cauchy sequence. By definition we have that for all > 0 there exists an N such that for
all m, n > N ,
||fn fm ||X <
Fix x X. From the definition of the norm we have
|fn (x) fm (x)| = |(fn fm )(x)|
||fn fm ||X ||x||X = ||x||X

(4.1)
(4.2)

and hence { fn (x) } is a Cauchy sequence of complex numbers. Since C is complete, { fn (x) }
converges to some point f (x). We claim that f (x) is the limit point of the original Cauchy
sequence. We must first show that the function is linear. Let x1 , x2 X and , C.
Then
f (x1 + x2 ) = lim [fn (x1 + x2 )]
n

= lim [fn (x1 ) + fn (x2 )]

n

= f (x1 ) + f (x2 )
as required. Now passing to the limit m in Equation 4.2 we have
|fn (x) f (x)| ||x||X
for all n N . Now since x is arbitrary, f (x) must be bounded. Hence
||fn f ||X
for all n N . Hence the original Cauchy sequence { fn } converges to the function f whence
X is complete.
Remark. If H is a Hilbert space, given any element g H the map f 7 (f, g) is a
continuous functional. By the Riesz representation theorem, all functionals on H have this
form and thus H = H .
Definition 4.3. Let X and Y be vector spaces over R (or C) with norms given by ||||X and
|| ||Y respectively. We say that a linear map f : X Y is an isometry if ||f (x)||Y = ||x||X
for all x X. We say that X and Y are isometric if there exists a bijective isometry
between them.
Lemma 4.4. Let X and Y be vector spaces over R (or C) with norms given by || ||X and
|| ||Y respectively. If f : X Y is an isometry then f is injective.
Proof. Since f is a linear map, it suffices to show that if T (x) = 0 for some x X then
x = 0. We have that
||x||X = ||T (x)||Y = ||0|| = 0
By the properties of the norm, it follows that x = 0 as required.
Proposition 4.5. (`1 ) is isometric to ` .

CHAPTER 4. BANACH SPACES

37

Proof. We claim that the following mapping is an isometry

J : ` (`1 )

y 7 y (x) =

yn xn

n=1

where x `1 . We must first show that this map is well defined. y is clearly linear. Indeed,
given w, v `1 and , C we have
y (w + v) =

yn (wn
n=1

+ vn )

yn wn +

n=1

yn vn

n=1

= y (w) + y (v)
y is also bounded in `1 :
|y (x)| = |

yn xn |

n=1

|yn ||xn |

||y|| |xn | = ||y|| ||x||1

n=1

n=1

Now by the definition of the norm on (`1 ) we have ||y ||(`1 ) ||y|| . To show the opposite
inequality, we fix some y ` . Fix > 0 and n N be such that |yn | ||y|| . Then
|y (en )| = |yn | ||y|| (||y|| )||en ||1
This implies that
||y ||(`1 ) =

|y (x)|
|y (en )|

||y||
||en ||`1
xX\{ 0 } ||x||`1
sup

But is arbitrary whence ||y ||(`1 ) = ||y|| .

Since J is an isometry, it must be injective so it remains to show that J is surjective.
Fix some (`1 ) . Setting yn = (en ) we can see that y = (y1 , y2 , . . . ) ` . Consider
= y (`1 ) . Clearly, (en ) = 0 for all n. Hence vanishes on the set of all linear
combinations of en . But such a set is dense in `1 and, since is continuous, we must have
that 0 whence = y .
Proposition 4.6. Let p, q (1, ) be such that

1
p

1
q

= 1. Then (`p ) is isometric to `q .

Proof. We claim that the following map is an isometry

J : `q (`p )
y 7 y (x) =

y n xn

n=1

where x `p . We first claim that y is linear. To this end, let w, v `p and , C. Then
y (w + v) =

X
n=1

yn (wn + vn ) =

X
n=1

yn wn +

X
n=1

yn vn = y (w) + y (v)

CHAPTER 4. BANACH SPACES

38

To see that it is bounded, we apply Holders inequality:



! 1q

 X
X
X


y n xn 
|yn xn |
|y (x)| = 
|yk |q
 n=1
 n=1

k=1

! p1
|xk |p

k=1

= ||y||q ||x||p
It thus follows that
||y ||(`p ) ||y||q
Now fix non-zero y `p and define

xn =

yn |yn |q2 if yn =
6 0
0
if yn = 0

Then
||x||pp =

|xn |p =

n=1

||yn |yn |q2 |p =

n=1

|yn |p |yn |pq2p

n=1

=
=

X
n=1

|yn |p |yn |qp

|yn |q

n=1

= ||y||qq
Furthermore,
y (x) =

X
n=1

y n xn =

yn yn |yn |q2 = |yn |2 |yn |q2 = |yn |q = ||y||qq

n=1

Now applying this to the definition of the norm, we see that

||y (x)||(`p )

||y||qq
q pq
|y (x)|
=
=
||y||
= ||y||q
q
q
p
||x||p
||y||q

whence ||y (x)||(`p ) = ||y||q . Hence J is an isometry and must be injective. It remains to
show that J is surjective.
Fix some (`p ) . Set yn = (en ) and let xn be as previously defined. We claim
that J(y) = . We must first show that y `q . Choose some N N and let x =
(x1 , x2 , . . . , xN , 0, . . . ). Then
! pq
N
N
X
X
|yn |q = (x) ||||(`p ) ||x||p = ||||(`p )
|yn |q
n=1

n=1

PN

whence n=1 |yn |q ||||q(`p ) . Now since N is arbitrary, we see that y `q . Consider
= y . We have that (en ) = 0 for all n. thus vanishes on the set of all finite linear
combinations of en . But this set is dense in `1 and is continuous so we must have that
0 whence = y .
Example 4.7. Similarly, (Lp [a, b]) is isometric to Lq [a, b] when 1 < p < and

1
p

1
q

= 1.

CHAPTER 4. BANACH SPACES

4.2

39

Hahn-Banach Theorem

Theorem 4.8. Let X be a Banach space and Y X a linear subspace. Let : Y C be

a linear functional such that |(x)| a||x||X for all x Y . Then there exists some X
such that ||||X a and (x) = (x) for all x Y . In other words, any bounded linear
functional on a subspace of a Banach space can be extended onto the whole Banach space
without increasing the norm.
Proof. For simplicity, we only consider the case where a = 1 and X is separable. We first
reduce to the real case. Set `(x) = Re(x). Then `(x) is a bounded R-linear functional.
Clearly, (x) = `(x) i`(ix). Hence, if we can extend ` to L without increasing the norm
then (x) = L(x) iL(ix) extends (x). Furthermore, would be bounded with norm at
most 1. Indeed, given x X, set = (x)/|(x)| then
|(x)| = (x) = (x) = L(x) ||ax|| = ||x||
Were we have used the fact that |(x)| is real to discard the imaginary part when going
from (ax) to L(ax). We now extend ` by one real dimension. Fix some z X\Y . We
shall extend ` onto the R-linear set
Z = { y + tz | y Y, t R }
without increasing the norm. Since ` is R-linear, we have that `(y + tz) = `(y) + t`(z).
Hence it suffices to define the value of `(z). For convenience, we set p = `(z). We need to
find a p such that
|`(y) + tp| ||y + tz||

(4.3)

for all y Y and t R. To this end, let y1 , y2 Y and , R positive. Then

|`(y1 y2 )| ||y1 y2 || ||y1 z|| + ||z y2 ||
= ||y1 z|| + ||z y2 ||
From which we get
1
1
(`(y1 ) ||y1 z||) (`(y2 ) + ||y2 z||)

Now, y1 , y2 , and are arbitrary so

sup
>0,y1 Y

1
1
(`(y1 ) ||y1 z||) inf
(`(y2 ) + ||y2 z||)
>0,y2 Y

Hence we may replace y1 and y2 with y and and with t to get

1
1
(`(y) ||y tz||) p (`(y) + ||y tz||)
t
t

(4.4)

for some p R. Replacing y by y also yields

1
1
(`(y) + ||y + tz||) p (`(y) ||y + tz||)
t
t

(4.5)

Multiplying Equations (4.4) and (4.5) through by t we have

`(y) ||y tz|| tp `(y) + ||y tz||
`(y) ||y + tz|| tp `(y) + ||y + tz||

(4.6)
(4.7)

CHAPTER 4. BANACH SPACES

40

Multiplying Equation (4.6) through by 1 we get

`(y) ||y tz|| tp `(y) + ||y tz||

(4.8)

Now, Equations (4.7) and (4.8) imply that the same inequality holds for all positive and
negative real numbers. Since the case where t = 0 is trivial, the inequality holds for all
t R. It is clear that this inequality is equivalent to Equation (4.3) and hence we have
extended ` by one real dimension.
With this in mind, we can easily extend ` onto the set
{ y + itz | y Y, t R }
without increasing the norm. This is done in exactly the same way as the extension by the
one real dimension by replacing z with iz.
To finish off the proof, we fix some countable dense subset x1 , x2 , . . . in X. We can
define a sequence of linear setsSY Y1 Y2 X such that xn Yn .SBy the previous
results, we can extend ` onto n Yn without increasing the norm. Clearly, n Yn is dense in
X. Now since ` is continuous, we may extend ` onto X and we are done.
Corollary 4.9. Let X be a Banach space and let x X. Then there exists an ` X such
that ||`||X = 1 and `(x) = ||x||X .
Proof. Let Y be the one dimensional subspace of X spanned by x and define `(tx) = t||x||X
for all t C. We can then extend ` to all of X by the Hahn-Banach theorem.

4.3

Second Dual space

Proposition 4.10. Let X be a Banach space. Denote X as the dual of X . Then there
exists an isometric embedding J : X X .
Proof. Fix x X. Then x generates a linear functional x X given by x (`) = `(x) for
some ` X . We have that |x (`)| = |`(x)| ||l||X ||x||X whence ||x ||X ||x||.
Now we may choose ` by Corollary 4.9 such that `(x) = ||x||X and ||`||X = 1. It follows
that |x (`)| = |`(x)| = ||x||X . Hence
||x ||X =

sup |x (`)| =
||`||X =1

sup |`(x)| ||x||X

||`||X =1

If we write J(x) = x then ||J(x)||X = ||x ||X = ||x||X whence J is an isometry.

Definition 4.11. Let X be a Banach space. We say that X is reflexive if X is isometric
to its second dual space X . In other words, X is reflexive if the above map J is surjective.
Example 4.12. For all 1 < p < , `p and Lp are reflexive. However they are not reflexive
for p = 1 and . Consider the subspace c0 of ` consisting of all sequences whose limit is
0. Then it can be shown that c0 = `1 but (`1 ) = ` as we have seen.

4.4

Bounded linear operators

Definition 4.13. Let X and Y be normed vector spaces over R (or C) with norms given
by || ||X and || ||Y respectively. We say that f : X Y is a bounded linear operator if

CHAPTER 4. BANACH SPACES

41

1. f (x + y) = f (x) + f (y) for all x1 , x2 V and C

2. f is bounded. In other words, there exists a C R such that ||f (x)||Y C||x||X for
all x X
We denote the vector space of all bounded linear operators from X to Y by B(X, Y ). We
can make B(X, Y ) into a normed vector space with the operator norm
||f || = sup
x6=0

||f (x)||Y
||x||X

Proposition 4.14. Let X and Y be normed vector spaces over R (or C) with norms given
by || ||X and || ||Y respectively. Let f B(X, Y ) be a bounded linear operator. Then the
following are equivalent definitions for ||f ||:
1. I = inf { c | ||f (x)||Y c||x||X x X }
2. S1 = sup||x||X 1 ||f (x)||Y
3. S2 = sup||x||X =1 ||f (x)||Y
4. S3 = supx6=0

||f (x)||Y
||x||X

Proof. This proof follows the same argument as the proof for Proposition 4.1.
Theorem 4.15. Let X be a normed vector space and Y a Banach space over R (or C).
Then B(X, Y ) is a Banach space.
Proof. This proof follows the same argument as the proof for Theorem 4.2.
Theorem 4.16 (Banach-Steinhaus theorem). Let X be a Banach space and F a collection
of bounded linear operators from X into a normed space Y such that the set { T x | T F }
is bounded for all x X. Then the set of norms {||T || | T F} is bounded.

4.5

Weak convergence

Definition 4.17. Let X be a Banach space and { xn } X a sequence. We say that { xn }

weakly converges to x X, denoted w-lim xn = x, if `(xn ) `(x) for all ` X .
Remark. Let { xn } X be a sequence in a Banach space. If ||xn x|| 0 as n
then clearly, xn converges weakly to x. Indeed, let f X be a continuous linear functional.
Since f is continuous, we have
lim f (xn ) = f ( lim xn ) = f (x)

and so w-limn xn = x.
The converse does not hold. Indeed, consider the sequence of basis vectors { en } `2 .
Since `2 is an inner product space, any linear continuous functional takes the form (y, ) for
some y `2 . Recall that y can be expressed as an infinite sum
y=

X
i=1

(y, ei )ei

CHAPTER 4. BANACH SPACES

42

and so
lim (y, en ) = lim

!
(y, ei )ei , en

= lim

i=1

(y, ei )(ei , en ) = lim

i=1

(y, ei )in = lim (y, en )

n

i=1

Now let ci = (y, ei ). y `2 means that

2

||y|| = (y, y) =

X
i=1

(y, ei )ei ,

X
i=1

!
(y, ei )ei

X
i=1

(y, ei )

X
j=1

(y, ei )(ei , ej ) =

|(y, ei )|2 <

i=1

and so limi (y, ei ) = 0. Hence w-limn (en ) = 0.

On the other hand, ||en ||2 = 1 for all n and so { en } does not converge in `2 .
Theorem 4.18. Let X be a Banach space and { xn } X a sequence. If { xn } converges
weakly then { xn } is bounded.
Proof. Recall that X embeds isometrically into X so we may consider each xn and x as
elements of X . Fix an ` X . Then `(xn ) = xn (`) converges and is thus bounded. Now
the Banach-Steinhaus theorem implies that the norms ||xn ||X = ||xn ||X are bounded.

Chapter 5
Compactness in Banach Spaces
5.1

Preliminaries

Definition 5.1. Let (X, ) be a metric space and K X a subset. A collection S of subsets
of X is said to be a cover of K if
[
K
S
SS

If every member of S is open then S is said to be an open cover. If a subset S0 S is also

a cover of K then S0 is said to be a subcover of K.
Definition 5.2. Let (X, ) be a metric space and K X a subset. We say that K is
compact if every open cover of K has a finite subcover.
Definition 5.3. Let (X, ) be a metric space and K X a subset. We say that X is
sequentially compact if any sequence of elements in K has a subsequence that converges
to a limit in K.
Theorem 5.4. Let (X, ) be a metric space and K X a subset. Then K is compact if
and only if K is sequentially compact.
Theorem 5.5. Let (X, ) be a metric space and K X a subset. If K is compact then K
is closed and bounded.
Theorem 5.6. Let K Rn (or Cn ) be a subset. Then K is compact if and only if K is
closed and bounded.
Proposition 5.7. Let (X, ) be a metric space and K X a compact set. If K0 K is
closed then K0 is compact.
Theorem 5.8. Let (X, ) and (Y, ) be metric spaces and f : X Y a continuous mapping.
If K X is compact then f (X) is compact in Y .
Corollary 5.9. Let (X, ) be a metric space and f : X R a continuous mapping. If
K X is compact then f attains its maximum and minimum on K.
Proof. The image of K in R is compact and is thus closed and bounded. Hence max f (K)
and min f (X) exist and are finite.

43

CHAPTER 5. COMPACTNESS IN BANACH SPACES

44

Definition 5.10. Let (X, ) be a metric space and f : X R (or C) a function. We say
that f is uniformly continuous if for all > 0, there exists a > 0 such that |x y| <
implies |f (x) f (y)| < .
Theorem 5.11. Let (X, ) be a metric space and K X a compact subset. If f : K R
is continuous then it is uniformly continuous.
If a subset of a metric space is compact then it is closed and bounded. We shall soon see
that the converse holds in a Banach space if and only if the metric space is finite dimensional.
The following are counter examples for the infinite dimensional cases.
Example 5.12. We claim that for all p [1, ], the closed unit ball B centered at zero in
`p is not compact. Consider the sequence { en } consisting of the standard basis elements.
Clearly, en B for all n. Now if p is finite, we have
||en em ||p =

!1/p
(i) p
|e(i)
n em |

= 21/p

i=1

If p is infinite then
(i)
||en em || = sup |e(i)
n em | = 1
iN

Hence the sequence is not Cauchy and cannot admit a convergent subsequence.
Example 5.13. We claim that the unit ball B centered at zero is not compact in C[a, b].
Let In [a, b] be a sequence of disjoint open intervals. For each n, let fn be a function that
is zero on [a, b]\In satisfying 0 f (x) 1 for all x In and f (x) = 1 for at least one x In
(for example, let f (x) be a piecewise linear function in the shape of a triangle). We have
that
||fn || = sup |fn (x)| = 1
x[a,b]

and so each fn B. Furthermore,

||fn fm || = sup |fn (x) fm (x)|| = 1
x[a,b]

Hence the sequence { fn } is not Cauchy and cannot admit a convergent subsequence.

5.2

Finite Dimensional Subspaces

Lemma 5.14. Let V be an n-dimensional vector space over Rn (or Cn ). Then all norms
on V are equivalent.
Proof. We shall prove the lemma for Cn . Let || || be a norm on Cn . We claim that
Pn|| || is
n
equivalent to the norm || ||1 . To this end, fix an x C . We can always write x = i=1 xi ei
for some xi C where the ei are the standard basis vectors. Then


! n
!
n
n
n
X
 X
X
X


||x|| = 
xi e i  
|xi | ||ei ||
|xi |
||ei || = C||x||1


i=1

i=1

i=1

i=1

CHAPTER 5. COMPACTNESS IN BANACH SPACES

45

P
where C =
i=1 ||ei ||.
Conversely, let K be the set
K = { x Cn | ||x||1 = 1 }
Then K is clearly closed and bounded and is thus compact. Since || || is a norm, it is
a continuous function and so || || attains its minimum, say m, on the compact set K.
Furthermore, 0 6 K so, necessarily, m > 0. Hence for all x K we have
||x|| m
Now let x Cn . Clearly, x/||x||1 K and so


 x 


 ||x||1  m
and so ||x|| m||x||1 for all x Cn . Hence || || is equivalent to || ||1 .
Proposition 5.15. Let (V, ||||) be an n-dimensional Banach space over C. Then any closed
ball in V is compact.
Proof. We may assume that V is isomorphic to Cn . Let B be a closed ball in V with respect
to || ||. By Lemma 5.14, all norms on V are equivalent so || || induces the same topology
on V as || ||2 . Since B is closed and bounded in the || ||2 topology, it is compact. Hence
B is compact with respect to || ||.
Proposition 5.16. Let (V, || ||) be a Banach space over C. Then any finite dimensional
subspace of V is closed.
Proof. We may assume that V is isomorphic to Cn . Then any linear subspace of V will be
isomorphic to Cm for some 1 m n. Since Cm is closed with respect to || ||2 , it is then
closed with respect to all norms.
Lemma 5.17 (Riesz Lemma). Let (V, || ||) be a normed vector space over R (or C) and
V0 ( V a closed linear subspace. Then for all (0, 1), there exists x0 V \V0 such that
||x0 || = 1 and ||x0 x|| 1 for all x V0 .
Proof. Fix some x1 V \V0 . The following number is greater than 0:
d = inf ||x1 x||
xV0

Indeed, if d were to equal zero then there would exist a sequence yn V0 such that ||x1
yn || 0. Since V0 is closed, this would imply that x1 V0 which is a contradiction. Now,
d < d/(1 ) so there exists y V0 such that d ||x1 y|| d(1 ). Define
x0 =

x1 y
||x1 y||

We claim that x0 is the desired element. Clearly, ||x0 || = 1. Furthermore, since V0 is a linear
subspace, y + ||x1 y||x V0 for all x X0 . Hence
||x0 x|| =

x1 (y + ||x1 y||x)
1
1
=
||x1 (y + ||x1 y||x)||
=1
||x1 y||
d
d

CHAPTER 5. COMPACTNESS IN BANACH SPACES

46

Theorem 5.18. Let (V, || ||) be a Banach space. Then a closed ball in V is compact if and
only if V is finite dimensional over its base field.
Proof. Assume that V is infinite dimensional. We shall exhibit the existence of a bounded
sequence in the unit ball about zero that does not have any Cauchy subsequences. To this
end, fix x1 V such that ||x1 || = 1. Denote by V1 the one-dimensional linear subspace
spanned by x1 . By Proposition 5.16, V1 is closed and V1 6= V since V is infinite dimensional.
By Riesz Lemma with = 1/2, there exists x2 V such that ||x2 || = 1 and ||x2 x|| 1/2
for all x V1 . In particular, ||x2 x1 || 1/2.
Proceeding by induction, suppose that we have already constructed x1 , . . . , xn V such
that
||xk || = 1,

||xk xj || 1/2

for all j, k = 1, . . . , n and j 6= k. Denote by Vn the n-dimensional linear subspace spanned

by x1 , . . . , xn . Proposition 5.16 again implies that Vn is closed and Vn 6= V since V is infinite
dimensional. Appealing again to Riesz Lemma with = 1/2, there exists xn+1 V such
that ||xn+1 || = 1 and ||xn+1 x|| 1/2 for all x Vn . We have thus constructed an infinite
sequence xk V such that ||xk || = 1 for all k and ||xk xj || 1/2 for all j, k = 1, . . . , n
and j 6= k. Clearly, this sequence cannot contain any Cauchy subsequences. Furthermore,
the sequence is contained in the unit ball centered at zero. Such a ball is therefore not
sequentially compact whence it is not compact.
The converse is exactly Proposition 5.15.

5.3

Total Boundedness

Definition 5.19. Let (X, ) be a metric space and K X a subset. We say that K is
totally bounded if, given any > 0, K is contained in a union of finitely many balls of
radius .
Example 5.20. Consider the subset of ` given by all the standard basis vectors ei . Then
the distance between any two ei is 1 meaning the set is bounded. However, any open unit
ball around an ei will not contain any of the other ei and so this set cannot possibly be
totally bounded.
Theorem 5.21. Let (X, ) be a complete metric space and K X a subset. Then K is
compact if and only if it is closed and totally bounded.
Proof. First suppose that K is compact. Then K is closed. Now given > 0, consider the
open cover { B (x) | x K } of K. Since K is compact, such a cover necessarily has a finite
subcover. But this is exactly what it means for K to be totally bounded.
Conversely, assume that K is closed and totally bounded. Fix a sequence { yn } in K.
We need to construct a subsequence of { yn } that converges to an element of K. Since K
is totally bounded, there exists a finite cover of K by balls of radius 1. At least one of
these balls contains infinitely many points of { yn }, label it B1 (x1 ). Let K1 = K B1 (x1 ).
K1 is also clearly totally bounded so there exists a finite cover of K by balls of radius 1/2.
At least one of these balls contains infinitely many points of { yn }, label it B1/2 (x2 ). Let
K2 = K1 B1/2 (x2 ). Continuing in this fashion, we construct a sequence of sets Kn satisfying
1. diam Kn 0 as n
2. for all n, Kn contains infinitely many points of the sequence { yn }

CHAPTER 5. COMPACTNESS IN BANACH SPACES

47

The second property allows us to select a subsequence { ynk } such that yni Ki for all
i N. The first condition ensures that { ynk } is Cauchy. Since X is complete, this sequence
must converge to some limit in X. But K is closed and must contain such a limit point.
Therefore K is sequentially compact whence it is compact.
Example 5.22 (Hilberts Brick). Consider the subset K `2 consisting of sequences satisfying
|x1 | 21 , |x2 | 22 , |x3 | 23 , . . .
We claim that K is compact. It suffices to show that K is closed and totally bounded. K
is clearly closed as it is the infinite intersection of the closed sets
Kn = { x `2 | |xn | 2n }
Now fix > 0. We need to find a finite set S `2 such that K
n N such that 2n1 . Consider the mapping

xS

B (x). Choose

x = (x1 , x2 , . . . , xn , . . . ) 7 x = (x1 , x2 , . . . , xn , 0, 0, . . . )
We first observe that ||x x ||2 /2. Indeed, we have
!1/2
!1/2
!1/2

X
X
X
||x x ||2 =
|xi xi |2
=
|xi |2

|2i |2
=
i=1

i=n+1

i=1

X
i=1
n

!1/2
i


=

4n
3

2n1

2
= =
2
3
2 3
The set K of all points x is totally bounded since it is a closed bounded
set in a finite
S

B
dimensional space. Hence there exists a finite
set
S
such
that
K

xS /2 (x). Since
S

||x x || /2, it is then clear that K xS B (x).

5.4

Arzela-Ascoli Theorem

Definition 5.23. Let C[a, b] be a family of functions. We say that is uniformly

bounded if there exists M R such that |(x)| M for all and x [a, b].
Definition 5.24. Let C[a, b] be a family of functions. We say that is equicontinuous if, given > 0, there exists > 0 such that |(x1 ) (x2 )| for all and
x1 , x2 [a, b] such that |x1 x2 | < .
Theorem 5.25 (Arzela-Ascoli). Let C[a, b] be a subset of the continuous real valued
functions on [a, b]. Then is totally bounded if and only if it is uniformly bounded and
equicontinuous.
Proof. First suppose that is totally bounded. Then, clearly, is bounded. In other
words, Br (0) for some r > 0. Hence is uniformly bounded. We must now show
that is equicontinuous. Fix some > 0. S
Since is totally bounded, there exist finitely
many functions f1 , . . . , fn ) such that nk=1 B/3 (fk ). Now, each fk is continuous by
hypothesis and [a, b] is compact. Hence each fk is uniformly continuous and thus, for each
k, there exists a k > 0 such that
|fk (x) fk (y)| /3

whenever |x y| k

Let = mink k . Then for all k we have

|fk (x) fk (y)| /3

whenever |x y|

1/2