M2206

Introduction to Functional Analysis

(Operators and Hilbertian Analysis)

Lecture Notes

A. Kachmar & B. Kojok & H. Abbas

Lebanese University, Faculty of Sciences, Department of Mathematics
Email address: [email protected]
Email address: [email protected]
Email address: [email protected]
 Abstract. This document contains a short introduction to some notions of functional
analysis in normed vector spaces. Topics are selected to fit the requirements of the course
M2206 at Lebanese University. Every chapter contains a collection of basic definitions,
properties, examples and exercises. A set of comprehensive exercises is given right after
each chapter.

c This document can be reproduced in its entirely for non commercial use.
 Contents

Chapter 1. NORMED VECTOR SPACES 5

1. Vector spaces 5
2. Normed spaces 6
3. Bound sets in normed space 9
4. Product of normed spaces 11
5. Equivalent Norms 11
6. Compact sets in normed spaces 14
7. Banach spaces 16
Exercises 20
Chapter 2. BOUNDED LINEAR TRANSFORMATIONS 27
1. On continuous functions between normed spaces 27
2. Linear continuous functions 28
3. The space of bounded linear transformations 31
Exercises 34
Chapter 3. HILBERT SPACES 41
1. Inner product spaces 41
2. Orthogonality 45
3. The Projection theorem 46
4. Applications of the projection theorem 50
5. Hilbert basis 53
6. Separable spaces 55
Exercises 60
Chapter 4. FOURIER SERIES 69
1. Density in the space of continuous functions 69
2. The trigonometric system 71
3. Uniform convergence of Fourier series 73
4. The Dirichelt theorem 75

3
 CHAPTER 1

NORMED VECTOR SPACES

In metric spaces, we studied notions such as convergence and continuity in abstract

sets. The only requirement was having a specific distance function in the set in question.
That is, we do not demand that the set has an algebraic structure.
In this course we will consider spaces having a richer structure compared to abstract
metric spaces. Besides the distance that will allow us to measure the nearness of two
points, the space is supposed to carry an algebraic structure: elements of the space can
be added and multiplied by real or complex numbers to produce new elements. For such
spaces, it is interesting to study distance functions that are compatible with the algebraic
operations. This compatibility will be described by precise terms below.

1. Vector spaces
It is assumed that the reader has already encountered vector spaces before but a brief
summery of the theory is provided here for reference purposes. In this section, X denotes
a vector space over K, (K = R or C). In this context the real or the complex numbers
will be called scalars.
Definition 1.1. A vector space X is a non empty set on which are defined an
operation of vectors + : X × X 7→ X satisfying associativity, commutativity, zero and
inverse axioms, and an operation of scalar multiplication × : K × X → X that satisfies
the following laws, for every x, y, z ∈ K and λ, µ ∈ K,
(1) x + (y + z) = (x + y) + z,
(2) x + y = y + x,
(3) 0X + x = x,
(4) x + (−x) = 0X ,
(5) λ(x + y) = λx + λy,
(6) (λ + µ)x = λx + µx,
(7) (λµ)x = λ(µx),
(8) 1x = x.

Definition 1.2. A nonempty subset F of a vector space X is called a subspace of

X if F is a vector space under the operations addition and scalar multiplication defined
in X.
Theorem 1.3. Suppose X is a vector space and F ⊂ X is a nonempty subset of X.
Then F is a subspace of X if and only if the following two closure conditions hold:
(1) If u, v ∈ F then u + v is in F ;
5
6 1. NORMED VECTOR SPACES

(2) If u ∈ F and λ ∈ K is a scalar, then λu ∈ F .

We end this section by the following theorem.
Theorem 1.4. Every vector space has a basis.
We remind the reader that a basis of X is a set S ⊂ X such that
• Every finite collection of vectors in S are linearly independent ;
• Every x ∈ X can be expressed as a linear combination of a finite number of
vectors of S.

2. Normed spaces
In this section, we assume that K denotes one of the fields R or C, and that X is a
vector space over K.
2.1. Norms.
Definition 2.1. We call a norm on a vector space X any mapping N : X 7−→ R+
(in general we write N (x) = kxk) verifying, for every x, y ∈ X and λ ∈ K, the following
three properties:
(N1 ) N (x) = 0 ⇔ x = 0 .
(N2 ) N (λx) = |λ|N (x).
(N3 ) N (x + y) ≤ N (x) + N (y).
A vector space X endowed with a norm N is called a normed vector space (n.v.s.) and
we denote it by the pair (E, N ). If necessary, norms on different spaces are distinguished
by subscript such that as k · kX . A positive function that satisfies (N 2) and (N 3) is called
semi-norm.

Example 2.2.
(1) [Usual norm in R or C] The absolute value function x 7→ |x| defines a norm
on R and the modulus function z 7→ |z| defines a norm on C. These are called
the usual norms.
(2) The mappings defined, for every x = (x1 , · · ·, xn ) ∈ Rn :
n n
!1/2
X X
kxk1 = |xi | , kxk2 = |xi |2 and kxk∞ = max |xi | .
1≤i≤n
i=1 i=1

are norms on Rn .
(3) Let a, b ∈ R (a < b) and X = C([a, b], R) the vector space of real valued contin-
uous functions over [a, b]. The mappings defined, for every f ∈ X, by
Z b Z b 1/2
2
kf k1 = |f (x)| dx , kf k2 = |f (x)| dx and kf k∞ = max |f (x)| .
a a x∈[a,b]

are norms on X.
Remark 2.3.
(1) We denote by C([a, b], K) = {f : [a, b] → K : f is continuous } with K = R or C.
In the case of K = R, we write C[a, b] instead of C([a, b], K) for simplicity.
(2) [Induced norm] Let F be a vector subspace of a normed space (X, N ). The
restriction of N to F is a norm on F called the induced norm of (X, N ) on F .
 2. NORMED SPACES 7

2.2. Topology induced by the norm. Using the properties of the norm in Defini-
tion 2.1, it is easy to verify that (the proof is left as an exercise)
Proposition 2.4. Normed spaces are metric spaces: Let (X, k · k) be a normed
space. The function d : X × X → [0, ∞) defined by d(x, y) = kx − yk defines a distance
on X. This distance is called the distance induced by the norm k · k.
The previous proposition guaranties that every normed vector space is a metric space.
However, the converse is not necessary true. To clarify this, we make the following
definition.
Definition 2.5. Distance induced by a norm: Suppose that (X, d) is a metric
space and X is a vector space on K. The distance d is said to be induced by a norm if
there exists a norm k · k in X such that, for all (x, y) ∈ X × X, dk·k (x, y) = kx − yk.
It is worth to mention that under some conditions, a distance can provide a norm.
Proposition 2.6. Let E be a vector space endowed with a distance d verifying two
additional properties:
(1) d is homogeneous i.e. ∀x, y ∈ E, ∀λ ∈ K, d(λx, λy) = |λ|d(x, y).
(2) d is invariant by translation i.e. ∀x, y, a ∈ E, d(x + a, y + a) = d(x, y).
Then the mapping defined by
N : E −→ R+
x 7−→ N (x) = d(x, 0)

is a norm over E.
Proof. Left as an exercise. 
Example 2.7.
(1) [Usual norm in R] The distance induced by the the usual norm | · | on R is the
usual distance on R, that is (x, y) 7−→ d(x, y) = |x − y|.
(2) The discrete distance in R is not induced by a norm.
Since every normed space is a metric space, we get a notion of open sets in normed
spaces, as well as many other topological notions.
Definition 2.8. [Topology induced by the norm] Let (X, k · k) be a normed
space and dk·k the distance induced by the norm k · k. We introduce the following class of
subsets of X:
Tk·k = {A ⊂ X : A is open in (X, dk·k )} .
The class Tk·k is called the topology induced by k · k.
Remark 2.9. [Topological notions in normed spaces] Let (X, k · k) be a normed
vector space and dk·k be the distance induced by k · k. Since (X, dk·k ) is a metric space,
then we can talk about various topological notions in X, for example:
(1) [Convergence] A sequence (un ) is convergent in (X, k · k) if ∃ u ∈ X such that
kun − uk → 0 ;
(2) [Cauchy sequences] A sequence (un ) is Cauchy in (X, k · k) if, for all ε > 0,
∃ n0 ∈ N, ∀ p ≥ q ≥ n0 =⇒ kup − uq k < ε.
(3) [Continuity] A function f : (X, k · k) → (R, | · |) is continuous at u0 ∈ X if
∀ ε > 0, ∃ δ > 0, ku − u0 k < δ =⇒ |f (u) − f (u0 )| < ε ;
8 1. NORMED VECTOR SPACES

(4) Interior, adherent, boundary and limit points of subsets of X ;

(5) Closure and interior of subsets of X ; dense subsets, etc.
(6) Completeness, compactness and connectedness of subsets of X.
Next we state a useful theorem on the calculus of convergent sequences. The proof is
left as Exercises to the reader.
Theorem 2.10. Let (X, k · k) be a normed space.
(1) [Sum rule] Let (xn ) ⊂ X, (yn ) ⊂ X, a ∈ X and b ∈ X. If
xn → a and yn → b in (X, k · k) ,
then xn + yn → a + b in (X, k · k).
(2) Let (xn ) ⊂ X and a ∈ X. It holds:
(a) xn → a in (X, k · k) if and only if xn − a → 0 in (X, k · k) ;
(b) if (αn ) ⊂ C, αn → 0 in R and xn → a in (X, k · k), then αn xn → 0 ;
(c) [Product rule] if (λn ) ⊂ C, λn → λ in C and xn → a in (X, k · k), then
λn xn → λa .
(3) [Continuity of the norm] Let (xn ) ⊂ X and a ∈ X. If
xn → a in (X, k · k) ,
then
lim kxn k = kak in C .
n→∞
Consequently, the function
(X, k · k) → (C, | · | = usual norm)
x 7→ kxk ,
is a continuous function (from (X, k · k) to (C, | · |).
2.3. Sequences of functions. Here, we will clarify some notions related to the space
of continuous functions C([a, b]). This is a normed space as seen in Example 2.2. However,
there are many notions of convergence associated with this space. We discuss two of these
notions here.
Definition 2.11. [Pointwise and uniform convergence] Assume that −∞ < a <
b < ∞ and for all n ≥ 1, fn : [a, b] 7−→ R is a continuous function.
(1) We say that the sequence (fn ) converges pointwise to a function f on [a, b] if for
every t ∈ [a, b] (fixed) we have fn (t) → f (t) as n → ∞.
(2) Let (fn ) ⊂ C([a, b], R), we say that the sequence (fn ) converges uniformly to a
function f on [a, b] if
kfn − f k∞ −→ 0 as n → ∞.
Remark 2.12.
(1) Notice that the uniform convergence implies the pointwise convergence. The
opposite is not necessary true. Consequently, in order to study the uniform
convergence of a sequence of functions it is recommended to follow the following
steps:
(a) Find the pointwise limit f of the sequence (fn ).
(b) Study the limit of kfn − f k∞ as n tends to +∞.
 3. BOUND SETS IN NORMED SPACE 9

(2) It is worth noting that if the sequence (fn ) ⊂ C([a, b], R) converges uniformly to
a function f on [a, b], then the then f ∈ C([a, b], R) (see Theorem 7.4). Con- 
sequently, the uniform convergence means convergence in C([a, b], R), k · k∞ ,
where k · k∞ is the norm introduced in Example 2.2.
(3) The norm k · k∞ is also called the norm of uniform convergence.


3. Bound sets in normed space

Since normed spaces are metric spaces (and the converse, in general, is not true), we
get the important class of open and closed balls.
Definition 3.1. [Balls]
Let (X, k · k) be a normed space, a ∈ X and r > 0.
(1) The open ball of center a and radius r (in (X, k · k)) is the following subset
Bk·k (a, r) = {x ∈ X : kx − ak < r} .
(2) The closed ball of center a and radius r (in (X, k · k)) is the following subset
Bk·k [a, r] = {x ∈ X : kx − ak ≤ r} .
(3) The sphere of center a and radius r (in (X, k · k)) is the following subset
Sk·k [a, r] = {x ∈ X : kx − ak = r} .

Remark 3.2.
(1) In many situations, it is customary to skip the reference to the norm k · k from
the notation of the balls and simply write B(a, r) and B[a, r].
(2) [The unit balls] For a = 0 and r = 1, the balls B(0, 1) and B[0, 1] are called
the unit open and closed balls respectively.
The algebraic structure in a vector space induce a geometric property which is the
convexity. Recall that X is assumed a vector space (on the field K ∈ {R, C}).
Definition 3.3. [Convex set]

(1) [Intervals in X] Let a ∈ X and b ∈ X. The interval [a, b] in X is the following

subset:
[a, b] = {ta + (1 − t)b : t ∈ R and 0 ≤ t ≤ 1} .
(2) Let A ⊂ X. The set A is said to be convex if for all a ∈ A and b ∈ A, [a, b] ⊂ A.
Notice that the notions of convex domain is defined through the vector space structure
only, i.e. norms are not involved. In that respect, these notions are understood to be
geometric/algebraic properties (i.e. not analytic properties).
In the next theorem, we see that convexity is a necessary condition for a subset to
be a ball in a normed space. Put another way, balls in a normed space have to obey a
geometric condition, namely convexity.
Theorem 3.4. In a normed space, the open and closed balls are convex subsets.
10 1. NORMED VECTOR SPACES

Proof. Let (X, k · k) be a normed space, a ∈ X and r > 0. We will prove that B(a, r)
and B[a, r] are convex. Let x ∈ B(a, r), y ∈ B(a, r) and t ∈ [0, 1]. We have,


tx+(1−t)y −a = t(x−a)+(1−t)(y−a) ≤ tkx−ak+(1−t)ky−ak < tr+(1−t)r = r ,
i.e. tx + (1 − t)y ∈ B(a, r). This proves that B(a, r) is convex. The proof of the convexity
of B[a, r] is similar. 
Remark 3.5. A potential application of Theorem 3.4 can be as follows. If A ⊂ X is
not convex, then A can not be a (closed or open) ball in X relative to any norm. Now if
d is a distance in X such that Bd (a, r) is not convex for some a ∈ X and r > 0, then the
distance d is not induced by a norm.
In a normed space, the property of boundedness takes the pleasant form described in:
Theorem 3.6. [Bounded subsets] Let (X, k · k) be a normed space. The following
two statements are equivalent:
(1) A ⊂ X is bounded in (X, k · k) ;
(2) ∃ C > 0, ∀ x ∈ A, kxk ≤ C .
Proof. Suppose that A ⊂ X. If A is bounded in (X, k · k), then A ⊂ B[a, M ] for
some a ∈ X and M > 0. Let C = kak+M . We have, for all x ∈ A, kxk ≤ kx−ak+kak ≤
M + kak = C. Conversely, suppose that, for all x ∈ A, kxk ≤ C. We have A ⊂ B[0, C].
Hence A is bounded in (X, k · k). 
Exercise 1.1. Let (X, k · k) be a normed space.
(1) Let S ⊂ X and λ ∈ C. Show that, if S is bounded in (X, k · k), then the set
λS = {λx : x ∈ S} is bounded in (X, k · k).
(2) Let S ⊂ X and T ∈ X. Show that, if S and T are bounded in (X, k · k), then
the set S + T = {x + y : x ∈ S and y ∈ T } is bounded in (X, k · k).
In the next theorem, we will see that the algebraic operations in a normed space will
force it to be unbounded.
Theorem 3.7. [A normed space is unbounded]
Let (X, k · k) be a normed space and X 6= {0}.
(1) For every α > 0, there exists x ∈ X such that kxk = α.
(2) X is not bounded in (X, k · k).
(3) Every non-null vector subspace of X is not bounded in (X, k · k).
Proof. The proof involves the trick of normalization of a non-zero vector. Let α > 0,
α
y ∈ X and y 6= 0. Let x = y (notice the use of the algebraic operations in X). Clearly,
kyk
x ∈ X and kxk = α.
If X is bounded, then there exists M > 0 such that, for all x ∈ X, kxk ≤ M . Select
x ∈ X such that kxk = M + 1. Consequently, M + 1 = kxk ≤ M , i.e. 1 ≤ 0 which is a
contradiction.
Every non-null subspace is a normed space, hence not bounded by the above consid-
erations. 
Remark 3.8. Since compact metric spaces are bounded, we get in light of Theorem 3.7
that every normed space (X, k · k) is not compact.
 5. EQUIVALENT NORMS 11

4. Product of normed spaces

In this section we show that the (cartesian) product of a finite collection of normed
spaces is a normed space too, called product normed space henceforth. However, there
are various ways to define a norm in the product space (all of which lead to equivalent
norms, see Section 5 below). We will focus on three possible constructions.
Theorem 4.1. [Three norms in the product space] Let (X1 , N1 ), (X, N2 ),· · · ,
(X` , N` ) be normed spaces, and X = X1 × X2 × · · · × X` . The mappings defined, for every
u = (u1 , · · ·, x` ) ∈ X, by:
N∞ (u) = max{Ni (ui ) : 1 ≤ i ≤ `} , (4.1)
`
X
N1 (u) = Ni (ui ) , (4.2)
i=1
v
u `
uX
2
N2 (u) = t Ni (ui ) . (4.3)
i=1

are norms on X.
The proof of Theorem 4.1 is left as an exercise.

Example 4.2. [Fundamental example in the case of a finite dimensional

space] Let (R, | · |) be the n.v.s endowed with its usual norm, i.e. the absolute value
function | · |, and let E = Rn = R | × ·{z
· · × R}. The mappings defined, for every x =
n times
(x1 , · · ·, xn ) ∈ E, by
N∞ (x) = max{|xi | : 1 ≤ i ≤ n} (norm infinity); , (4.4)
v
u n
uX
N2 (x) = t |xi |2 (Euclidian norm); , (4.5)
i=1
n
X
N1 (x) = |xi | (norm sum). (4.6)
i=1

are norms on E, called the usual norms on Rn .

5. Equivalent Norms
In this section, we will study the notion of equivalent norms. These are norms having
distinct expressions but defined on the same vector space. We shall see that this notion
is highly non-trivial in infinite dimensional spaces.
5.1. Generalities.
Definition 5.1. [Equivalent norms]
Let N1 and N2 be two norms in a vector space X. The norms N1 and N2 are said to
be equivalent if there exist two constants 0 < c1 < c2 such that,
∀ x∈X, c1 N1 (x) ≤ N2 (x) ≤ c2 N1 (x) . (5.1)
12 1. NORMED VECTOR SPACES

Example 5.2. The norms N∞ , N1 and N2 , defined in Theorem 4.1, are equivalent on
the product space X. More precisely we have
√
∀ x ∈ X , N∞ (x) ≤ N1 (x) ≤ ` N2 (x) ≤ ` N∞ (x) .
Remark 5.3.
(1) In (5.1), the constants c1 and c2 are fixed while the point x is allowed to vary in
X. The strength of (5.1) is that the inequality remains true when x varies. In
other words, the constants c1 and c2 do not depend on the variable x.
(2) If N1 and N2 are norms on X, then they induce two distances dN1 and dN2
respectively. The norms N1 and N2 are equivalent if and only if the distances
dN1 and dN2 are equivalent too.
(3) Consequently, when N1 and N2 are equivalent norms, then they induce the same
topology, i.e. TN1 = TN2 . The converse is also true. We will present an elegant
proof of this in Chapter 2. Another elementary proof will be given in Exercise 3.
5.2. Case of finite dimension. In this section we discuss finite dimensional normed
vector spaces. These spaces share many properties with the Euclidean spaces, in particular
Theorems 5.5 and 6.1.
Our discussion of finite dimensional spaces relies on:
Lemma 5.4. Let (X, k·k) be a normed vector space of finite dimension d. Suppose that
{e1 , · · · , ed } is a basis of X. For all u = (u1 , · · · , ud ) ∈ Rd , define f (u) = u1 e1 +· · ·+ud ed .
(1) The function N (u) = kf (u)k is a norm on Rd ;
(2) ∃ c1 , c2 > 0 such that, ∀ u ∈ Rd , c1 kuk2 ≤ N (u) ≤ c2 kuk2 , where k · k2 is the
Euclidean norm in Rd . ;
(3) The function f : (Rd , k · k2 ) → (X, k · k) is a homeomorphism, i.e. f and f −1 are
continuous.
Proof. That N is a norm on Rd is left as an exercise to the reader. The existence of
the constants c1 and c2 follows from Theorem 5.5 which states that the norms N (·) and
k · k2 are equivalent.
Since {e1 , · · · , ed } is a basis of X, then f is bijective. The continuity of f and f −1
follows from the inequality in Item (2) above; actually, f and f −1 are Lipschitz functions.

Theorem 5.5. [Equivalent norms in finite dimension spaces] Let N be a norm
in Rd . The norm N and the Euclidean norm
q
x = (x1 , x2 , · · · , xd ) 7→ kxk2 = x21 + x22 + · · · + x2d
are equivalent norms.
Proof. Step 1. Let
e1 = (1, 0, · · · , 0) , e2 = (0, 1, 0, · · · , 0) , ed = (0, · · · , 0, 1) ,
be the canonical basis of R . For all x = (x1 , x2 , · · · , xd ) ∈ Rd , we have
d

d
X
x= xk ek ,
k=1
and by triangle’s inequality we have,
X
N (x) ≤ |xk |N (ek ) .
k=1
 5. EQUIVALENT NORMS 13

Let M = max N (ek ). We have,

1≤k≤d
X X
N (x) ≤ |xk |N (ek ) ≤ M |xk |
k=1 k=1
√ (5.2)
≤ M d kxk2 [By Minkowski’s inequality] .
Step 2. Let S = {x = (x1 , x2 , · · · , xd ) ∈ Rd : kxk2 = 1}. By Theorem 6.1, the
metric space (S, d2 ) is compact. By Step 1, the function
N : (S, d) → (R, | · |) ,
is a Lipschitz function, hence continuous. Since (S, d) is compact and N is continuous on
S, the function N attains its minimum, hence
∃ a∈S, N (a) = inf{N (x) : x ∈ S} .
Since a ∈ S, then kak2 = 1 > 0 and a 6= 0. Thus N (a) > 0. Let m = N (a). We have,
∀ u∈S, N (u) ≥ N (a) = m > 0 .
1
Now, let x ∈ Rd and x 6= 0. Let u = kxk2
x. We have kuk2 = 1, i.e. u ∈ S. Thus, we may
write
N (u) ≥ m > 0 . (5.3)
1
But, since N is a norm, N (u) = kxk2
N (x).That way, we deduce from (5.3),
N (x) ≥ mkxk2 ,
and this inequality is true for all x ∈ Rd \ {0}. When x = 0, the inequality is obviously
true. 
Theorem 5.6. [Norms on finite dimensional spaces]
If X is a finite dimensional vector space, then all the norms in X are equivalent.
Proof. Let N1 and N2 be two norms on X. We will prove that N1 and N2 are
equivalent. Let B = {e1 , · · · , ed } be a basis of X. The function f : Rd → X defined by
f (u1 , · · · , ud ) = u1 e1 + · · · + ud ed is bijective. By Lemma 5.4, the functions
n1 (u) = N1 (f (u)) and n2 (u) = N2 (f (u))
are norms in Rd , hence equivalent by Theorem 5.5. This yields that the norms N1 and
N2 are equivalent as well, since the function f is bijective. 
5.3. Case of infinite dimension. Here we claim that, in infinite dimensional normed
spaces, norms may not be equivalent in general. To prove that, we consider the space
X = C[0, 1] of real-valued continuous functions on the interval [0, 1] with that the follow-
ing norms:
Z 1
kf k1 = |f (x)| dx , kf k∞ = max |f (x)| . (5.4)
0 x∈[0,1]

By considering the sequence of functions (fn ) ⊂ X defined by,

1 − nx if 0 ≤ x ≤ n1 ,

fn (x) =
0 if x ≥ n1 ,
kfn k∞
we can prove that k · k1 and k · k∞ are not equivalent. Actually, lim = ∞.
n→∞ kfn k1
14 1. NORMED VECTOR SPACES

6. Compact sets in normed spaces

In this section, we discuss the compactness in a normed vector spaces. In what follows,
we will distinguish between the finite and infinite dimensional vector spaces.
6.1. Case of finite dimension. We start by recalling the:
Theorem 6.1. [Hein-Borel theorem]
Let S ⊂ Rd and d2 be the Euclidean distance in Rd defined as follows,


(x, y) = (x1 , x2 , · · · , xd ) , (y1 , y2 , · · · , yd )
p
7→ d2 (x, y) = (x1 − y1 )2 + (x2 − y2 )2 + · · · + (xd − yd )2 .
The metric space (S, d2 ) is compact if and only if S is a closed and bounded set in (R, d2 ).
The Hein-Borel theorem admits the following extension:
Theorem 6.2. [Compact subsets of finite dimensional spaces]
Let (X, k·k) be a finite dimensional normed space. A set S ⊂ X is compact in (X, k·k)
if and only if it is closed and bounded in (X, k · k).
Proof. It is necessary to prove the sufficient condition. Let S ⊂ X be closed and
bounded in (X, k · k). We will show that S is compact in (X, k · k).
Let B = {e1 , · · · , ed } be a basis of X and define the function f : Rd → X by
f (u1 , · · · , ud ) = u1 e1 + · · · + ud ed . By Lemma 5.4, f is bijective and the functions f
and f −1 are continuous. Furthermore, Items (2) and (3) in Lemma 5.4 yields that the set
A = f −1 (S) is bounded and closed in Rd . By Theorem 5.5, we get that A is compact in
Rd . The continuity of f yields that f (A) = S is compact in (X, k · k). 
Our next result concerns finite dimensional subspaces.
Theorem 6.3. [Finite dimensional subspaces are closed]
Let (X, k · k) be a normed space and M ⊂ X be a subspace of X. If the dimension of
M is finite, then M is closed in (X, k · k).
Proof. Suppose that d = dim(M ) < ∞. There exists a basis B = {e1 , · · · , ed } of M .
The function
f : Rd → M , f (u1 , · · · , ud ) = u1 e1 + · · · + ud ed
is bijective and linear. By Lemma 5.4, the function
N (u) = kf (u)k
is a norm on Rd and
∃ c1 , c2 > 0 , ∀ u ∈ Rd , c1 kf (u)k ≤ kuk2 ≤ c2 kf (u)k , (6.1)
d
where k · k2 is the Euclidean norm on R .
Now we prove that M is closed in (X, k · k). It is enough to show that M ⊂ M . Let
u ∈ M . There exists a sequence (un ) ⊂ M such that un → u. In particular, (un ) is a
Cauchy sequence. By (6.1), we get that (f (un )) ⊂ Rd is a Cauchy sequence in (Rd , k · k2 ),
hence convergent to some x ∈ Rd . Since the function f is bijective, we may define
w = f −1 (x) ∈ M . Again, (6.1) yields that un → w in (X, k · k), since f (un ) → x = f (w)
in (R2 , k · k2 ). Since the limit of a sequence is unique, we get that u = w. As w ∈ M , we
get that u ∈ M . 
Complementary to Theorem 6.3 is:
 6. COMPACT SETS IN NORMED SPACES 15

Theorem 6.4. [Closure of a subspace]

Let (X, k · k) be a normed space and M ⊂ X. If M is a subspace of X, then its closure
M is a subspace of X and dim(M ) = dim(M ).
Proof. That M is a subspace is left as an Exercise for the reader. We prove that
dim(M ) = dim(M ). If dim(M ) = ∞, the equality is obvious, since M ⊂ M .
If dim(M ) < +∞, then M is closed by Theorem 6.3. Consequently, M = M and
dim(M ) = dim(M ). 

6.2. Case of infinite dimension. In the Euclidean space Rd , the Hein-Borel theo-
rem tells us that the unit ball B[0, 1] is compact. This fact will not be true in an infinite
dimensional space. The Riesz compactness theorem will clarify this, as it will state that
the unit ball in a normed space is compact if and only if the space is finite dimensional.
The lengthy proof of this theorem will occupy the whole of this section.
Theorem 6.5. [Riesz’s theorem]
Suppose that (X, k · k) is a normed space. The unit closed ball B[0, 1] in (X, k · k) is
compact if and only if the vector space X has a finite dimension (i.e. dim(X) is finite).
The proof of Theorem 6.5 is split into several lemmas. In the rest of this section,
(X, k · k) denotes a normed space.
Lemma 6.6. Suppose that M is a closed proper subspace of X. For all  ∈]0, 1[, there
exists y ∈ X such that
kyk = 1 and dist(y, M ) ≥ 1 −  .
Proof. Since M 6= X, there exists x ∈ X \ M . Let α = dist(x, M ). Since M is closed
and x 6∈ M , M = M and α = dist(x, M ) > 0.
α
Let  ∈]0, 1[ be fixed. Note that α = dist(x, M ) = inf{kx−mk : m ∈ M } and α ≤ 1− .
By the definition of the infimum, we get m0 ∈ M such that α ≤ kx − m0 k ≤ α/(1 − ).
Now, we define y = kx − m0 k−1 (x − m0 ). Clearly kyk = 1. Let m ∈ M . We will prove
that ky − mk ≥ 1 − . Since M is a subspace, m0 + kx − m0 km ∈ M . Consequently, by
definition of α = dist(x, M ),

ky − mk = kx − m0 k−1 x − m0 − kx − m0 km ≥ kx − m0 k−1 α .

Since kx−m0 k ≤ α/(1−), kx−m0 k−1 ≥ α−1 (1−) and consequently, ky−mk ≥ 1−. 
Lemma 6.7. Suppose that 0 <  < 1. If dim(X) = +∞, there exist a sequence (un )
in X and a sequence of subspaces (Mn ) such that, for all n ≥ 2,
• kun k = 1 ;
• un ∈ Mn ;
• dist(un , Mn−1 ) ≥ 1 −  ;
• Mn ⊂ Mn+1 and dimMn = n .
Proof. Choose u1 ∈ X such that ku1 k = 1. Let M1 = Span(u1 ). M1 is a subspace
of X, dim(M1 ) = 1 and M1 6= X since dim(X) = +∞. Furthermore, M1 is closed
by Theorem 6.3. Applying Lemma 6.6, we obtain u2 ∈ X such that ku2 k = 1 and
dist(u2 , M1 ) ≥ 1 − ε.
Next, we define M2 = Span(u1 , u2 ). Clearly, M1 ⊂ M2 , dim(M2 ) = 2 and M2 is closed
by Theorem 6.3. Furthermore, M2 6= X since dim(X) = +∞. Applying Lemma 6.6,
16 1. NORMED VECTOR SPACES

we get u3 ∈ X such that ku3 k = 1 and dist(u3 , M2 ) ≥ 1 − ε. Next we define M3 =

Span(u1 , u2 , u3 ).
We continue this construction by induction as follows. Suppose that we constructed
u1 , u2 , · · · , un and Mk = Span(u1 , · · · , uk ) for k ∈ {1, · · · , n}. Applying Lemma 6.6, we
get un+1 ∈ X such that kun+1 k = 1 and dist(un+1 , Mn ) ≥ 1 − ε. We define Mn+1 =
Span(u1 , · · · , un+1 ). 
Proof of Theorem 6.5. Note that B[0, 1] is closed and bounded in (X, k · k). If
dimX < +∞, then B[0, 1] is compact by Theorem 6.2.
Next, we suppose that B[0, 1] is compact. We will prove that dim(X) < +∞. Suppose
by contradiction that dim(X) = +∞. Let  ∈]0, 1[ and (un ) be the sequence constructed
in Lemma 6.7. Since kun k = 1 for all n, (un ) ⊂ B[0, 1]. Since B[0, 1] is compact, (un ) has
a convergent subsequence (unk ).
The subsequence (unk ) is a Cauchy sequence. In particular, there exists k0 such that
1−
for p = nk0 and q = nk0 +1 , kuq − up k <.
2
But, since q > p and up ∈ Mp , kuq − up k ≥ dist(uq , Mp ) ≥ 1 − . Consequently,
1−
0<1−< .
2
This is the desired contradiction. 

7. Banach spaces
In this section, we discuss the completeness in the case of infinite dimensional normed
vector spaces. More precisely, we study the convergence of Cauchy sequences in (C[0, 1], k · kp )
for p ∈ {1, 2, ∞}.
We begin by the following definition:
Definition 7.1. [Banach spaces] Let (X, k · k) be a normed space and dk·k be the
distance induced by k · k. The space (X, k · k) is said to be a Banach space if the metric
space (X, dk·k ) is complete, i.e. Cauchy sequences in (X, k · k) are convergent.
Now we recall the following theorem:
Theorem 7.2. In any normed vector space, we have
(1) Every convergent sequence is a Cauchy sequence. The opposite is not necessary
true.
(2) Every Cauchy sequence is bounded.
(3) Every Cauchy sequence that has an adherent value is convergent 1.
Remark 7.3. In Rn any bounded sequence has an adherent value (this is asserted
by the Bolzano-Weirstrass theorem). Consequently, the above theorem 7.2 ensures that
every Cauchy sequence in Rn is convergent i.e. Rn is a Banach space. On other hand,
due to the non-compactness of the closed ball in every infinite dimensional vector space
X (see Theorem 6.5), it can not be guaranteed that any bounded sequence in X has an
adherent value. Hence the question of convergence of Cauchy sequences in (the infinite
dimensional space) X is far from trivial.
1Apoint a is an adherent (or cluster) value of a sequence (xn ) if there exists a subsequence (xnk )
such that xnk → a.
 7. BANACH SPACES 17

In the following two theorems we will find that the space (C[0, 1], k · kp ) is a Banach
space for p = ∞ unlike the cases of p ∈ {1, 2} (see Example 2.2 for the definition of k · kp ,
for p = 1, 2). The main reason for this is that the space C([0, 1]) is an infinite dimensional
vector space.
Theorem 7.4. The space (C([0, 1]), k · k∞ ) is a Banach space, i.e. Cauchy sequences
are convergent.
Proof. Let (fn ) be a Cauchy sequence in C([0, 1]), k · k∞ ).
To prove Theorem 7.4,
we must find f ∈ C([0, 1]) such that fn → f in C([0, 1]), k · k∞ .
Step 1. Definition of the function f : [0, 1] → R :
Let ε > 0. Since (fn ) is a Cauchy sequence in (C([0, 1]), k · k∞ ), then there exists
n0 ∈ N such that,
ε
∀ n ≥ n0 , ∀ p ∈ N , kfn+p − fn k∞ < . (7.1)
6
By definition of k · k∞ , we infer from (7.1),
ε
∀ x ∈ [0, 1] , ∀ n ≥ n0 , ∀ p ∈ N , |fn+p (x) − fn (x)| ≤ kfn+p − fn k∞ < < ε . (7.2)
6
Consequently, for all x ∈ [0, 1], the sequence (fn (x)) is a Cauchy sequence in R, hence
convergent. Let us denote by f (x) the limit of the sequence fn (x) in R. That way, we get
a function f : [0, 1] → R.
Step 2. The function f is in the space C([0, 1]) :
Let x0 ∈ [0, 1]. We will prove that the function f defined is Step 1 is continuous at x0 .
Let ε > 0 and n0 ∈ N be as in Step 1. Since the function fn0 is continuous at x0 , then,
there exists δ > 0 such that,
  ε
x ∈ [0, 1] and |x − x0 | < δ =⇒ |fn0 (x) − fn0 (x0 )| < .
6
In the sequel, we suppose that x ∈ [0, 1] and |x − x0 | < δ. By (7.2), we have, for all p ∈ N,
|fn0 +p (x) − fn0 +p (x0 )| ≤ |fn0 +p (x) − fn0 (x)| + |fn0 (x) − fn0 (x0 )| + |fn0 (x0 ) − fn0 +p (x0 )|
ε ε ε ε
< + + = .
6 6 6 2
Since the inequality mentioned previously is true for all p ∈ N, then
ε
lim |fn0 +p (x) − fn0 +p (x0 )| ≤ < ε .
p→∞ 2
By Step 1, we deduce that,
|f (x) − f (x0 )| < ε .


Step 3. fn → f in C([0, 1]), k · k∞ :
By (7.2) ,
ε
∀ n ≥ n0 , ∀ x ∈ [0, 1] , lim |fn+p (x) − fn (x)| ≤ .
p→∞ 6
By Step 1, we get,
ε
∀ x ∈ [0, 1] , ∀ n ≥ n0 , |f (x) − fn (x)| ≤ .
6
By definition of the distance d, we deduce that,
ε
∀ n ≥ n0 , kf − fn k∞ ≤ < ε .
6
This proves that lim kf − fn k∞ = 0. 
n→∞
18 1. NORMED VECTOR SPACES

Theorem 7.5. Let p ∈ {1, 2}.

(1) The space (C([0, 1]), k·kp ) is not a Banach space, i.e. there exist divergent Cauchy
sequences.
(2) There exist bounded sequences in (C([0, 1]), k · kp ) that do not have convergent
subsequences in (C([0, 1]), k · kp ).
Here k · kp is defined in (5.4).
Proof. Consider the sequence of functions (fn )n≥3 ⊂ C([0, 1]) defined as follows:

1
 if 0 ≤ x ≤ 21 ,
2n
x − n1 if 12 < x ≤ 21 + n1 ,


fn (x) = n−2
if 21 + n1 < x ≤ 1 .

0

The reader is asked to verify that the sequence (fn ) is Cauchy in (C([0, 1]), k · kp ). We
will prove that (fn ) is not convergent in (C([0, 1]), k · kp ). If (fn ) where convergent, then
there exists f ∈ C([0, 1]) such that kfn − f kp → 0. However,
Z 1/2 Z 1
p
∀ n ≥ 3, 0 ≤ |1 − f (x)| dx ≤ |fn (x) − f (x)|p dx .
0 0
Z 1
Sending n to infinity, we get that |1 − f (x)|p dx = 0. Consequently, the continuity of
0
f yields that f (x) = 1 for all x ∈ [0, 12 ].
Now, let a ∈] 21 , 1]. We will prove that f (a) = 0. Select n0 ≥ 3 such that 21 + n10 < a ≤ 1.
Consequently,
Z 1 Z 1
p
∀ n ≥ n0 , 0 ≤ |f (x)| dx ≤ |fn (x) − f (x)|p dx .
a 0

By sending n to infinity, we get that f (a) = 0. This is true for all a ∈] 12 , 1]. By continuity
of f , we get f ( 12 ) = lima→ 1 + f (a) = 0. On the other hand, we proved that f ( 12 ) = 0. This
2
yields the desired contradiction.
The same proof can be applied on every subsequence of (fn ). That is, if (fnk ) is a
convergent subsequence of (fn ), then it will converge to a function f ∈ C([0, 1]) that
+
satisfies f ( 12 ) = 1 and f ( 12 ) = 0. 
Remark 7.6. (The space L2 ([0, 1])
According to Theorem 7.5, the space (C([0, 1], k · k2 ) is not a Banach space. However,
this issue can be resolved as follows. One can construct a larger space of functions,
denoted by L2 ([0, 1]) and called the space of square integrable functions, such that, for
every function f ∈ L2 ([0, 1]), the integral
Z 1
|f (x)|p dx exists .
0

This space is needed in Chapter 4 later. Obviously, the space L2 ([0, 1]) contains C([0, 1]).
All we need to know about L2 ([0, 1]) is the following property. If f ∈ L2 ([0, 1]), then there
exists a sequence of continuous functions (fn )n≥1 ⊂ C([0, 1]) such that
• (fn ) Z
is a Cauchy sequence in (C([0, 1]), k · k2 ) ;
1
• lim |fn (x) − f (x)|2 dx → 0 .
n→∞ 0
 7. BANACH SPACES 19

A deep theorem asserts that (L2 ([0, 1]), k·k2 ) is a Banach space. However, the construction
of the space L2 is a subtle problem that is beyond the scope of this course. One way to
define L2 is through a completion process by collecting all Cauchy sequences in C([0, 1])
in equivalent classes.
20 1. NORMED VECTOR SPACES

Exercises

Exercise 1. In each of the following cases, check if the function N defines a norm in the
space X.
(1) X = R and N (x) = sin(x) ;
(2) X = R and N (x) = x2 + |x| ;
(3) X = R2 and N (x, y) = 2|x| + 3|y| ;
(4) X = R2 and N (x, y) = sup |x + ty| .
t∈[0,1]

Exercise 2. Let (E, k · k) be a normed space, x, y ∈ E.

(1) Show that kxk − kyk ≤ kx − yk.

(2) Let λ ∈ R+ . Show that, if x 6= 0 and y 6= 0, then the following inequality holds
true:
kλxk
λx − y ≤ 2λkx − yk.
kyk

Exercise 3. [A Geometric way to visualize equivalence norms]

Let N1 and N2 be two norms in a vector space X, and 0 < c1 < c2 . Show that the
following three statements are equivalent:
(a) ∀ v ∈ X,
c1 N1 (v) ≤ N2 (v) ≤ c2 N1 (v)

.
1 1
(b) BN1 0, c2 ⊂ BN2 (0, 1) ⊂ BN1 0, c1 .
(c) BN1 0, c12 ⊂ BN2 [0, 1] ⊂ BN1 0, c11 .
   

Exercise 4. Suppose that (X, k · k) is a normed space, p and q the two functions defined
by,
kxk
p(x) = , q(x) = ln(1 + kxk) , (x ∈ X) .
1 + kxk
(1) Are p and q norms in X ?
(2) Are the functions dp (x, y) = p(x − y) and dq (x, y) = q(x − y) metrics in X ?

Z 1 1/2
2
Exercise 5. Let X = C([0, 1]) be normed by kf k2 = |f (x)| dx . Consider the
√ 0
two subsets Y = {ax : a ∈ R} and B = {ax + x : 0 ≤ a ≤ 1} .
(1) Show that Y is not bounded.
(2) Let f0 (x) = x2 . Find dist(f0 , Y ) and a function g ∈ Y such that kf0 − gk2 =
dist(f0 , Y ) .
 EXERCISES 21

(The function g is called the best approximation in Y of the function f0 .)

(3) Show that B is bounded and compute its diameter relative to k · k2 .

Exercise 6. Let (X, k · k) be a metric space, a ∈ X and r > 0. Show that:

(1) B(a, r) = B[a, r] ;
(2) B[a, r]◦ = B(a, r) ;
(3) ∂B(a, r) = ∂B[a, r] = {u ∈ X : ku − ak = r} ;
(4) The diameters of B(a, r) and B[a, r] are equal to 2r ;

Exercise 7. Suppose that (X, k · k) is a normed space and V a subspace of X such that
V 6= X. Show that V ◦ = ∅. (Hint: if x ∈ V ◦ , y ∈ V and y 6= 0, then w = x + 2kyk
r
y∈
B(x, r)).

Exercise 8. Let (X, k · k) be a normed space.

(1) Show that:

(a) B(0, r) + B(0, r0 ) = B(0, r + r0 ) ; B(x, r) + B(y, r0 ) = B(x + y, r + r0 ) ;
B(λx, λr) = λB(x, r) ;
(b) If A ⊂ X is convex, then A◦ is convex too.
(2) Let A ⊂ X be a convex set. Show that A is convex.
(3) Let A be a subspace of X, show that A is a subspace of X.
(4) Suppose that A = B(0, 1) \ (B[0, 1/2]). Show that:
(a) A is open ;
(b) A is not convex.

Exercise 9. Consider X = C[0, 1] endowed with the norm of uniform convergence

kf k∞ = sup |f (x)|.
x∈[0,1]

Study in (X, k · k∞ ) the convergence of the following sequences:

(a) fn (t) = tn − tn+1 , (b) gn (t) = (n + 1)(tn − tn+1 ).

Exercise 10. Let X = C[−1, 1] endowed with the norm of uniform convergence. Study
22 1. NORMED VECTOR SPACES

in (X, k · k∞ ) the convergence of the following sequences:

 −n
(t + n1 ) if −1 ≤ t ≤ −1

−1 ,

 −1 if −1 ≤ t ≤ n
, 
 n−1 n

 

 
fn (t) = nt if −1 n
< t < 1
n
, gn (t) = 0 if −1
n
< t < n1 ,

 

 
1  n (t − 1 ) if 1
 
 1 if n
≤ t ≤ 1. n−1 n n
≤ t ≤ 1.

Exercice 11. Let X = C[0, 1] and F = {f ∈ X : f (0) = 1} a subset of X. We recall the

following two norms on X

Z 1
kf k∞ = sup |f (x)| and kf k1 = |f (x)| dx.
x∈[0,1] 0

(1) Show that F is closed in (X, k · k∞ ). What can we say about (X, k · k∞ )?
(2) Let (fn )n≥1 the sequence of real functions defined on [0, 1] by:

 nt + 1 if t ∈ [0, n1 ],
fn (t) =
2 if t ∈] n1 , 1].


(a) Show that (fn )n≥1 converges in (X, k · k1 ) to the function f = 2.

(b) Deduce that F is not closed in (X, k · k1 ).
n Z 1 o
Exercice 12. Let X = C[0, 1] and let F = f ∈ C[0, 1]; f (x)dx = 0 .
0
(1) Show that, for every f ∈ X, kf k1 ≤ kf k∞ .
(2) Deduce that if F is closed in (X, k · k1 ) then F is closed in (X, k · k∞ ).
(3) Show that F is closed in (X, k · k1 ) and in (X, k · k∞ ).

Exercise 13. Let E = C([0, 1], R) the vector space of continuous functions over [0, 1].
Consider the sequence of functions (fn ) defined by:

 1 − nx if 0 ≤ x ≤ n1
fn (x) = .
0 if n1 ≤ x ≤ 1


(1) Show that (fn ) is not a Cauchy sequence in (E, k · k∞ ).

(2) Show that (fn ) is a Cauchy sequence in (E, k · k1 ).

Exercise 14. Let X = C([0, 1]) be the space of continuous functions from [0, 1] to R.

Define the three functions in X,

Z 1
kf k1 = |f (x)| dx ,
0

kf k∞ = sup |f (x)| ,
x∈[0,1]
 EXERCISES 23

and
Z 1 1/2
2
kf k2 = |f (x)| dx .
0
(1) Show that k · k1 and k · k∞ are norms in X.
(2) Show that k · k2 is a norm in X. (Hint: Prove the following Cauchy-Schwarz
R1
inequality, 0 f (x)g(x) dx ≤ kf k2 kgk2 ).
(3) Give the distances assigned to each of the norms k · k1 , k · k∞ , k · k2 .
(4) Let n ∈ N∗ and define the function fn : [0, 1] → R as follows,
1 − nx if 0 ≤ x ≤ n1 ,

fn (x) =
0 if x ≥ n1 .
(a) Show that fn ∈ X ;
(b) Compute kfn k1 and kfn k∞ .
(c) Deduce that k · k1 and k · k∞ are not equivalent.

Exercise 15. Let X = C 1 [0, 1] be the set of real valued continuously differentiable

functions defined on the closed bounded interval [0, 1]. Consider the following mappings
on X Z 1
kf k1 = |f (t)|dt, kf k∞ = sup |f (t)|
0 t∈[0,1]

and
0
kf kX = kf k∞ + kf k1
(1) Show that k · kX is a norm on X.
(2) Show that for every f ∈ X we have
kf k∞ ≤ kf kX
(3) Show that k · k∞ is not equivalent to k · kX . (Hint: you may consider fn (x) =
sin(2πnx), n ∈ N)

Exercise 16. Let E = {f : [0, 1] → R of class C 2 : f (0) = f 0 (0) = 0}. Let N : E → R

be defined by  
N (f ) = sup |f 00 (x)| + |f (x)| .
[0,1]

(1) Show that N is a norm in E.

(2) Can we define N in C([0, 1]) ?
(3) Let n ≥ 2, fn (x) = xn and x ∈ [0, 1]. Calculate kfn k∞ and N (fn ).
(4) Deduce that k · k∞ and N are not equivalent norms in E.

Exercise 17. Let X = {C 1 [0, 1] : f (0) = 0}. For every f ∈ X, we define:

0 0
kf k1 = kf k∞ + kf k∞ and kf kX = kf + f k∞
24 1. NORMED VECTOR SPACES
0
(1) Integrate the differential equation f + f = 0. Deduce that k · kX is a norm over
X.
(2) By integrating by parts, show that, for every f ∈ X, we have
Z x
−x 0
f (x) = e (f + f )(t)et dt, ∀ x ∈ [0, 1].
0

Deduce that
0
kf k∞ ≤ (e − 1)kf kX and kf k∞ ≤ ekf kX .
(3) Deduce that the two norms k · k1 and k · kX are equivalent over X.
Z 2
Exercise 18. Let X = C[0, 2] endowed with the norm kf k1 = |f (x)|dx. Consider
0
the sequence


 0 if 0 ≤ t ≤ 1 − n1 ,



fn (t) = nt − n + 1 if 1 − n1 < t < 1,




 1 if 1 ≤ t ≤ 2.
(1) Show that (fn )n is a Cauchy sequence in (X, k · k1 ).
(2) Suppose that (fn )n converges to some function f in (X, k · k1 ).
(a) Show that f (x) = 1 for every x ∈ [1, 2].
Z 1− 1
n
(b) Show that |f (x)|dx −→ 0 as n −→ +∞. Deduce that f (x) = 0 for all
0
x ∈ [0, 1[.
(c) Deduce that (fn )n does not converge in (X, k · k1 ).
(3) What can you say about the space (X, k · k1 )?

Exercise 19. Let X = C([0, 1]). Show that (X, k · k∞ ) is a Banach space.

Exercise 20. Let X = C[0, 1] and let B = {f ∈ X, kf k∞ ≤ 1} and we define the

sequence of functions (fn ) ⊂ B by fn (x) = xn , n ∈ N.

(1) Show that B is closed and bounded in (X, k · k∞ ).
(2) Show that the sequence (fn ) does not converge uniformly on [0, 1].
(3) Deduce that B is not compact in (X, k · k∞ ).

Exercise 21. Consider the vector space `1 (R) defined by

( ∞
)
X
`1 (R) = x = (xk ) : xk ∈ R, |xk | < +∞ ,
k=1
 EXERCISES 25

∞
X
equipped with the norm kxk1 = |xk |. Let B = {x = (xk ) ∈ `1 (R), kxk1 ≤ 1} and we
k=1
define the sequence (xn ) ⊂ `1 by
1 if i = n

xni =
0 6 n.
if i =
(1) Show that B is closed and bounded in (`1 (R), k · k1 ).
(2) Show that for every n, we have xn ∈ B.
(3) Show that kxn − xm k1 = 2 for n 6= m. Deduce that (xn ) is not a Cauchy sequence
in B.
(4) Deduce that B is not compact in (`1 (R), k · k1 ).
(5) What can you say about the dimension of `1 (R)?

Exercise 22.

Consider the vector space `1 defined by

( ∞
)
X
`1 (R) = x = (xk ) : xk ∈ R, |xk | < +∞ ,
k=1
∞
X
equipped with the norm kxk1 = |xk |. Show that (`1 (R), k · k1 ) is a Banach space.
k=1
 CHAPTER 2

BOUNDED LINEAR TRANSFORMATIONS

This chapter is devoted to the study of continuous linear functions between normed
spaces and the spaces of such functions. Recall that a normed space carries an algebraic
and an analytic structure. These two structures are compatible in the following sense.
The analytic structure is defined by the distance function, the distance function is induced
by a norm, and the norm respects the vector space operations.
In the same way, when one considers functions between two normed spaces, a special
attention is given to those functions that are compatible with the algebraic structure,
i.e. linear functions. For these functions, it is natural to study their analytic properties,
i.e. continuity. This is a rich material in many aspects: There are examples of linear
functions that fail to be continuous; the study of spaces of continuous linear functions
is interesting in its own; spaces of continuous linear functions lead us to dual spaces; by
inspecting the dual of a given space, one may extract an additional information about the
space. In summary, the theory of bounded linear transformations plays a central role in
the development of functional analysis and its applications.

1. On continuous functions between normed spaces

It is assumed that the reader is familiar with the notion of continuous functions be-
tween two metric spaces and their basic properties. Here we stick to the special case of
continuous functions between two normed spaces. We review some concepts and facts
here.
Let (X, k · kX ) and (Y, k · kY ) be two normed spaces and F : X → Y be a function.
Some notions:
(1) Let a ∈ X. We say that F is continuous at a if
∀ ε > 0, ∃ δ > 0, ku − akX < δ =⇒ kF (u) − F (a)kY <  .
(2) We say that F is continuous (on X) if F is continuous at a, for all a ∈ X.
(3) We say that F is uniformly continuous if
∀ ε > 0, ∃ δ > 0, ku − vkX < δ =⇒ kF (u) − F (v)kY <  .
(4) We say that F is k-Lipschitz if, ∀ u, v ∈ X, kF (u) − F (v)kY ≤ kku − vkX .
(5) We say that F is a contraction if there exists k ∈ [0, 1[ such that F is k-Lipschitz.
(6) We say that F is a homeomorphism if F is invertible and the functions F and
F −1 are continuous.

Continuity and topology:

Continuity can be expressed using the topology (i.e. the class of open sets). Let
TX and TY be the classes of open subsets of X and Y respectively. The following two
statements are equivalent:
(a) F : X → Y is continuous ;
27
28 2. BOUNDED LINEAR TRANSFORMATIONS

(b) ∀ G ∈ TY , F −1 (G) ∈ TX .
Using the above property, we can verify the result summarized in the following exercise:
Exercise 2.1. Let (X, k · kX ) and (Y, k · kY ) be two n.v.s. and let F be a linear
surjective application from X into Y . Show that the following statements are equivalent:
(1) F is a homeomorphism from X into Y ;
(2) There exist α, β > 0 such that ∀ x ∈ X, αkxkX ≤ kF (x)kY ≤ βkxkX .
Consequently, in the case when N1 and N2 are two distinct norms in X, we get
TN1 = TN2 ⇐⇒ I : (X, N1 ) → (X, N2 ) is a homeomorphism ,
where I : u ∈ X 7→ I(u) = u ∈ X is the identity function, TN1 and TN2 denote the classes
of open subsets of (X, N1 ) and (X, N2 ) respectively.

The Banach fixed point theorem:

We recall one important fixed point theorem here. Let (X, k · k) a Banach space and
assume that F : (X, k · k) 7−→ (X, k · k) is a contraction. There exists a unique u ∈ X
such that F (u) = u.

This theorem have important applications to the theory of differential and integral
equations, namely on existence of solutions. Some exercises will be devoted to applications
of this sort.
2. Linear continuous functions
In this section, we specialize in studying continuity for linear functions between two
normed spaces. We will see that the algebraic property of a function being linear is not
enough to guarantee continuity.
Let us start by recalling the notion of a linear function:
Definition 2.1. Let X and Y be two vector spaces on the some field K = (R or C).
A function L : X → Y is said to be linear if,
∀ x, y ∈ X , ∀ α, β ∈ K, L(αx + βy) = αL(x) + βL(y) .
Remark 2.2.
(1) If L is a linear function, then L(0) = 0.
(2) If L1 : X → Y and L2 : X → Y are two linear functions, then the function
L1 + L2 : X → Y
x 7→ L1 (x) + L2 (x)
is a linear function.
(3) If L : X → Y is a linear function and α ∈ C, then the function
αL : X → Y
f 7→ α L(f )
is a linear function.
(4) The space of all linear functions from X to Y is a vector space.
(5) The identity function
I: X → X
f 7→ I(x) = x
is a linear function.
 2. LINEAR CONTINUOUS FUNCTIONS 29


Linear functions enjoy the property that continuity at one point guarantees continuity
in the space. That is a consequence of the algebraic property of linearity.
Theorem 2.3. Let (X, k · kX ) and (Y, k · kY ) be normed spaces and L : X → Y be a
linear function. The following four statements are equivalent:
(a) The function L is continuous at 0 (from (X, k · kX ) to (Y, k · kY )) ;
(b) The function L is continuous everywhere (from (X, k · kX ) to (Y, k · kY )) .
(c) The function L is uniformly continuous (from (X, k · kX ) to (Y, k · kY )) .
(d) ∃ M > 0, ∀ x ∈ X, kL(x)kY ≤ M kxkX .
Proof. First we prove (a)⇐⇒(b)⇐⇒(c). The implication (c) =⇒ (b) =⇒ (a) is
obviously true. So we will prove that (a) =⇒ (b) and (b) =⇒ (c).
Suppose that (a) is true. Let a ∈ X and ε > 0. Since L is continuous at 0 and
L(0) = 0, there exists δ > 0 such that
∀u ∈ X, kukX < δ =⇒ kL(u)kY < ε .
In particular, for u = x − a and x ∈ X, we have
kx − akX < δ =⇒ kL(x − a)kY < ε =⇒ kL(x) − L(a)kY < ε .
Here we used the linearity of L to write L(x − a) = L(x) − L(a). This proves that (b)
is true and consequently (a)⇐⇒(b). The implication (b) =⇒ (c) follows similarly as
(a) =⇒ (b).
Now it remains to show that (c)⇐⇒(d). Indeed, suppose that (c) is true. We will
prove (d) by contradiction. Suppose that (d) is false. Consequently, the negation of (d)
is true, i.e.
∀ M > 0 , ∃ a(M ) ∈ X , kL(a(M ))kY > M ka(M )kX . (2.1)
For all n ∈ N, define xn = a(n). Clearly, xn 6= 0 since, if xn = 0, then by (2.1),
1

0 = kL(xn )kY > M kxn kX = 0, i.e. 0 > 0. Consider the sequence, un = xn .
nkxn kX
1
It is clear that, for all n ∈ N, kun kX = n → 0, i.e. un → 0 in (X, k · kX ). By the
continuity of L, L(un ) → L(0) = 0 in (Y, k · kY ). By (2.1), for all n ∈ N, kL(un )kY > 1.
By taking n → ∞ on both sides of this inequality, we get 0 = lim kL(un )kY ≥ 1. This is
n→∞
a contradiction.
Now, suppose that (d) is true. By the linearity of L, for all x ∈ X and y ∈ X, (d)
asserts that kL(x) − L(y)kY = kL(x − y)kY ≤ M kx − ykX , i.e. L is M -Lipschitz. This
proves that L is uniformly continuous. Thus (a) is true.

Remark 2.4. The above theorem gives a characterization of linear continuous func-
tions. In addition to the algebraic property of linearity, continuity needs an extra condi-
tion, namely, L : X → Y is continuous if kL(·)kY and k · kX are comparable i.e.
L : (X, k · kX ) 7−→ (Y, k · kY ) is continuous ⇐⇒ ∃ M > 0, ∀ x ∈ X, kL(x)kY ≤ M kxkX
Theorem 2.3 validates the following definition:
Definition 2.5. Let (X, k · kX ) and (Y, k · kY ) be normed spaces.
(1) L : X → Y be a linear function.
30 2. BOUNDED LINEAR TRANSFORMATIONS

(a) If L is continuous (from (X, k · kX ) to (Y, k · kY )), then we define the norm
of L to be
kL(x)kY
kLk = sup .
x∈X kxkX
x6=0

(b) We say that L is a bounded linear transformation if L is a continuous

linear function (from (X, k · kX ) to (Y, k · kY )).
(2) The space of continuous linear function
from (X, k · kX ) to (Y, k · kY ) is denoted
by L(X, Y ) (or L (X, k · kX ), (Y, k · kY ) or B(X, Y )).
In the special case (X, k · kX ) = (Y, k · kY ), we simply write L(X) instead of
L(X, Y ).

Remark 2.6.
(1) Theorem 2.3 yields that L(X, Y ) is a vector space.
(2) The norm of a linear continuous function L from (X, k · kX ) to (Y, k · kY ) is also
denoted by
kLkL(X,Y ) or kLk
.
L (X,k·kX ),(Y,k·kY )

This notation is helpful to remind the reader of the spaces and the norms.
(3) The notion of a bounded linear transformation/operator is not to be con-
fused with bounded functions. In the former, bounded-ness signifies that the norm
of the linear function is finite (hence meaning continuity).
(4) For bounded linear transformations, Theorems 2.3 states the following
continuity at a point ⇔ continuity everywhere ⇔ uniform continuity ⇔ Lipschitz function .
(5) One can extract useful information about the norm of a continuous linear function
L : X → Y as follows (these simply come from the definition of the sup and kLk).
Let L : X → Y be a linear function. It holds:
(a) If
∃ M > 0 , ∀ x ∈ X, kL(x)kY ≤ M kxkX ,
then we deduce that L is continuous and kLk ≤ M . (This remark is helpful
to extract an upper bound on kLk).
(b) For all x ∈ X \ {0}, we have
kL(x)kY
kLk ≥ .
kxkX
If x is selected properly, then this remark is helpful to extract a lower bound
on kLk.
(c) Let ε > 0 (no matter how small it is). There exists xε ∈ X such that,
kL(xε )kY
xε 6= 0 and ≥ kLk − ε .
kxε k
(d) The inequality kL(x)kY ≤ kLk kxkX is true for all x ∈ X.
(6) The norm of a linear continuous function can be expressed in the equivalent forms
kLk = sup kL(x)kY = sup kL(x)kY .
x∈X x∈X
kxkX ≤1 kxkX =1
 3. THE SPACE OF BOUNDED LINEAR TRANSFORMATIONS 31

(7) [Composition of continuous linear functions] If L and T are two continuous

linear functions such that the composition L ◦ T is defined, then L ◦ T is a
continuous linear function whose operator norm satisfies kL ◦ T k ≤ kLk kT k.
Actually, the composition of two continuous functions is a continuous func-
tion. Also, in linear algebra, the composition of two linear functions is a linear
function. The statement about the norm of L ◦ T follows from statement (5)
above.

3. The space of bounded linear transformations

In this section we study spaces of bounded linear transformations L(X, Y ). Theo-
rem 3.1 below says that the notion of the norm of a linear bounded transformation is
indeed a norm in the (vector) space L(X, Y ).
Theorem 3.1. Let (X, k · kX ) and (Y, k · kY ) be normed spaces, and L(X, Y ) be the
space of all continuous linear functions from (X, k · kX ) to (Y, k · kY ). The function (see
Definition 2.5)
k · k : L(X, Y ) → [0, ∞[
kL(x)kY
L 7→ kLkL(X,Y ) = sup ,
x∈X kxkX
x6=0

is a norm in L(X, Y ). (For instance, we get a notion of convergence in L(X, Y ), as well

as many topological notions).
Proof. We will check that k · k satisfies the properties (N1)-(N3) in Definition 2.1.
Proof of (N1):
Let T ∈ L(X, Y ) and kT k = 0. By Remark 2.6, for all x ∈ X, 0 ≤ kT (x)kY ≤
kT k kxkX = 0, i.e. T (x) = 0. This proves that T = 0.
Proof of (N2):
Let T ∈ L(X, Y ) and λ ∈ C. By (N2) (for k · kY ) and Remark 2.6, for all x ∈ X,
kλT (x)kY = |λ| kT (x)kY ≤ |λ| kT k ||xkX . This proves that kλT k ≤ |λ| kT k.
Now we prove that kλT k ≥ |λ| kT k. Let ε > 0. By Remark 2.6, there exists xε ∈
X such that xε 6= 0 and kT (xε )kY ≥ (kT k − ε)kxε kX . Consequently, kλT (xε )kY =
|λ| kT (xε )kY ≥ |λ|(kT k − ε)kxε kX and
kλT (xε )kY
≥ |λ|(kT k − ε) . (3.1)
kxε k
But, by Remark 2.6,
kλT (xε )kY
≤ kλT k . (3.2)
kxε k
Combining (3.1) and (3.2), we get,
∀ ε > 0, |λ|(kT k − ε) ≤ kλT k .
By taking ε → 0+ , we arrive to |λ|kT k ≤ kλT k.
Proof of (N3):
Let T ∈ L(X, Y ) and S ∈ L(X, Y ). By the triangle inequality and Remark 2.6, for
all x ∈ X,
k(S + T )(x)kY = kS(x) + T (x)kY ≤ kS(x)kY + kT (x)kY


≤ kSk kxkX + kT k kxkX = kSk + kT k kxkX .
32 2. BOUNDED LINEAR TRANSFORMATIONS

By Remark 2.6, this proves that kS + T k ≤ kSk + kT k. 

Next we study completeness of spaces of continuous linear functions. This is the
content of:
Theorem 3.2. Let (X, k · kX ), (Y, k · kY ) be normed spaces and L(X, Y ) be the normed
space of linear continuous functions from (X, k · kX ) to (Y, k · kY ).
If (Y, k · kY ) is a Banach space, then (L(X, Y ), k · kL(X,Y ) ) is a Banach space too.
Proof of Theorem 3.2. Let us recall the definition of the norm in L(X, Y ). If
T ∈ L(X, Y ), then the norm of T is
kT (x)kY
kT k = kT kL(X,Y ) ≤ sup .
x∈X kxkX
x6=0

Let (Tn ) be a Cauchy sequence in L(X, Y ). To prove Theorem 3.2, we must find a
continuous linear function T : X → Y such that Tn → T in L(X, Y ). We will do this in
three steps.
Step 1. Definition of the function T :
Let ε > 0. Since (Tn ) is a Cauchy sequence in L(X, Y ), then there exists n0 ∈ N such
that,
ε
∀ n ≥ n0 , ∀ p ∈ N , kTn+p − Tn k < . (3.3)
2
Let x ∈ X. As a consequence of (3.3) and the definition of the norm in L(X, Y ), we have,
ε
∀ n ≥ n0 , ∀ p ∈ N , kTn+p (x) − Tn (x)kY < kf kX . (3.4)
2
By Remark ??, the sequence (T (xn )) is a Cauchy sequence in (Y, k · kY ). Since (Y, k · kY )
is a Banach space, we deduce that (T (xn )) is convergent in (Y, k · kY ). Let us denote by
T (x) the limit of T (xn ) in (Y, k · kY ). That way, to every x ∈ X, we assigned T (x) ∈ Y ;
this defines a function T : X → Y .
Step 2. The function T is in the space L(X, Y ):
Here we prove that T ∈ L(X, Y ). Let x ∈ X, y ∈ X and k ∈ R or C. We will prove
that T (x + y) = T (x) + T (y) and T (kx) = kT (x). Since Tn is a linear function, then
Tn (x + y) = Tn (x) + Tn (y) and Tn (kx) = kTn (y) . (3.5)
Now, by Step 1,
Tn (x + y) → T (x + y) , Tn (x) → T (x) , Tn (y) → T (y) in (Y, k · k) .
Furthermore, we get by Theorem 2.10
Tn (x) + Tn (y) → T (x) + T (y) and kTn (x) → kT (x) in (Y, k · k) .
We take n → ∞ in (3.5) and use that the limit is unique to write
T (x + y) = T (x) + T (y) and T (kx) = kT (y) .
Next we prove that T is continuous. By (3.4) used with p ∈ N and n = n0 , we have,
ε
∀ x ∈ X , ∀ p ∈ N , kTn0 +p (x) − Tn0 (x)kY < kxkX .
2
By sending p → ∞ and Step 1, we get,
ε
∀ x ∈ X , kT (x) − Tn0 (x)kY ≤ kxkX .
2
 3. THE SPACE OF BOUNDED LINEAR TRANSFORMATIONS 33

By the triangle inequality and the fact that Tn0 ∈ L(X, Y ), we deduce further,
ε
∀ x ∈ X , kT (x)kY ≤ kTn0 (x)kY + kT (x) − Tn0 (x)kY ≤ (kTn0 + )kxkX .
2
By Theorem 2.3, we deduce that T is continuous.
Step 3. The convergence Tn → T holds in L(X, Y ):
By (3.4) and Step 2, we have,
ε
∀ x ∈ X , ∀ n ≥ n0 , kT (x) − Tn (x)k = lim kTn+p (x) − Tn (x)kY ≤ kxkX .
p→∞ 2
By definition of the norm in L(X, Y ) (cf. Remark 2.6, Item (5)-(a)), we get,
ε
∀ n ≥ n0 , kT − Tn k ≤ < ε .
2
This proves that lim kT − Tn k = 0. 
n→∞
34 2. BOUNDED LINEAR TRANSFORMATIONS

Exercises

Exercise 1. Let E = C[0, 1] be equipped with the norm of uniform convergence k · k∞ .

For every k > 0, we denote by Hk the subspace of E constituting of k-Lipschitz functions
i.e. n o
Hk = f ∈ E : |f (x) − f (y)| ≤ k|x − y|, ∀ x, y ∈ [0, 1] .
We also denote by H the subspace of E constituting of all Lipschitz functions i.e.
n o
H = f ∈ E : ∃ K > 0 s.t. |f (x) − f (y)| ≤ k|x − y|, ∀ x, y ∈ [0, 1] .

(1) Show that for every k > 0, Hk is closed in (E, k · k∞ ). What can we say about
Hk ?
(2) We define the sequence of functions (fn )n≥1 by
r
1
fn (x) = x +
n
(a) Show that the sequence (fn ) converges uniformly on [0, 1] to a limit f to be
determined. √
n
(b) Show that for all n ∈ N, fn is -Lipschitz.
2
(c) Show that f 6∈ H.
(3) Is H closed in (E, k · k∞ )? Is (H, k · k∞ ) a Banach space? Justify your answers.

Exercise 2. We denote by H the subspace of E = C[0, 1] constituting of Lipschitz

functions i.e.
n o
H = f ∈ E : ∃K > 0 s.t. |f (x) − f (y)| ≤ k|x − y|, ∀ x, y ∈ [0, 1] .

For every f ∈ H, we denote

|f (x) − f (y)|
kf k = kf k∞ + sup .
x6=y x,y∈[0,1] |x − y|

(1) Show that k · k is a norm on H.

(2) Show that (H, k · k) is a Banach space.

Exercise 3. Let (X, N1 ) and (Y, N2 ) be normed spaces and f a mapping from X into Y .

(1) Show that if f is continuous then G = {(x, f (x)) : x ∈ X} is closed in X × Y . G

is called the graph of f .
(2) Show that the converse is not  necessary true. Hint: take X = Y = R and
1
x
if x 6
= 0
consider the mapping f (x) =
0 if x = 0.
(3) Let A a compact of Y . Show that if G is closed then f : A 7−→ Y is continuous.
 EXERCISES 35

Exercise 4. Let K be a compact subset of Rn and x0 ∈ Rn . Show that the set

Kx0 = {x0 + x : x ∈ K} is compact.

Exercise 5. Let (R, | · |) endowed with the usual norm.

(1) Consider f and g two real valued continuous functions on [0, 1]. We define the
function h : R 7−→ R by
h(x) = sup |f (t) + xg(t)|.
t∈[0,1]

(a) Show that h is Lipshitz.

(b) Deduce that h is uniformly continuous on R. √
(2) Show that the function f : [0, +∞[7−→ R defined by f (x) = x is uniformly
continuous but it is not Lipshitz.

Exercise 6. Let E = C[0, 1] be equipped with the norm of uniform convergence k · k∞ .

For every f ∈ E, define
Z 1
ψ(f ) = (f (t))2 dt.
0
Let A ⊆ E be the set defined by A = {f ∈ E, kf k∞ ≤ 1}.
(1) Show that A is closed in (E, k · k∞ ).
Show that ψ : (A, k · k∞ ) 7−→ R is Lipschitz. 
(2) 
Hint: we can write f 2 − g 2 = (f − g)(f + g)
(3) Let f (x) = x and gn (x) = nx, n ∈ N.
(a) Calculate ψ(f ), ψ(gn ) and kf − gn k∞ .
(b) Deduce that ψ : (E, k · k∞ ) 7−→ R is not Lipschitz.

Exercise 7. Let (X, k.k) be a normed space. Show that the mapping f : X 7−→ B(0, 1)
x
defined by f (x) = 1+kxk is a homeomorphism.

1
Exercise 8. Let E = Q ∩ [0, 1] and f : E 7−→ E the mapping defined by f (x) = .
x+2
Show that f is a contraction and that f does not have a fixed point in X. What can you
say about X?

Exercise 9. Consider the space E = C[0, 1] equipped with the norm of uniform conver-
gence k · k∞ .
We define the application T : (E, k · k∞ ) → (E, k · k∞ ) by
 Z 1 
1 xt
T (f )(x) = 1+ xe f (t)dt .
2 0

(1) Show that T is a contraction.

36 2. BOUNDED LINEAR TRANSFORMATIONS

(2) Deduce that the integral equation

Z 1
xext f (t)dt = 2f (x) − 1, ∀ x ∈ [0, 1]
0
unique solution f0 ∈ E.
admits a 
1
(3) Let B = f ∈ E; kf k∞ ≤ .
3−e
(a) Show that B is closed and complete in (E, k · k∞ ).
(b) Show that
1 e−1
kT (f )k∞ ≤ + kf k∞ .
2 2
Deduce that the application T : B → B is well defined.
1
(c) Deduce that kf0 k∞ ≤ .
3−e

Exercise 10.
(1) Let α ≥ 0 and consider the space C[0, 1] equipped with the two norms:
kf k∞ = sup |f (t)| and kf kα = ke−αt f (t)k∞ , α ≥ 0.
t∈[0,1]

(a) Show that the norms k · k∞ and k · kα are equivalent over C[0, 1].
(b) Deduce that (C[0, 1], k · kα ) is a Banach space.
(2) Let f : [0, 1] → R be a α-Lipschitz function; that is, for all y, z ∈ R:
|f (y) − f (z)| ≤ α|y − z|.
The aim of this problem is to show that the following differential equation
( 0
y (t) = f (y(t)),
(∗)
y(0) = y0 ∈ R,
has a unique solution y ∈ C 1 [0, 1] = {f ∈ C[0, 1], f 0 ∈ C[0, 1]}.
(a) Show that if y ∈ C[0, 1] is a solution of (∗) then y ∈ C 1 [0, 1].
(b) Show that y ∈ C 1 [0, 1] is a solution of (∗) Zif and only if y ∈ C[0, 1] is a
t
solution of the integral equation y(t) = y0 + f (y(s))ds.
0
(c) Let T : (C[0, 1], k · kα ) 7−→ (C[0, 1], k · kα ) be the mapping defined by
Z t
T (y)(t) = y0 + f (y(s))ds, (t ∈ [0, 1]).
0
Show that T is a contraction.
(d) Deduce that there exists a unique solution y ∈ C 1 [0, 1] of (∗).

Exercise 11. Let E = {f ∈ C(R) : (1 + x2 )|f (x)| is bounded}. Consider the mappings
N : E 7−→ R+ and L : E 7−→ R defined by:
Z
2
N (f ) = sup(1 + x )|f (x)|, L(f ) = f (x)dx
x∈R R

(1) Show that N is a norm on E.

 EXERCISES 37

(2) Show that the linear form L is continuous.

(3) Calculatethe norme
Z of L. 
(4) Let A = f ∈ E : f (x)dx = 1 . Show that, A is closed in (E, N ).
R

Exercise 12. Let X = C[0, 2]. Using linear form show that the following spaces are
closed in (X, k · k∞ ):
 Z 2 
E = {f ∈ X : f (0) = 0}, and F = f ∈X: f (t) dt = 0 .
0

Exercise 13. Let E = C 1 [0, 1] and F = C[0, 1] endowed with the norm
Z 1
kf k1 = |f (x)|dx.
0
Consider the mappings I : E 7−→ E and D : E 7−→ F defined by
Z t
I(f )(t) = f (x)dx, (t ∈ [0, 1]); D(f ) = f 0 .
0
(1) Show that I is linear and continuous for the norm k · k1 .

(2) Show that D is linear but not continuous for the norm k · k1 .

(Hint: consider the sequence of functions fn (t) = (n + 1)tn , (t ∈ [0, 1]) .
(3) We define on E the mapping N by
N (f ) = kf k1 + kf 0 k1 .
Show that N is a norm on E and if E is endowed with this norm then D is
continuous.
(4) Let (gn ) the sequence of functions defined by:
∀t ∈ [0, 1], gn (t) = (n + 1)(1 − t)n
(a) Calculate kgn k1 et kI(gn )k1 .
(b) Deduce that kIk = 1.
(5) Let (hn ) be the sequence of functions defined by:
n n−1
∀t ∈ [0, 1], hn (t) = t
n+1
(a) Calculate N (hn ) et kD(hn )k1 .
(b) Deduce that kDk = 1.

Exercise 14. For f ∈ C[0, 1], we define the norms kf k∞ = sup |f (t)| and kf k1 =
t∈[0,1]

Z 1
|f (t)|dt. Are the following linear mappings continuous? Eventually, compute their
0
norms.
(1) φ : (C([0, 1]), k · k∞ ) −→ R, φ(f ) = f (0);
38 2. BOUNDED LINEAR TRANSFORMATIONS

(2) φ : (C([0, 1]), k · k1 ) −→ R, φ(f ) = f (0);

(3) φ : (C([0, 1]), k · k1 ) −→ (C([0, 1]), k · k∞ ), φ(f ) = f ; Z
x
(4) φ : (C([0, 1]), k · k1 ) −→ (C([0, 1]), k · k∞ ), φ(f )(x) = f (x)dx.
0

Exercise 15. Consider the vector space `∞ defined by

 
∞
` = x = (xk ) : xk ∈ R, sup |xk | < +∞ ,
k∈N

equipped with the norm kxk∞ = sup |xk |. Consider the mapping f : `∞ 7−→ R defined by
k∈N

f (x) = x0 , ∀ x = (xk ) ∈ `∞ .
Show that f is a linear form and continuous. Calculate kf k.

Exercise 16. Consider E = C[−1, 1] endowed with the norm kf k∞ = sup |f (t)|. Let
t∈[−1,1]
F the space of 2π-periodic continuous functions from R into R, endowed with either the
Z π 1/2
norm N2 such that N2 (f ) = |f (t)|2 dt , or the norm N∞ : N∞ (f ) = sup |f (t)|.
−π t∈R
Let L : E 7−→ F be the mapping defined by
(L(f ))(t) = f (cos(t)).
(1) Show that L is well defined, linear and injective.
(2) Show that L is continuous for each of the norms N2 and N∞ of F , and calculate
their corresponding norms kLk2 and kLk∞ .

Exercise 17. Let E and F be two n.v.s. and let T ∈ L(E, F ) such that kLk ∈ [0, 1[.
Show that T is a contraction.

Exercise 18. Let X = C([0, 1], R) endowed with the norm kf k∞ = sup |f (x)| and let
x∈[0,1]
I : X 7−→ X be the mapping defined, for any f ∈ X, by
Z x
I(f )(x) = sin(x − t)f (t)dt, ∀ x ∈ [0, 1].
0
(1) Show that is well defined.
(2) Show that L is linear and continuous. Calculate kIk.
(3) Deduce that I has a unique fixed point.
(4) Determine the fixed point of I.

Exercise 19.
A. Let E be a Banach space and f : E 7−→ E a continuous mapping. Show that if
f 2 = f ◦ f is a contraction then f has a unique fixed point.
 EXERCISES 39

B. Let f : R −→ R a differentiable function. Show that f is a contraction if and

only if there exists 0 < k < 1 such that |f 0 (x)| ≤ k, ∀x ∈ R.
C. Application: Let E = R.
(a) Show that the function defined by f (x) = cos x is not a contraction.
(b) Show that the function defined by g(x) = cos(cos(x)) is a contraction.
(c) Deduce that f has a unique fixed point.

Exercise 20. Let (E, k · k) be a normed space and C a convex compact of E. Consider
f : C 7−→ C an application verifying
kf (x) − f (y)k ≤ kx − yk, ∀ x, y ∈ C.
∗
Let x0 ∈ C. For every n ∈ N , consider the application fn : C 7−→ C defined by
f (x0 )  1
fn (x) = + 1− f (x).
n n
(1) Show that fn is a contraction.
(2) Show that there exists a unique xn ∈ C such that
fn (xn ) = xn .
(3) Show that kfn (xn ) − f (x)k ≤ n1 kx − x0 k + 1 − n1 kxn − xk.
(4) Deduce that f has a fixed point.
 CHAPTER 3

HILBERT SPACES

In this chapter, we will discuss spaces having a richer algebraic/geometric structure

than the normed spaces discussed before. We can take the Euclidean space R2 as a
starting point and make the following observations:
(1) R2 is a vector space ;
(2) If u = (u1 , u2 ) ∈ R2 , v = (v1 , v2 ) ∈ R2 and k ∈ R, the following operations are
defined:
• Addition: u + v ;
• Multiplication by a scalar: ku
• Inner product: hu, vi=u · v = u1 v1 + u2 v2 p
(3) R2 is a normed space with the Euclidean norm kuk = u21 + u22 .
In the earlier chapters discussed so far, the roles of the vector space and norm structures
have been discussed thoroughly. Here we will discuss the operation of the inner product.
We notice that:
The Euclidean norm in R2 can be expressed using the inner product: kuk =
• √
u·v;
• The inner product allows us to define the notion of orthogonality in R2 , namely,
u ∈ R2 and v ∈ R2 are orthogonal if u · v = 0.
We will generalize these properties to more general spaces, namely the Hilbert spaces.
The Euclidean space R2 (and more generally, Rd with d ≥ 1) is an example of a Hilbert
space. In other terms, a Hilbert space accommodates three different notions: vector spaces
(algebra), inner products (geometry/orthogonality) and norms (analysis). Our discussion
of Hilbert spaces will highlight the interplay between these three notions.

1. Inner product spaces

In this section, H denotes a vector space over the field K, where K is either the field
of real number R or the field of complex numbers C. The complex conjugate of a complex
number z will be denoted by z.
Definition 1.1. [Inner product]
An inner product in the space H is a function h·, ·i : H × H → K satisfying the
following properties:
(IP1) (positive definite)
• For all f ∈ H, hf, f i ≥ 0 ; and
• hf, f i = 0 =⇒ f = 0.
(IP2) For all f, g ∈ H, hf, gi = hg, f i.
(IP3) (semi-linearity) For all f, g, h ∈ H and λ, µ ∈ K,
hλf + µg, hi = λhf, hi + µhg, hi .
41
42 3. HILBERT SPACES

The space H endowed with an inner product h·, ·i is called an inner product space
(or a pre-Hilbert space). This is usually denoted by the pair (H, h·, ·i). If furthermore
K = R, then H is called a real inner product space.

Remark 1.2. Assume that (H, h·, ·i) is an inner product space. Then the following
consequences are satisfied:
(1) For all λ, µ ∈ K and h, g, h ∈ H, we have
hh, λf + µgi = λhh, f i + µhh, gi .
(2)
hf, f i = 0 ⇔ f = 0 .
(3) Polarization identity: For all f, g ∈ H, we have
1  i 
hf, gi = kf + gk2 − kf − gk2 + kf + igk2 − kf − igk2 .
4 4
If H is a real inner product space, the polarization identity becomes
1 
hf, gi = kf + gk2 − kf − gk2 .
4
(4) Parallelogram identity: For all f, g ∈ H, we have
1 
kf k2 + kgk2 = kf + gk2 + kf − gk2 .
2
Next we point two typical examples of inner product spaces.

Example 1.3.
(1) [Complex Euclidean Spaces] Let d ≥ 1. The Euclidean space Cd is an inner
product space with the usual inner product defined as follows. For all u =
(u1 , · · · , ud ) ∈ Cd and v = (v1 , · · · , vd ) ∈ Cd ,
d
X
hu, vi = u · v = ui vi . (1.1)
i=1

(2) [Euclidean Spaces] Let d ≥ 1. The Euclidean space Rd is a real inner product
space with the usual inner product defined as follows. For all u = (u1 , · · · , ud ) ∈
Rd and v = (v1 , · · · , vd ) ∈ Rd ,
d
X
hu, vi = u · v = ui vi . (1.2)
i=1

(3) [Space of continuous functions] The space C([0, 1]) of continuous real-valued
functions on the interval [0, 1] is a real inner product space with the inner product
defined as follows
Z 1  
hf, gi = f (x)g(x) dx f, g ∈ C([0, 1]) . (1.3)
0

Theorem 1.4. [Inner product spaces are normed spaces] p

Let (H, h·, ·i) be an inner product space. For all f ∈ H, we define kf k = hf, f i. The
following two statements hold.
(1) [Cauchy-Schwarz Inequality] For all f, g ∈ H, |hf, gi| ≤ kf k kgk .
 1. INNER PRODUCT SPACES 43

(2) The function f 7→ kf k is a norm in H.

Proof.
Proof of (1): If hf, gi = 0 then the inequality is true. Assume that hf, gi 6= 0 and
take λ ∈ C, we may write
 
0 ≤ hf + λg, f + λgi = kf k2 + 2 Re λhf, gi + kgk2 |λ|2 .
|hf, gi|
We choose λ = t with t ∈ R an arbitrary real number. We get
hf, gi
0 ≤ hf + λg, f + λgi = kf k2 + 2|hf, gi|t + kgk2 t2 .
By completing the square in the quadratic expression in t, we deduce that
|hf, gi|2 − kf k2 kgk2 ≤ 0 .
This yields the inequality in (1).

Proof of (2):
Clearly, kf k ≥ 0 for all f ∈ H. Moreover, if kf k = 0, then hf, f i = 0 and by (IP1),
f = 0. p
Using (IP3), we get for all λ ∈ C and f ∈ H, kλf k = |λ|2 hf, f i = |λ| kf k.
Finally, we prove the triangle inequality. Let f, g ∈ H. We have
kf + gk2 = hf + g, f + gi
= kf k2 + kgk2 + 2 Re hf, gi

= kf k2 + kgk2 + 2|hf, gi [ since Re(z) ≤ |z| ]

≤ kf k2 + kgk2 + 2kf k kgk [By Cauchy-Schwartz Inequality]
 2
= kf k + kgk .
This yields that kf + gk ≤ kf k + kgk. 
Remark 1.5. Let g ∈ H be fixed and define the function Lg : H → R by Lg (f ) =
hf, gi. Using (IP3), we find that Lg is a linear function. Using the Cauchy-Schwarz
inequality (Theorem 1.4), we find that Lg is continuous and its operator norm satisfies
kLg k ≤ kgk.
Definition 1.6. [Hilbert spaces]
Let (H, h·, ·i) be an inner product space.
p
(1) The norm k · k = h·, ·i is called the norm induced by the inner product.
(2) If (H, k · k) is a Banach space, then (H, h·, ·i) is said to be a Hilbert space.

Example 1.7.
(1) The Euclidean space Rd with the usual inner product in (1.2) is a Hilbert space.
The complex Euclidean space Cd is a Hilbert space too.
(2) The space of continuous functions C([0, 1]) with the inner product in (1.3) is
not a Hilbert space. In fact, the norm induced by the inner product is kf k2 =
Z 1 1/2
2
|f (x)| dx , and (C([0, 1]), k · k2 ) is not a Banach space as pointed in
0
Theorem 7.5 (Chapter 1).
44 3. HILBERT SPACES

(3) The space L2 ([0, 1]) introduced in Remark 7.6 (Chapter 1) is a Hilbert space with
the inner product defined by the formula in (1.3). Note that L2 ([0, 1]) is loosely
defined in the following manner:
• L2 ([0, 1]) is a space of functions larger than
R 1 C([0,2 1]) ;
2
• If f ∈ L ([0, 1]), then we may compute 0 |f (x)| dx ;
• (L2 ([0, 1]), k · k2 ) is a Banach space ;
• C([0, 1]) is dense in L2 ([0, 1]) .
Theorem 1.8. [Continuity of the inner product]
Assume that (H, h·, ·i) is an inner product space. Let k · k be the norm in H induced
by the inner product.
Suppose that:
• (fn ) and (gn ) are two sequences in H ;
• fn → f and gn → g in (H, k · k) .
Then the sequence (hfn , gn i) ⊂ C is convergent and
hfn , gn i → hf, gi in (C, usual distance) .
Proof. For all n ≥ 1, we write
hfn , gn i − hf, gi = hfn , gn − gi + hfn − f, gi . (1.4)
We will prove that hfn , gn − gi → 0 and hfn − f, gi → 0 in C. Since fn → f in (H, k · k),
Remark 1.5 yields that
hfn − f, gi = Lg (fn − f ) → Lg (0) = 0 .
For the term hfn , gn − gi, we note that (fn ) is a convergent sequence, hence bounded.
Consequently,
∃ M > 0, ∀ n ≥ 1, kfn k ≤ M .
Next, we use the Cauchy-Schwarz inequality (Theorem 1.4) to write, for all n ≥ 1,
|hfn , gn − gi| ≤ kfn k kgn − gk ≤ M kgn − gk .
Since gn → g in (H, k · k), kgn − gk → 0 and we deduce that hfn , gn − gi → 0 in C.
Returning back to (1.4), we find
lim hfn , gn i − hf, gi = lim hfn , gn − gi + lim hfn − f, gi = 0 .
n→∞ n→∞ n→∞


Remark 1.9.
(1) We record the following special case of Theorem 1.4. If g ∈ H and fn → f in H,
then lim hfn , gi = hf, gi and lim hg, fn i = hg, f i.
n→∞ n→∞
(2) The inner product in H induces a norm k·k, hence a notion of convergence/continuity
in H. This in turn induces a norm k·kprod in the product space H ×H, hence a no-
tion of convergence/continuity in H × H (Theorem 4.1; Chapter 1). Theorem 1.4
states that the inner product
h·, ·i : (H × H, k · kprod ) → (H, k · k)
is a continuous function.
 2. ORTHOGONALITY 45

2. Orthogonality
In this section, (H, h·, ·i) denotes an inner product space.
Definition 2.1. [Orthogonal vectors]
Two vectors f ∈ H and g ∈ H are said to orthogonal in (H, h·, ·i) if hf, gi = 0. We
will express this notion by f ⊥g.
Remark 2.2.
(1) For all f ∈ H, f ⊥0 since hf, 0i = 0.
(2) If f ⊥f , then hf, f i = 0 and f = 0.
(3) [Pythagorean identity] If f ⊥g, then kf + gk2 = kf k2 + kgk2 .
Definition 2.3. [Orthogonal complements]
Let M ⊂ H. The orthogonal complement of M in (H, h·, ·i) is the new subset
M ⊥ = {f ∈ H : ∀ g ∈ M , hf, gi = 0} .
Remark 2.4.
(1) Let f ∈ H. By definition 2.3, f ∈ M ⊥ if and only if f is orthogonal to every
vector in M .
(2) The orthogonal complement of {0} is {0}⊥ = H.
(3) H ⊥ = {0}. Actually, if f ∈ H ⊥ , then f ∈ H and hf, f i = 0, hence f = 0.
(4) For every nonempty subset M of H we have 0 ∈ M ⊥ , hence M ⊥ 6= ∅.
(5) M ⊥ is a subspace of H, since M ⊥ 6= ∅ and
∀ f1 , f2 ∈ M ⊥ , ∀ k ∈ R, kf1 ∈ M ⊥ and f1 + f2 ∈ M ⊥ .
(6) If N ⊂ M ⊂ H, then M ⊥ ⊂ N ⊥ .
(7) M ⊥ ∩ M = {0}.
Theorem 2.5. [The orthogonal complement is closed]
Suppose that M ⊂ H. The orthogonal complement M ⊥ is a closed set.
Proof. If M = ∅ then M ⊥ = ∅ is a closed set. Suppose M 6= ∅ and let N = M ⊥ .
We should prove that N ⊂ N . Let f ∈ N . There exists a sequence (fn ) ⊂ N such that
fn → f . Let g ∈ M . By Remark 1.9 (2), we get
lim hfn , gi = hf, gi .
n→∞

But, for all n ≥ 1, fn ∈ N = M ⊥ , thus hfn , gi = 0, because g ∈ M . Consequently, for all

g ∈ M,
hf, gi = lim hfn , gi = 0 .
n→∞
⊥
This proves that f ∈ M = N . 
Theorem 2.6. [Orthogonal complement of the closure]
Suppose that M ⊂ H is a non-empty set. The orthogonal complement of the closure
of M satisfies
⊥
M = M⊥ .
⊥
Proof. Since M ⊂ M , M ⊂ M ⊥ by Remark 2.4 (6). So it remains to prove the
⊥
inclusion M ⊥ ⊂ M .
⊥
Let f ∈ M ⊥ . We have to prove that f ∈ M . Pick any g ∈ M , we will prove that
hf, gi = 0.
46 3. HILBERT SPACES

Since g ∈ M , there exists a sequence (gn ) ⊂ M such that gn → g in H. For all n ≥ 1,

gn ∈ M , hence hf, gn i = 0, because f ∈ M ⊥ .
Now, by Theorem 1.8 (or the special case in Remark 1.9 (1)), we get
hf, gi = lim hf, gn i = 0 .
n→∞


3. The Projection theorem

In this section, we assume that Hp is a Hilbert space, i.e. there is an inner product
h·, ·i in H such that the norm k · k = h·, ·i induced by the inner product turns H to a
Banach space. Thus, Cauchy sequences in (H, k · k) are convergent.
We will prove that, if M is a closed subspace of H, then we can decompose H into the
direct sum of the spaces M and M ⊥ . That is1, we will prove that every element f ∈ H
can be decomposed in a unique manner in the form u + v, where u ∈ M and v ∈ M ⊥ .
In the particular case where H = R2 and M is the x-axis say, we can decompose
f ∈ H in the form u + v, where u is the orthogonal projection on M and v = f − u
is orthogonal to M . We will prove that such a geometric construction holds in general
Hilbert spaces. A necessary ingredient in the construction is the property that
Cauchy sequences are convergent.
Theorem 3.1. [The projection theorem]
Suppose that M is a closed subspace of H. Then there exists a function
PM : H → H
satisfying the following properties:
(1) PM is a linear function ;
(2) For all f ∈ M , PM (f ) = f ;
(3) For all f ∈ H, PM (f ) ∈ M and f − PM (f ) ∈ M ⊥ ;
(4) For all f ∈ H, kf − PM (f )k = inf{kf − gk : g ∈ M } ;
(5) If M 6= {0}, then kPM kL(H) = 1 .
Remark 3.2.
(1) The function PM constructed in Theorem 3.1 is called the orthogonal projec-
tion on M .
(2) [Direct sum decomposition of H] When the assumptions in Theorem 3.1
hold, for all f ∈ H, f = u + v with u = PM (f ) ∈ M and v = f − PM (f ) ∈ M ⊥ .
Thus H = M + M ⊥ . Since M ⊥ ∩ M = {0}, we get that H = M ⊕ M ⊥ .
The direct sum decomposition of H may be of practical interest. If f ∈ H is
written in the form f = u + v with u ∈ M and v ∈ M ⊥ , then this decomposition
is unique. Consequently, PM (f ) = u and f − PM (f ) = v.
(3) If M is a subspace of H but not a closed set, then we can not apply Theorem 3.1
on M . But we can still write a useful decomposition of H in the form H =
M ⊕ M ⊥ by noting that:
• M is a subspace of H and is a closed set (see Remark 2.4 and Theorem 2.5) ;
⊥
• Theorem 3.1 now yields H = M ⊕ M ;
⊥
• Theorem 2.6 yields M = M ⊥ .
1We
say that H is the direct sum of the two subspaces E and F and we write H = E ⊕ F if:
H = E + F and E ∩ F = {0}.
 3. THE PROJECTION THEOREM 47

(4) [The orthogonal projection on a uni-dimensional space] Suppose that

u ∈ H, u 6= 0 and M = Span(u) is the vector space generated by u. The
hf, ui
orthogonal projection on M is given by PM (f ) = u. This is true in light
kuk2
of the statement (3) above since PM (f ) ∈ M and f − PM (f ) ∈ M ⊥ (actually, for
all g = αu ∈ M , hf − PM (f ), gi = 0).
The proof of Theorem 3.1 is lengthy and contains many tricks. We will split it into
several lemmas for the sake of clarity.
Lemma 3.3. [Construction of the projection]
Suppose that M 6= ∅ is a convex subset of H (see Definition 3.3 - Chapter 1). If
furthermore M is closed, then for every f ∈ M , there exists a unique element PM (f ) ∈ M
such that
kf − PM (f )k = dist(f, M ) .
Remark 3.4.
(1) Recall that dist(f, M ) = inf{kf − gk : g ∈ M }.
(2) The uniqueness of PM (f ) yields that, if g ∈ M and kf − gk = dist(f, M ), then
g = PM (f ).
(3) The uniqueness of PM (f ) ensures that the binary relation f 7→ PM (f ) is a func-
tion.
(4) If f ∈ M , then dist(f, M ) = 0 = kf − f k. The result in Lemma 3.3 then yields
that PM (f ) = f .
(5) In particular, since PM (f ) ∈ M , we get that PM ◦ PM (f ) = PM (f ).
(6) Every subspace is a convex set. Consequently, the result in Lemma 3.3 holds
when M is a closed subspace of H (as in the assumption of Theorem 3.1).

Proof of Lemma 3.3.

Let f ∈ H be fixed. Define λ = inf{kf −gk : g ∈ M }. Since M 6= ∅ then 0 ≤ λ < +∞
and hence, for each n ∈ N there exists gn ∈ M such that
1
kf − gn k2 < λ2 + . (3.1)
n
We will prove that the sequence (gn ) is a Cauchy sequence. By applying the parallelogram
identity to the vectors f − gn and f − gm we get, for any n, m ∈ N,
kf − gn − (f − gm )k2 + kf − gn + (f − gm )k2 = 2kf − gn k2 + 2kf − gm k2
 
2 1 1
< 4λ + 2 + .
n m
We can rearrange this inequality to get
  2
2 2 1 1 gn + gm
kgn − gm k < 4λ + 2 + −4 f − .
n m 2
gn +gm
Since M is convex and gn , gm ∈ M , we get 2
∈ M and therefore
2
gn + gm
f− ≥ λ2 .
2
48 3. HILBERT SPACES

Hence,
   
2 2 1 1 2 1 1
∀ m, n ≥ 1, kgn − gm )k < 4λ + 2 + − 4λ = 2 + .
n m n m

It follows that (gn ) is a Cauchy sequence, and so the completeness of H yields that (gn )
is convergent. Thus there exists g ∈ H such that gn → g in H. Since M is closed and
(gn ) ⊂ M , we get that g ∈ M .
We set PM (f ) = g. As for the uniqueness of PM (f ), if h ∈ M satisfies kf − hk = λ,
we get by the parallelogram identity and (3.1)

1 1
2λ2 ≥ kf − PM (f )k2 + kf − hk2 = k2f − PM (f ) − hk2 + kPM (f ) − hk2
2 2
1 2 1
= 2 f − (PM (f ) + h) + kPM (f ) − hk2
2 2
2 1 2 1
≥ 2λ + kPM (f ) − hk since (PM (f ) + h) ∈ M .
2 2
This proves that kPM (f ) − hk = 0, i.e. PM (f ) = h. 

Remark 3.5. [A useful inequality]

Assume that M 6= ∅ is a convex and closed subset of H. Then
 
∀ f ∈ H, ∀ g ∈ M , Re hf − PM (f ), g − PM (f )i ≤ 0 ,

where PM is the projection constructed in Lemma 3.3.

Proof. Let f ∈ H and g ∈ M . For all t ∈ ]0, 1], ut := (1 − t)PM (f ) + tg ∈ M by the

convexity of M . Consequently, kf − ut k ≥ dist(f, M ) = kf − PM (f )k. Therefore,

kf − Pm (f )k2 ≤ kf − ut k2

2
2
= f − (1 − t)PM (f ) + tg = (f − PM (f )) + t PM (f ) − g
h i
= kf − PM (f )k2 + t2 kPM (f ) − gk2 + 2t Re hf − PM (f ), PM (f ) − gi .
h i
Thus, for all t ∈ ]0, 1], 2Re hf − PM (f ), g − PM (f )i ≤ tkPM (f ) − gk2 . Sending t to 0+
h i
in this inequality, we get that Re hf − PM (f ), g − PM (f )i ≤ 0. 

In the sequel, we assume that M is a closed subspace of H (as in Theo-

rem 3.1). Also, we denote by f 7→ PM (f ) the function defined in Lemma 3.3.

Lemma 3.6. [The projection is orthogonal] Let M be a closed subspace of H.

Then, for all f ∈ H, we have f − PM (f ) ∈ M ⊥ .

Proof. Let f ∈ H and g ∈ M . We will prove that hf − PM (f ), gi = 0. The proof

involves a tricky application of the inner product identities.
 3. THE PROJECTION THEOREM 49

Let λ ∈ C. Since M is a subspace and f, g ∈ M , PM (f ) + λg ∈ M and

kf − PM (f ) − λgk ≥ dist(f, M ). We have
 2
dist(f, M ) ≤ kf − PM (f ) − λgk2
h i
= kf − PM (f )k2 + |λ|2 kgk2 − 2 Re λhf − PM (f ), gi
 2 h i
= dist(f, M ) + |λ|2 kgk2 − 2 Re λhf − PM (f ), gi .
 
This yields the inequality |λ|2 kgk2 − 2Re λhf − PM (f ), gi ≥ 0, for all λ ∈ C. In
particular, we choose λ = t−1 z where t > 0 is arbitrary and the complex number z is
defined as follows

1 if hf − PM (f ), gi = 0
z = |hf − PM (f ), gi|
 if hf − PM (f ), gi =
6 0
hf − PM (f ), gi
so that z satisfies zhf − PM (f ), gi = |hf − PM (f ), gi|. Hence, for all t > 0, we have
1
|hf − PM (f ), gi| ≤ tkgk2 .
2
Letting t → 0+ , we get that hf − PM (f ), gi = 0. 
Lemma 3.7. [The projection is self-adjoint] For all f, g ∈ M ,
hPM (f ), gi = hf, PM (g)i.
Proof. We write
hPM (f ), gi = hPM (f ), g − PM (g)i + hPM (f ), PM (g)i
= hPM (f ), PM (g)i ,
because PM (f ) ∈ M (by Lemma 3.3) and g − PM (g) ∈ M ⊥ (by Lemma 3.6). In a similar
fashion, we prove that
hf, PM (g)i = hf − PM (f ), PM (g)i + hPM (f ), PM (g)i = hPM (f ), PM (g)i .

Lemma 3.8. [The projection is linear] The two identities PM (kf ) = kPM (f ) and
PM (f1 + f2 ) = PM (f1 ) + PM (f2 ) hold for all k ∈ K, f ∈ H and f1 , f2 ∈ H.
Proof. The proof is based on the trick that H ⊥ = {0} (see Remark 2.4). We have
∀ g ∈ H , hPM (kf ), gi = hkf, PM (g)i [by Lemma 3.7]
= khf, PM (g)i
= khPM (f ), gi = hkPM (f ), gi .
This proves that PM (kf ) − kPM (f ) ∈ H ⊥ = {0}, hence PM (kf ) = kPM (f ). The identity
PM (f1 + f2 ) = PM (f1 ) + PM (f2 ) can be proved in a similar fashion. 

Proof of Theorem 3.1.

The five statements in Theorem 3.1 follow simply by collecting the results proved in
the foregoing lemmas. Lemma 3.3 yields the existence of the function f 7→ PM (f ) along
with Statement (4) in Theorem 3.1.
50 3. HILBERT SPACES

That PM is linear is a consequence of the result in Lemma 3.8. This proves statement
(1) in Theorem 3.1. Statement (2) follows by Remark 3.4. Statement (3) is the content
of Lemma 3.6.
Finally, for statement (5), we pick u ∈ M such that kuk > 0 (this is possible since
M 6= {0}). Consequently, PM (u) = u (by Statement (2)) and the operator norm satisfies
(see Definition 2.5)
kPM (u)k kuk
kPM kL(H) ≥ = = 1.
kuk kuk
On the other hand, for all f ∈ H, PM (f ) ∈ M and f −PM (f )⊥PM (f ). By the Pythagorean
identity (Remark 2.2), kf k2 = kPM (f )k2 + kf − PM (f )k2 ≥ kPM (f )k2 , i.e. kPM (f )k ≤
kf k. This yields that kPM kL(H) ≤ 1. 

4. Applications of the projection theorem

4.1. The second orthogonal complement. The projection theorem has a number
of important applications. One of them regards the second orthogonal complement of
subsets. If M ⊂ H, the second orthogonal complement of M is (M ⊥ )⊥ . In Theorem 4.1,
we will prove that the second orthogonal complement of a subset M is larger than M . If
M is a subspace, we will say more on (M ⊥ )⊥ in Theorem 4.2.
Theorem 4.1. [The second orthogonal complement]
Suppose that M ⊂ H. The orthogonal complement of M ⊥ satisfies
M ⊂ (M ⊥ )⊥ .

Proof. Let N = M ⊥ . We will prove the inclusion M ⊂ N ⊥ . By Theorem 2.6,

⊥
N = M . Consequently,
 
f ∈ M & u ∈ N =⇒ hf, ui = 0 . (4.1)

By Remark 2.4, N is a closed subspace of M . We can apply Theorem 3.1 to get

H = N ⊕ N ⊥ . Let f ∈ M . There exists u ∈ N and v ∈ N ⊥ such that f = u + v. Since
u ∈ N and v ∈ N ⊥ , hu, vi = 0. By (4.1), hf, ui = 0. Consequently, the identity f = u + v
yields
0 = hf, ui = hu + v, vi = hu, vi + hv, vi = 0 + kvk2 = kvk2 .
This proves that v = 0 and f = u + v = u ∈ N . 
Theorem 4.2. [The second orthogonal complement of subspaces]
Suppose that M is a subspace of H. The orthogonal complement of M ⊥ is given by
(M ⊥ )⊥ = M .
In particular, if M is a closed set, then (M ⊥ )⊥ = M .

Proof. Let N = M ⊥ . We will prove that N ⊥ = M . The inclusion M = N ⊥ holds

by Theorem 4.1. So we need to prove the inclusion N ⊥ ⊂ M . Note that:

(*) H = M ⊕ M ⊥ by Theorem 3.1 (see Remark 3.2, (3)) ;

⊥
(**) N = M by Theorem 2.6.
 4. APPLICATIONS OF THE PROJECTION THEOREM 51

Let f ∈ N ⊥ . Using (*), we get u ∈ M and v ∈ M ⊥ = N such that f = u + v. Using

⊥
(**), we get that v ∈ M . Now, since f ∈ N ⊥ and v ∈ N , hf, vi = 0. Also, since u ∈ M
⊥
and v ∈ M , hu, vi = 0. From the identity f = u + v, we get
0 = hf, vi = hu + v, vi = hu, vi + hv, vi = 0 + kvk2 = kvk2 .
Thus, v = 0 and f = u + v = u ∈ M . This proves the inclusion N ⊥ ⊂ M . 

4.2. The Riesz representation theorem. The projection theorem has another
important application, namely the representation of linear forms on a Hilbert space that
will allow us to identify a Hilbert space with its dual. This is the Riesz representation
theorem that has a fundamental role in the development of functional analysis and its
applications to other fields, notably the theory of linear partial differential equations.
In this section, H is a Hilbert space, h·, ·i denotes the inner product in H and k · k
denotes the norm induced by the inner product.
Definition 4.3. [Linear forms and topological dual]
(1) A linear form on H is a function L : H → K such that
• L is linear ;
• L is continuous .
(2) The (topological) dual of H, denoted by H ∗ , is the space of all linear forms in
H, i.e. H ∗ = L(H; K).
Remark 4.4.
(1) H ∗ is endowed with the operator norm: kLkH ∗ = sup kL(f )k. Since R is
f ∈H,kf k≤1
complete, it follows from Theorem 3.2 that (H ∗ , k · kH ∗ ) is a Banach space.
(2) [The fundamental example] We can embed H into H ∗ in the following simple
manner. For all g ∈ H, we define the linear function Lg : H → K by Lg (f ) =
hf, gi. It is easy to check that Lg ∈ H ∗ (see Remark 1.5) and kLg kH ∗ = kgk.
Now we can state the Riesz representation theorem, which explains how we can identity
the spaces H and H ∗ .
Theorem 4.5. [The Risez representation theorem]
7 Lg (·) := h·, gi ∈ H ∗ is an invertible linear isometry. More
The function g ∈ H →
precisely:
(1) For all g ∈ H, kLg kH ∗ = kgk ;
(2) For all L ∈ H ∗ , there exists a unique g ∈ H such that L = Lg ;
(3) For all k ∈ R and g ∈ H, Lkg = kLg ;
(4) For all g1 , g2 ∈ H, Lg1 +g2 = Lg1 + Lg2 .
Analogy with the Euclidean case. Before moving to the proof of Theorem 4.5, we
discuss an analogy between its content and the simple case of Euclidean spaces (or more
generally, finite dimensional inner product spaces). Actually, the content of Theorem 4.5
is of algebraic character in the case where H = Rd , d ≥ 1.
Consider Rd with the usual inner product in (1.2). Let L : Rd → R be a linear
function. The function L admits a canonical representation relative to the canonical basis
B = {e1 = (1, 0 · · · , 0), e2 = (0, 1, 0·, 0), · · · , ed = (0, · · · , 0, 1)} ,
52 3. HILBERT SPACES

namely

 
u1

 u2 

L(u) = [L]B 
 · ,
 u = (u1 , u2 , · · · , ud ) ∈ Rd , (4.2)
 · 
ud

where [L]B is the matrix of L relative to the basis B (this is a 1 × d matrix). We have

 
[L]B = L(e1 ) L(e2 ) · · · L(ed ) .

Using the inner product in (1.2), we express the formula in (4.2) as follows

L(u) = u · g = hu, gi

where g = (L(e1 ), L(e2 ), · · · , L(ed )) ∈ Rd . This is the content of Theorem 4.5 for H = Rd .
Also, we actually proved that every linear function L : Rd → R is continuous (hence a
linear form).
The proof of Theorem 4.5 is of an analytic nature and valid for general Hilbert spaces
(typically infinite dimensional ones). It is one out of many cases where analysis is the
right tool for the passage from finite dimensional settings to infinite dimensional ones.
In proving Theorem 4.5, we will need the following remarkable conclusion:

Lemma 4.6. [Orthogonal complement of a proper subspace]

Suppose that M is a closed subspace of H. If M 6= H, then M ⊥ 6= {0}.

Proof. By the projection theorem (Theorem 3.1 and Remark 3.2), H = M ⊕ M ⊥ .

If M ⊥ = {0}, then H = M ⊕ {0} = M , thereby violating the assumption M 6= H. 
 5. HILBERT BASIS 53

Proof of Theorem 4.5.

We will prove the four statements (1)-(4) in Theorem 4.5 separately.
Proof of (1):
Let g ∈ H. That Lg ∈ H ∗ and kLg kH ∗ ≤ kgk follow from Remark 1.5. In particular,
if g = 0, we deduce that kLg kH ∗ = 0 = kgk. If g 6= 0, the operator norm of Lg satisfies
kLg (g)k hg, gi
kLg kH ∗ ≥ = = kgk .
kgk kgk
Proof of (3) and (4):
Let k ∈ R and g ∈ H. For all f ∈ H, Lkg (f ) = hf, kgi = khf, gi = kLg (f ). This
proves statement (3). In a similar fashion, we prove statement (4).
Proof of (2):
Suppose that L ∈ H ∗ . If L = 0, then L(·) = h·, 0i and we have nothing to prove. So,
we assume henceforth that L 6= 0.
Since L is a linear function, Ker(L) = {f ∈ H : L(f ) = 0} = L−1 ({0}) is a subspace
of H. Also, this is a proper subspace of H, because L 6= 0. The continuity of L yields
that Ker(L) is a closed set. We may apply the result of Lemma 4.6 on M := Ker(L).

⊥ 1
Pick h ∈ Ker(L) such that h 6= 0. Define a = khk h and w = L(a)f − L(f )a.

⊥
Clearly, a ∈ Ker(L) , kak = 1 and w ∈ Ker(L). Consequently
0 = hw, ai = L(a)hf, ai − L(f )ha, ai = L(a)hf, ai − L(f )
which yields L(f ) = hf, gi for g = L(a)a. 

5. Hilbert basis
5.1. Motivation. Consider the Hilbert space Rd with the usual inner product intro-
duced in (1.2) (this inner product induces the Euclidean norm k · k). The space Rd has a
distinguished basis (called the canonical basis)
B = {e1 = (1, 0, · · · , 0), e2 = (0, 1, 0, · · · , 0), · · · , ed = (0, · · · , 0, 1)} .
Let us list some remarkable properties of the canonical basis:
(1) The basis B is orthonormal in the sense:
• For all n ∈ {1, 2, · · · , d}, ken k = 1 ;
• For all n 6= m, hen , em i = 0 .
(2) Every u ∈ Rd can be expressed in the pleasant form u = dn=1 hu, en ien .
P
(3) (Parseval’s identity) The norm of every u ∈ Rd can be expressed in the form
kuk2 = dn=1 (hu, en i)2 .
P

The canonical basis B is an example of a Hilbert basis. We will discuss the existence of
such a basis in a general Hilbert space.
5.2. Orthonormal subsets. Here we introduce many notions (from linear algebra)
that will lead us to the remarkable Bessel’s inequality in inner product spaces.
In this section, we assume that H is an inner product space whose inner product is
h·, ·i. The norm in H induced by the inner product is denoted by k · k.
Definition 5.1. Let S ⊂ H.
(1) [Orthogonal subsets] The set S is said to be orthogonal if, for all f, g ∈ H, it
holds
f 6= g =⇒ hf, gi = 0 .
54 3. HILBERT SPACES

(2) [Orthonormal subsets] The set S is said to be orthonormal if:

• S is orthogonal ; and p
• ∀ f ∈ H, kf k = 1. (Here, k · k = h·, ·i ).
Remark 5.2. [Linear algebra]
(1) [Orthogonality and linear independence] Suppose that S is an orthogonal
subset and 0 6∈ S. Then S is linearly independent. Actually, if f1 , f2 , · · · , fn are
distinct elements in S and
λ1 f1 + λ2 f2 + · · · + λn fn = 0
for some λ1 , λ2 , · · · , λn ∈ K, then, for all m ∈ {1, 2, · · · , n}, we take the scalar
product with fm to obtain
n
X
0 = hλ1 f1 + λ2 f2 + · · · + λn fn , fm i = λk hfk , fm i = λm kfm k2 ,
k=1

which yields that λm = 0, since fm 6= 0 (here we used that fm ∈ S and 0 6∈ S).

(2) By the foregoing remark, we find that every orthonormal subset is linearly inde-
pendent.
(3) [Generalized Pythagorean identity] Suppose that S = {f1 , f2 , · · · , fn } is an
orthogonal subset of H. A straightforward computation using the properties of
the inner product yields, for all λ1 , λ2 , · · · , λn in K,
n 2 n
X X
λk fk = |λk |2 kfk k2 . (5.1)
k=1 k=1

More generally, we can verify that, for all α1 , α2 , · · · , αn in K,

* n n
+ n
X X X
λk f k , αk f k = λk αk kfk k2 . (5.2)
k=1 k=1 k=1

(4) [Subspaces spanned by a subset] Let S ⊂ H. The (vector) subspace spanned

(generated) by S is the subset of H denoted by span(S) and defined as follows:
f ∈ span(S) if and only if f is a (finite) linear combination of elements of S, i.e.
n
X
n n
f ∈ span(S) ⇔ ∃ n ∈ N , ∃ (f1 , f2 , · · · , fn ) ∈ S , ∃ (λ1 , λ2 , · · · , λn ) ∈ K , f = λk f k .
k=1

(5) [The Gram-Schmidt process] Suppose that B = {u1 , u2 , · · · } ⊂ H is a linearly

independent set. Let M = span(B). We can construct a new set S = {f1 , f2 , · · · }
such that
• S is orthonormal ; and
• M = Span(S).
Actually, for all n ≥ 1, we define the vector fn as follows
1
fn = wn
kwn k
where the vectors w1 , w2 , · · · are defined by recursion as follows
n−1
X hun , wk i
w 1 = u1 and ∀ n ≥ 2 , wn = un − wk .
k=1
kwk k2
 6. SEPARABLE SPACES 55

Theorem 5.3. [Bessel’s inequality]

Assume that S = {f1 , f2 , · · · , fn , · · · } is an orthonormal subset of H. Then, for all
f ∈ H, the following holds
X∞
(1) The series |hf, fn i|2 is convergent ;
n=1
∞
X
(2) [Bessel’s inequality] |hf, fn i|2 ≤ kf k2 .
n=1
n
X
Proof. For all n ≥ 1, let rn = f − hf, fk ifk . For 1 ≤ j ≤ n, we have
k=1
n
X
hrn , fj i = hf, fj i − hf, fk ihfk , fj i = hf, fj i − hf, fj i = 0.
k=1
Pn
Hence rn is orthogonal to S and then from f = rn + k=1 hf, fk ifk and by using the
generalized Pythagorean identity, we have

n
X n
X n
X n
X
kf k2 = krn + hf, fk ifk k2 = krn k2 + k hf, fk ifk k2 ≥ k hf, fk ifk k2 = |hf, fk i|2
k=1 k=1 k=1 k=1
so n
X
2
kf k ≥ |hf, fk i|2 .
k=1
∞
X
That is, the sequence of partial sums of the series |hf, fn i|2 is bounded. Since this is
n=1
a series of positive terms we conclude that it is convergent. Sending n to ∞, in the above
inequality, yields the Bessel’s inequality.
X∞
|hf, fn i|2 ≤ kf k2 .
n=1


6. Separable spaces
Here we introduce the notion of separable Hilbert spaces. In such spaces, we will
see that Paseval’s identity holds (this is where equality occurs in Bessel’s inequality -
Corollary ??). Bessel’s inequality is valid for any choice of an orthonormal subset S = (fn ).
However, for equality to occur, we need that S is in addition a Hilbert basis, a term
that we will define below.
Definition 6.1. [Hilbert basis]
Suppose that (H, h·, ·i) is an inner product space and S ⊂ H. The set S is said to be
a Hilbert basis of H if
• S is orthonormal ;
• span(S) is dense in H.
Note that we deviated from the standard algebraic definition of a basis. Actually, S
is a basis of H if span(S) = H. To be a Hilbert basis, we need only that span(S) = H,
that is equality is relaxed and replaced by density.
56 3. HILBERT SPACES

Of special importance are separable Hilbert spaces since they resemble very much
the Euclidean space Rd . The precise definition is:
Definition 6.2. [Separable spaces]
An inner product space (H, h·, ·i) is said to be separable if it has a countable Hilbert
basis.
Next, we establish how and in what sense do separable Hilbert spaces resemble the
Euclidean spaces. But let us first clarify the following notion: If (un ) is a sequence in an
inner product space H and u ∈ H, by writing
X∞
u= un (in H)
n=1
n
X
we mean that the sequence sn = uk is convergent to u in H.
k=1
Now we are ready to state:
Theorem 6.3. Suppose that (H, h·, ·i) is a separable Hilbert space. Let S = {f1 , f2 , · · · }
be a Hilbert basis of H. Then it holds the following:
(1) (Hilbertian decomposition) For all f ∈ H,
X∞
f= hf, fn ifn in H .
n=1

(2) (Parseval’s identities)

∞
X
(a) For all f, g ∈ H, hf, gi = hf, fn ihfn , gi,
n=1
∞
X
(b) For all f ∈ H, kf k2 = |hf, fn i|2 .
n=1

The proof of Theorem 6.3 relies on the following lemma.

Lemma 6.4. Assume that S = {f1 , f2 , · · · } is an orthonormal subset of a Hilbert
Xn
space H. Let f ∈ H and define the sequence (sn ) ⊂ H by sn = hf, fk ifk , n ≥ 1.
k=1
The following statements hold true:
(1) The sequence (sn ) is a Cauchy sequence in H .
(2) There exists u ∈ H such that sn → u in H .
(3) If there exist  > 0 and u ∈ span(S) such that kf − u k < 3 , then kf − uk < .
Proof.
Proof of (1) and (2):
For all n, p ≥ 1, we write
ksn+p − sn k2 = ksn+p k2 + ksn k2 − 2hsn+p , sn i .
Using the the Pythagorean identities in (5.1) and (5.2), and the orthogonality of S = (fn ),
we get the following identity
n+p
X
2
ksn+p − sn k = |hf, fk i|2 .
k=n+1
 6. SEPARABLE SPACES 57

∞
X
By Bessel’s inequality, the series |hf, fk i|2 is convergent. Thus,
k=n+1
∞
X
∀  > 0, ∃ n0 ≥ 1 , ∀n ≥ n0 , ∀ p ≥ 1 , |hf, fk i|2 <  ,
k=n+1

and consequently
n+p ∞
X X
2 2
ksn+p − sn k = |hf, fk i| ≤ |hf, fk i|2 <  .
k=n+1 k=n+1

Thus, (sn ) is a Cauchy sequence in the Hilbert space H, hence convergent to some limit
u ∈ H.
Proof of (3):
Suppose that u ∈ span(S) and kf − u k < 3 . Since sn → u, there exists n0 ≥ 1 such
that, for all n ≥ n0 , ksn − uk < 3 .
Since u ∈ span(S), there exist N ≥ n0 and λ1 , λ2 , · · · , λN ∈ R such that
X N
u = λk fk . Using the Pythagorean identities in (5.1) and (5.2), we get
k=1
N 
X 2
ksN − u k2 = hf, fk i − λk . (6.1)
k=1
Using Bessel’s inequality,
N
X 2
|hf − u , fk i|2 ≤ kf − u k2 < . (6.2)
k=1
9
But, by the orthogonality of S = (fn )n≥1 ,
N
X
∀ k ≥ 1, hf − u i = hf, fk i − hλm fm , fk i = hf, fk i − λk .
m=1
N 
X 2 2
Thus, (6.2) becomes hf, fk i − λ2k < . Inserting this into (6.1), we get
k=1
9

ksN − u k <
.
3
But, since N ≥ n0 , we know that ksN − uk < 3 . Thus, by the triangle inequality,
  
kf − uk ≤ kf − u k + ku − sN k + ksN − uk < + + =  .
3 3 3

With Lemma 6.4 in hand, we can present the

Proof of Theorem 6.3.

Proof of (1):
Let f ∈ H be fixed. By Lemma 6.4, the sequence (sn ) ⊂ H defined by sn =
n
X
hf, fk ifk , n ≥ 1, is convergent to some u ∈ H.
k=1
58 3. HILBERT SPACES

Let  > 0 be arbitrary but fixed. Since S = (fn )n≥1 is a Hilbert basis of H, then,
span(S) is dense in H. Consequently, there exists u ∈ span(S) such that kf − u k < 3 .
By Lemma 6.4, kf − uk < . Since  > 0 is arbitrary, we get that u = f .
X∞
This proves that hf, fn ifn = f .
n=1
Proof of (2):
Let f, g ∈ H be fixed. We know that
X n n
X
sn = hf, fk ifk → f and tn = hg, fk ifk → g in H .
k=1 k=1

Using the orthonormality of the set S we get

* n n
+ n X
n n
X X X X
hsn , tn i = hf, fk ifk , hg, fj ifj = hf, fk ihg, fj ihfk , fj i = hf, fk ihg, fk i.
k=1 j=1 k=1 j=1 k=1

Now by continuity of the norm and the inner product (see Theorem 1.8) we have
n
X ∞
X ∞
X
2
hf, gi = lim hsn , tn i = lim hf, fk ihg, fk i = hf, fk ihfk , gi and kf k = |hf, fk i|2 .
n→∞ n→∞
k=1 k=1 k=1


We will end this chapter by TheoremX 6.5. In its statement, `2 (K) is the space of all
sequences (an ) ⊂ K such that the series |an |2 is convergent. This is a Hilbert space
n≥1
with the following inner product
X
h(an ), (bn )i`2 = an b n .
n≥1

Two Hilbert spaces H1 and H2 are said to be unitary isomorphic if there exists an
invertible function U : H1 → H2 such that
∀ f, g ∈ H1 , hU (f ), U (g)iH2 = hf, giH1 .
Theorem 6.5. Let H be a separable Hilbert space. Then H is unitarily isomorphic
either to Kn , for some n ∈ N, or to the space `2 (K).
Proof. Suppose H is a separable Hilbert space. Assume first that dim(H) = +∞.
Then H has a dense countable (and infinite) orthonormal basis S = {f1 , f2 , · · · }. Consider
the map
U : H → `2 (K)
defined by  
U (f ) = αn (f ) = hf, fn i .
n≥1
∞
X
For all f ∈ H, Bessel’s inequality yields that the sequence (αn ) ⊂ ` (K) since 2
|αn |2 =
n=1
∞
X
|hf, fn i|2 < +∞. And also Parseval’s identity gives
n=1

kU (f )k`2 = kf k. (6.3)
 6. SEPARABLE SPACES 59

That U is linear follows from the properties of the inner product h·, ·i. That U is injective
follows from the identity (6.3).
= (αn )n≥1 ∈ `2 (K). An easy adaptation
It remains to check that U is surjective. Let αP
of the proof of Lemma 6.4 yields that the series ∞ n=1 αn fn is convergent in H. Hence it
converges to some f ∈ H. Consequently, we find that U (f ) = α.
If the Hilbert space H is finite dimensional and dim(H) = n, then H has an or-
thonormal basis formed by n vectors so H is unitarily isomorphic to KN , see Lemma 5.4
(Chapter 1). 
Remark 6.6. We can just think about separable Hilbert spaces as `2 (K) or Kn when
we wish. To be more precise, the calculus in a general separable Hilbert space resembles
that in `2 (K) or Kn .
60 3. HILBERT SPACES

Exercises

Exercise 1.
Tell whether the following maps, are inner product or not on the vector space E :
n
X
(1) hx, yi = xi yi3 , E = Rn ;
i=1
∞
X ∞
X
2
(2) hx, yi = xn y n , E = ` (R) = {x = (xn )n : xn ∈ R and |xn |2 < +∞};
Zn=1 Z n=1
2 2
(3) hf, gi = f (x)g(x)e−x dx, E = {f ∈ C(R, R) : |f (x)|2 e−x dx < ∞};
ZR1 R
0 0 1
(4) hf, gi = f (x)g (x)dx, E = C ([0; 1], R);
0
Z 1
(5) hf, gi = f 0 (x)g 0 (x)dx, E = {C 1 ([0; 1], R) : f(0)=0}
Z0 1  
(6) hf, gi = f (x)g(x) + f 0 (x)g 0 (x) dx, E = C 1 ([0; 1], R).
0

Exercise 2.
Let x and y be two nonnull vectors of a real pre-Hilbert space (E, h·, ·i) . Show that
x y kx − yk
2
− 2
= .
kxk kyk kxkkyk

Exercise 3. Let (E, h·, ·i) be a real pre-Hilbert space and u : E 7−→ E be a map. We
designate by k · k the associated norm to h·, ·i.
(1) Suppose that u verify the following conditions:

u(0) = 0 and ku(x) − u(y)k = kx − yk, ∀x, y ∈ E.

(a) Show that hu(x), u(y)i = hx, yi, for all x, y ∈ E.

(b) Prove that ku(x + y) − u(x) − u(y)k = 0, for all x, y ∈ E.
(c) Deduce then that u is a linear mapping.
(2) Show the polarisation identity:
1
kx + yk2 + kx − yk2


hx, yi = ∀ x, y ∈ E.
4
(3) Deduce that if u is linear and verifies ku(x)k = kxk, for all x ∈ E, then
hu(x), u(y)i = hx, yi.

Exercise 4. Let (E, h·, ·i) be a real pre-Hilbert space and u : E 7−→ E be an application
verifing
hf (x), yi = hx, f (y)i
for all x, y ∈ E. Show that f is linear.
 EXERCISES 61

Exercise 5. Let (E, h·, ·i) be a real pre-Hilbert space. Show that , for all x, y ∈ E,
h i h
|hx, yi| = kxk · kyk ⇐⇒ x and y are two linearly dependent vectors
i
(i.e. there exist α ∈ R such that y = αx)

Exercise 6. Let E = C([0, 1], R) equipped with norm k · k∞ . Put f (x) = x and g(x) = 1.
(1) Show that kf − gk2∞ + kf + gk2∞ 6= 2(kf k2∞ + kgk2∞ ).
(2) what can be said about the space (E, k · k∞ )?

Exercise 7. [The Hahn-Banach theorem for a ball]

Let (H, h·, ·i) be a real pre-Hilbert space and let x0 ∈ H. Suppose for R > 0 the open
ball B = B(x0 , R) of H does not contain 0. The aim of this exercise is to show that there
exist a continueous linear form ϕ ∈ H ∗ such that, for all x ∈ B, we have ϕ(x) > 0. to do
that, we define ϕ by
ϕ(x) = hx0 , xi, ∀ x ∈ H
(1) show that, for all u ∈ B(0, 1), we have ϕ(x0 + Ru) ≥ kx0 k(kx0 k − R).
(2) Deduce that ϕ(x) > 0, for x ∈ B.

Exercise 8. Let (X, h·, ·i) be a real Hilbet space, a ∈ X and f : X 7−→ X be a mapping
such that:

• hf (x) − f (y), x − yi ≥ 0, ∀x, y ∈ X,

• There exist β > 0 such that kf (x) − f (y)k ≤ βkx − yk, ∀x, y ∈ X.

The aim is to show that the equation

x + f (x) = a (E)
has a unique solution α ∈ X.
(1) Suppose that β < 1.
Show the application g defined by g(x) = a − f (x) is a contraction. Then deduce
that the equation (E) admits a unique solution in X.
(2) Now suppose that β ≥ 1 and 0 < λ < 1.
(a) Show that x is a solution of E if and only if x is a solution of

x = (1 − λ)x − λf (x) + λa.

(b) Let gλ = (1−λ)x−λf (x)+λa. Find a value of λ such that gλ is a contraction.
(c) Deduce that the equation (E) admits a unique solution in X.

Exercise 9. Let (H, h·, ·i) be a real Hilbert space and f : H 7−→ R be a linear continueous
form (i.e f ∈ H ∗ ). We define the continueous map φ : H 7−→ R by:
1
φ(x) = kxk2 − f (x),
2
where k · k is the associated norm to the scalar product h·, ·i.
62 3. HILBERT SPACES

(1) Show that, for all, x, y ∈ H, we have

x + y  1  1
φ = φ(x) + φ(y) − kx − yk2 , ∀ x, y ∈ H.
2 2 8
(Hint: we can use the parallelogram law)
(2) Let F be a closed sub-vector space of H and put d = inf φ(y).
 x + y y∈F

(a) Show that, for all x, y ∈ F , we have d ≤ φ .
2
(b) Let (xn ) be a sequence in F such that lim φ(xn ) = d.
n→∞
(i) Deduce, from the previous parts, that
1
kxp − xq k2 ≤ φ(xp ) + φ(xq ) − 2d.
4
(ii) Show that (xn ) is a Cauchy sequence in H. Deduce that (xn ) converges
to a limit a ∈ F .
(iii) Show that φ(a) = d (i.e. φ attains its minimum on F ).

Exercise 10. Let E = C 1 ([0, 1], R) equipped with the inner product
Z 1 
hf, gi = f (x)g(x) + f 0 (x)g 0 (x) dx.
0

Let u(x) = sin x et v(x) = cos x. Prove that u and v are orthogonal with respect to h·, ·i.
Deduce then the value of ku + vk.

Exercise 11. Let (E, h·, ·i) be a real pre-Hilbert space. Let u and v be two vectors such

that, for all λ ∈ R, ku + λvk ≥ kuk. Show that u and v are orthogonal.
Z 1
Exercise 12. Let E = C([0, 1], R) equipped by the inner product hf, gi = f (x)g(x)dx.
0
We define the sub-vector space F by
F = {f ∈ E : f (0) = 0}.
Calculate hf, xf i for f ∈ F ⊥ . Deduce that F ⊥ = {0}.
Z b
Exercise 13. Let E = C([a, b], R) equipped with the inner product hf, gi = f (x)g(x)dx.
a
We designate by F = R[x] the sub-vector space of E formed by polynomials.
(1) Prove that, for f ∈ F ⊥ and for P ∈ R[x], we have
kf k2 ≤ (b − a)kf k∞ kf − P k∞
(2) Deduce, by using the theorem of approximation of Weierstrass, that (F )⊥ = {0}.

Exercise 14 We consider the Hilbert space

( ∞
)
X
`2 (R) = x = (xn )n≥1 : xn ∈ R, |xn |2 < +∞ ,
n=1
 EXERCISES 63

∞
X
equipped with the inner product hx, yi = xn yn . For N ∈ N∗ fixed, we define the map
n=1
fN : `2 → R by
N
X
fN (x) = xn ∀ x = (xn )n≥1 .
n=1
(1) Show that (
fN is linear and continueous. )
XN
(2) Let AN = x = (xn )n≥1 ∈ `2 (R) : xn = 0 .
n=1
(a) Show that AN is closed in `2 (R) and that `2 (R) = AN (AN )⊥ .
L
(b) What c an be said about (AN , h·, ·i)?
(c) Let
n o
E = y = (yn )n≥1 ∈ `2 (R) : (yn = y1 , for 1 ≤ n ≤ N ) and (yn = 0 for n > N ) .

(i) Verify that E ⊂ (AN )⊥ .

(ii) For all i ≥ 1, we consider the sequences ai = (ain ) ∈ `2 and bi = (bin ) ∈
`2 defined by

1
 if n = 1 
1 if n = i, and i > N
ain = − 1 if n = i, and 1 < i ≤ N , bin =

0 0 otherwise.
otherwise.
Verify that, for all i ≥ 1, we have ai , bi ∈ AN .
(iii) Calculate hai , xi for 1 < i ≤ N and hbi , xi for i > N
(iv) Deduce that (AN )⊥ = E.

Exercise 15. We consider the Hilbert space E = R4 equipped with the inner product
4
X
hx, yi = xi yi . Let E1 = {(x1 , x2 , x3 , x4 ) ∈ E : x1 = x3 , x2 = x4 }.
i=1
(1) Prove that E1 is a closed sub-vector space of E.
(2) Prove that the two vectors a1 = (1, 0, −1, 0) and a2 = (0, 1, 0, −1) are orthogonal
to E1 , and deduce that E1⊥ .
(3) Let A = {x ∈ E1 : kxk = 2} and B = {x ∈ E1⊥ : kxk = 2}.

(a) Show that A ∩ B = ∅.

(b) Show that d(A, B) > 0 and calculate d(A, B).

Exercise 16. Let H = L2 (]0, +∞[). We define

Z +∞
hf, gi = e−x f (x)g(x)dx, ∀ f, g ∈ E.
0

(1) Show that h·, ·i defines an inner product on H.

(2) We admit that (H, h·, ·i) is a Hilbert space and we designate by k · k the induced
norm. We consider the following functions f1 (x) = x, f2 (x) = x2 .
64 3. HILBERT SPACES
n o
(a) Let K = f :]0, +∞[7−→ R : f (x) = af1 (x) + bf2 (x), with a, b ∈ R . Say
why K is a closed sub-vector space of H.
(b) Find the values of the reals a and b such that the following integral is mini-
mal:
Z +∞  2
I(a, b) = e−x x3 − af1 (x) − bf2 (x) dx.
0
Z +∞
(We may use the formula: xn e−x dx = n!)
0
(c) Find then min I(a, b).
a,b∈R

Exercise 17. Let H = L2 (] − π, π[) be the Hilbert space equipped with the inner product
Z π
1
hf, gi = f (x)g(x)dx.
2π −π
We designate by k · k the induced norm. √ √
We consider the functions f0 (x) = 1, f1 (x) = 2 cos x and f2 (x) = 2 sin x.

(1) Show thatn the family f0 , f1 , f2 is orthonormal in H. o
(2) Let K = f :]−π, π[7−→ R : f (x) = af0 (x)+bf1 (x)+cf2 (x), with a, b, c ∈ R .
Say why K is a closed a sub-vector space of H?
(3) Let f (x) = x2 . Determine hf, f0 i, hf, f1 i and hf, f2 i.
(4) Deduce, from the previous parts, the values of the real a, b and c such that the
integral
Z π
1 2
J(a, b, c) = x2 − af0 (x) − bf1 (x) − cf2 (x) dx
2π −π
is minimal.

Exercise 18. Find the real numbers a and b so that the following integrals are minimal:
Z 1 Z 1
2 2
(1) (x − a − bx) dx (2) (ex − a − bx)2 dx.
0 0

Exercise 19. Let E = C 1 ([0, 1], R) equipped with the inner product

Z 1  
hf, gi = f (x)g(x) + f 0 (x)g 0 (x) dx.
0
Put
00
V = {f ∈ E : f (0) = f (1) = 0} and W = {f ∈ E : f is C 2 and f = f }.
(1) Prove that V and W are orthogonal.
f (1) − f (0)ch(1)
(2) For f ∈ E, we define g(x) = f (0)ch(x) + sh(x) and h = f − g.
sh(1)
(a) Verify that g ∈ W and h ∈ V .
(b) Deduce that V and W are complementary. Express the orthogonal projec-
tion on W .
 EXERCISES 65

(c) Let Eα,β = {f ∈ E : f (0) = α, f (1) = β}. By using the function g calculate
Z 1
inf (|f (t)|2 + |f 0 (t)|2 )dt.
f ∈Eα,β 0

Exercise 20. Let H be Hilbert space, and F be a closed sub-vector space of H, not
reduce to zero. We denote by p the orthogonal projection of H on F . Show that:
(1) p ◦ p = p.
(2) For all x, y ∈ H, we have hp(x), yi = hx, p(y)i.
(3) kpk = 1.

Exercise 21. Let E = L2 (]0, 1[) equipped with the inner product h·, ·i and let
Z 1
F = {f ∈ E : f (x)dx = 0}.
0
(1) Prove that F is a closed sub-space of E.
(2) Prove that F ⊥ = span{h}, where h ≡ 1.
(3) Let g(x) = x2 . Calculate the orthogonal projection of g on F .

Exercise 22. Let E = C[0, 1] equipped with the norm k · k∞ and let K be a sub-space
of E defined by
n Z 1 o
K = f ∈ E; f (0) = 0 and f (x)dx = 0 .
0
(1) Prove that K is a closed convex of (E, k · k∞ ).
(2) Take f0 (x) = x and the sequence of functions (fn ) defined, for all n ≥ 1, by

(n − 1)2
− x if 0 ≤ x ≤ n1



 2n − 1
fn (x) =
 n
 x− if n1 < x ≤ 1.


2n − 1
(a) Prove that fn ∈ K andZ f0 6∈ K.
1
(b) For f ∈ K, calculate (f (x) − f0 (x))dx . Deduce that
0
1
kf − f0 k∞ ≥ , ∀ f ∈ K.
2
n 1
(c) Prove that kfn − f0 k∞ = . Deduce that inf kf − f0 k∞ = .
2n − 1 f ∈K 2
(d) Suppose there exist a function f1 ∈ K such that kf1 −f0 k∞ = inf kf −f0 k∞ .
f ∈K
In this case, we say that f1 is the orthogonal projection of f0 on K.
(i) Put g(x) = kf1 − f0 k∞ − (f0 (x) − f1 (x)). Verify that
Z 1
g(x)dx = 0 et g ≥ 0.
0
(ii) Deduce that f1 = f0 and that f1 6∈ K. Give your conclusion.
(iii) Is (E, k · k∞ ) a Hilbert space? Justify.
66 3. HILBERT SPACES

Exercise 23.[Projection on a ball] Let (H, h·, ·i) be a real Hilbert space. Let B =
B(0, 1) be the unit closed ball of H. We denote by PB the orthogonal projection of H on
B.
(1) Le but de cette partie est de trouver l’expression de PB .
(a) Let x 6∈ B. Verify that, for all z ∈ B, we have
x x
hz − ,x − i≤0
kxk kxk
(b) Deduce that

x if x ∈ B
PB (x) = x
if x 6∈ B.
kxk


(2) Prove that if x and y are nonnull elements of H then

x y kx − yk
− ≤2 .
kxk kyk kxk + kyk
(3) Deduce that PB is 1-Lipschitz.
Z 1
Exercise 24. Let E = C([0, 1], R) equipped with the inner product hf, gi = f (x)g(x)dx.
Z 1/2 0

Let u be the map defined by u(f ) = x2 f (x)dx

0
(1) Show that u is a linear continueous form on E (i.e. u ∈ E ∗ ). Calculate kuk.
(2) We suppose that there exist g ∈ E such that u(f ) = hf, gi for all f ∈ E.
(a) We consider the sequence (fn )n ⊂ E defined by


 0 if 0 ≤ t ≤ 21 ,



fn (t) = ng(t)(t − 12 ) if 12 < t < 12 + n1 ,



if 12 + n1 ≤ t ≤ 1.

 g(t)

(i) Calculate u(fn ) and hfn , gi.

(ii) Deduce that g(t) = 0 if 21 ≤ t ≤ 1.
(b) Let f (x) = x2 − g(x). Show that g(x) = x2 if 0 ≤ x ≤ 12 .

(c) What can be said about (E, h·, ·i)?

Exercise 25.

Let E be the pre-Hilbert space of all sequences u = (un )n≥1 satisfying:

∃ N ∈ N such that ∀ n ≥ N, un = 0,
equipped with the inner product defined, for all u, v ∈ E, by
X∞
hu, vi = un vn .
n=1
 EXERCISES 67

We designate by k · k the induced norm.

∞
X un
(1) Show that the map φ : (E, k · k) → R defined by φ(u) = is linear and
n=1
n
continueous.
(2) We suppose that there exist a = (an )n ∈ E such that φ(u) = hu, ai for all u ∈ E.
(a) For all i ∈ N∗ , we consider the sequence ui = (uin )n ∈ E defined by:
1 if n = i

i
un =
0 if n 6= i.

(i) Calculate φ(ui ) and hui , ai.

1
(ii) Deduce that ai = .
i
(b) Is (E, h·, ·i) a Hilbert space?

Exercise 26. We consider the Hilbert space

( ∞
)
X
`2 (R) = x = (xn )n≥1 : xn ∈ R, |xn |2 < +∞ ,
n=1
∞
X
equipped with the inner product hx, yi = xn yn . We designate by k · k`2 the induced
n=1
norm la norm. Let N ∈ N∗ fixed and defined the map fN : `2 → R by
N
X √
fN (x) = n xn ∀ x = (xn )n≥1 .
n=1

(1) Show that fN is linear and continueous.

(√
n if 1 ≤ n ≤ N
(2) Let a = (an ) ∈ `2 such that an =
0 otherwise.
2
(a) Determine ha, xi for all x ∈ ` . Calculate
! then kfN k.
N
X N (N + 1)
Recall that n=
n=1
2

Exercise 27. Let E = C([−1, 1], R). For f, g ∈ E we define

Z 1
f (t)g(t)
hf, gi = √ dt.
−1 1 − t2
(1) Show that (f, g) 7−→ hf, gi is well defined and it defines an inner product on E.
(2) We admit that (L2 (] − 1, 1[), h·, ·i) is a Hilbert space. We consider the sequence
(Pn )n of polynomials defined by
 q
 Pn (x) = 2 cos(n arccos x) where n ≥ 1;
qπ
 P0 (x) = 1 .
π
68 3. HILBERT SPACES

(a) Recall the formula cos p + cos q = 2 cos( p+q 2

) cos( p−q
2
). Prove that for all
n ≥ 1 we have:
r
2
Pn+1 (x) + Pn−1 (x) = 2x Pn (x)
π
(b) Deduce that Pn is a polynomial of degree n for all n ∈ N.
(c) Prove that (Pn )n is orthonormal.
(d) Prove that the family (Pn ) is total in E.
(e) Prove that the family (Pn ) forms a Hilbert basis of (L2 (] − 1, 1[), h·, ·i).

Exercise 28.
(1) Let (H, h·, ·i) be Hilbert space and (ei )i∈I be an orthonormal family. Show that
(ei )i∈I is a hilbert basis of H if and only if
hx, ei i = 0, ∀ i ∈ I =⇒ x = 0
(2) For all i ≥ 1, we on consider the sequence ei = (ein ) ∈ `2 (R) defined by
1 if i = n

i
en =
0 if i 6= n.
Show that (ei )i≥1 is a Hilbert basis of `2 (R).
 CHAPTER 4

FOURIER SERIES

In this chapter, we will apply the theory of Hilbert spaces in the specific setting of the
spaces C([0, 2π], R) and L2 ([0, 2π], R), both endowed with the following inner product
Z 2π
hf, gi = f (x)g(x) dx .
0

This inner product induces the norm (commonly called L2 -norm)

Z 2π 1/2
2
kf k2 = |f (x)| dx ,
0

We remind the reader that C([0, 2π]) is the space of real-valued continuous functions
on the interval [0, 2π]. The space L2 ([0, 2π]) is the space of square integrable functions:
It is a Hilbert space containing C([0, 2π]) as a dense subset (relative to the L2 -norm).
More details regarding this loosely defined space are given in Remark 7.6.
We will prove that the space L2 ([0, 2π]) is separable (see Chapter 3 - Definition 6.2),
i.e. it has a countable Hilbert basis. We will do this by discussing a special Hilbert basis
in L2 ([0, 2π]), namely the trigonometric basis defined by orthogonal sine and cosine
functions. The Hilbertian decomposition of functions relative to the trigonometric basis
(in the sense described in Chapter 3 - Theorem 6.3) will lead us to the notion of Fourier
series. Our approach to the subject will show us immediately that the Fourier series
are convergent relative to the L2 -norm, but also we will elaborate the discussion of their
convergence in other directions (uniform and point-wise convergence).

1. Density in the space of continuous functions

In the space C([0, 2π]), we have two important norms, the L2 -norm
Z 2π 1/2
2
kf k2 = |f (x)| dx , (1.1)
0

and the norm of uniform convergence

kf k∞ = sup |f (x)| . (1.2)
x∈[0,2π]

Both norms have their own advantage. One useful remark is the inequality
√
kf k2 ≤ 2πkf k∞ , (1.3)
which simply says that, if a sequence (fn ) converges (to f ) relative to the norm of uniform
convergence, then it converges (to the same limit f ) relative to the L2 -norm.
As opposed to Euclidean spaces, it is a quite delicate task to answer the following
question in the space of continuous functions: Prove that a subset D of C([0, 2π]) is dense
in C([0, 2π]) (relative to one of the norms k · k∞ or k · k2 ). The Stone-Weirstrass theorem
provides us with a wide class of dense subsets in C([0, 2π]). We state this theorem below.
69
70 4. FOURIER SERIES

Theorem 1.1. Let D ⊂ C([0, 2π]). Suppose that:

(1) D is an algebra, i.e.
• ∀ f, g ∈ D, f + g ∈ D and f g ∈ D ;
• ∀ f ∈ D, ∀ λ ∈ R, λf ∈ D .
(2) D is unitary, i.e. the function 1 : x ∈ [0, 2π] 7→ 1(x) = 1 belongs to D .
(3) D separates points, i.e. for all a, b ∈ [0, 2π], if a 6= b, then there exists f ∈ D
such that f (a) 6= f (b).
Then D is dense in (C[0, 2π], k · k∞ ).
We will not write down the proof 1 of Theorem 1.1. Instead, we will apply Theorem 1.1
on some specific examples.
The first example is when D is the space of polynomials. It is quite simple to verify that
D is an algebra, it is unitary and separates points. Consequently, we get the Weirstrass
density theorem, namely that the polynomials are dense in (C([0, 2π]), k · k∞ ).
The second example involves the trigonometric sine and cosine functions. In the rest
of this chapter, we denote by
 1 1 1
S = φ0 (x) = √ , φn (x) = √ cos(nx) , ψn (x) = √ sin(nx) : n = 1, 2, · · · . (1.4)
2π π π
Note that span(S), the space of all finite linear combinations of elements in S, can be
written in the following simple form:
Xn
ak cos(kx)+bk sin(kx) : n ∈ N , (a0 , a1 , · · · , an , b1 , · · · , bn ) ∈ R2n+1 } .


span(S) = {a0 +
k=1
(1.5)
We will prove that:
Theorem 1.2. The set span(S) introduced in (1.5) is dense in (C([0, 2π]), k · k∞ ).
Proof. It is easy to check that span(S) is an algebra2. Taking n = 1, (a0 = 1, a1 =
0, b1 = 0), we get that
n
X

1(x) = a0 + ak cos(kx) + bk sin(kx) ,
k=1

hence span(S) is unitary. Finally, if a, b ∈ [0, 2π] and a 6= b, then we can find a function
f ∈ D such that f (a) 6= f (b). Indeed, we choose f as follows:
• If a, b ∈ [0, π] (or a, b ∈ [π, 2π]), we choose f (x) = cos x ;
• If a ∈ [0, π] and b ∈ [π, 2π] (or a ∈ [π, 2π] and b ∈ [0, π]), we choose f (x) = sin x .
Hence, span(S) separates points. Therefore, Theorem 1.1 yields that D = span(S) is
dense in (C([0, 2π]), k · k∞ ). 
Consequently, we get the following two density results relative to the L2 -norm:
Corollary 1.3. The set span(S) introduced in (1.5) is dense in (C([0, 2π]), k · k2 ).
1There are many many proofs of the theorem, one simple proof is by B. Brosowski and F.
Deutsch ‘An elementary proof of the Stone-Weirstrass theorem, Proceedings of AMS, Vol. 81,
No. 1, 1981 (http://www.ams.org/journals/proc/1981-081-01/S0002-9939-1981-0589143-8/S0002-9939-
1981-0589143-8.pdf).
2An algebra is a vector space which is closed under multiplication and satisfying the usual algebraic
rules.
 2. THE TRIGONOMETRIC SYSTEM 71

Proof. Let  > 0 and f ∈ C([0, 2π]). By Theorem 1.2, there exists f ∈ span(S)
such that kf − f k∞ < √2π . Using the inequality in (1.3), we get that kf − f k2 < . 

Corollary 1.4. The set span(S) introduced in (1.5) is dense in (L2 ([0, 2π]), k · k2 ).
Proof. Let  > 0 and f ∈ L2 ([0, 2π]). Since C([0, 2π]) is dense in L2 ([0, 2π]), there
exists u ∈ L2 ([0, 2π]) such that kf − u k2 < 2 .
By Theorem 1.3, there exists f ∈ span(S) such that ku −f k2 < 2 . Using the triangle
inequality, we get
 
kf − f k2 ≤ kf − u k2 + ku − f k2 < + = .
2 2


2. The trigonometric system

The set S introduced in (1.4) is a Hilbert basis of the space L2 ([0, 2π]). We record
this in:
Theorem 2.1. The set S introduced in (1.4) is a Hilbert basis of the Hilbert space
L ([0, 2π]). In particular, the space L2 ([0, 2π]) is separable (and Theorem 6.3 - Chapter 3
2

holds).
Proof. Straightforward computations yield that,
Z 2π
|φn (x)|2 dx = 1 (n = 0, 1, 2 · · · )
0
Z 2π
|ψn (x)|2 dx = 1 (n = 1, 2, · · · )
0
Z 2π Z 2π Z 2π
φn (x)φm (x) dx = ψn (x)ψm (x) dx = φn (x)ψm (x) = 0 (if n 6= m)
0 0 0
Z 2π
φn (x)ψn (x) dx = 0 n = 0, 1, 2, · · · ) .
0

Hence S is orthonormal. Furthermore, by Corollary 1.4, span(S) is dense in L2 ([0, 2π]).


In light of Theorem 6.3 (Chapter 3), it is natural to introduce:
Definition 2.2. [Fourier series]
Let f : [0, 2π] → R be a Riemann integrable function (or a square integrable
function in L2 ([0, 2π])). The Fourier coefficients of f are
Z 2π
1 2π
Z Z 2π
1 1
a0 = f (x) dx , an = f (x) cos(nx) dx , bn = f (x) sin(nx) dx ,
2π 0 π 0 2π 0
where n ∈ {1, 2, · · · }.
The Fourier series of f is
∞ 
X 
a0 + an cos(nx) + bn sin(nx) .
n=1

It will be denoted by FS(f (x)).

72 4. FOURIER SERIES

Remark 2.3.
(1) To stress their dependence on the function, the Fourier coefficients of the function
f are sometimes denoted by an (f ) and bn (f ).
(2) [Even and odd functions]
If the function f is even, a simple symmetry argument yields the following
formulas for the Fourier coefficients of f :
1 π 2 π
Z Z
a0 = f (x) dx , an = f (x) cos(nx) dx , bn = 0 (n ≥ 1) .
π 0 π 0
Similarly, if f is an odd function, the Fourier coefficients of f are given by
2 π
Z
a0 = an = 0 , bn = f (x) sin(nx) dx (n ≥ 1) .
π 0
(3) Using the functions {φn , ψn } introduced in (1.4), we may express the Fourier
series of a function f as follows
1 1 1
a0 = √ hf, φ0 i , an = √ hf, φn i , bn = √ hf, ψn i ,
2π π π
where n ∈ {1, 2, · · · }. Furthermore, the Fourier series of f can be expressed as
follows (compare with Theorem 6.3 - Chapter 3):
∞ 
X 
hf, φ0 iφ0 + hf, φn iφn + hf, ψn iψn .
n=1

(4) The othonormal set S = {φn , ψn } introduced in (1.4) can be expressed using a
single sequence of functions (fn )n≥0 , where f0 = φ0 and for all n ≥ 1,
(
φ(n+1)/2 if n is odd
fn = ,
ψn/2 if n is even
so that f1 = φ1 , f2 = ψ2 , f3 = φ2 , f4 = ψ2 , etc.
The Fourier series of f becomes
∞
X
hf, fn ifn .
n=0

Using Remark 2.3 and Theorem 2.1, we infer from Chapter 3 - Theorem 6.3:
Theorem 2.4. [Convergence of Fourier series in L2 -norm]
If f ∈ L2 ([0, 2π]), then
(1) [Fourier series decomposition]
The Fourier series of f is convergent in (L2 ([0, 2π]), k · k2 ) and
∞ 
X 
f (x) = a0 + an cos(nx) + bn sin(nx) in L2 ([0, 2π]) .
n=0

(2) [Parseval’s identity]

∞ 2π
a20 X  2  1 Z
2
+ an + b n = |f (x)|2 dx .
2 n=1
π 0
 3. UNIFORM CONVERGENCE OF FOURIER SERIES 73

3. Uniform convergence of Fourier series

In this section, we will discuss the convergence of Fourier series relative to the norm
k · k∞ rather than the L2 -norm (see (1.2)). In light of (1.3), convergence relative to k · k∞
(uniform convergence) is stronger than that relative to the L2 -norm. Hence, for the sake
of obtaining the uniform convergence of the Fourier series, we will need to work with
functions having a fair amount of regularity, that is functions that are not simply square
integrable (as in Theorem 2.4).
Let us recall the notion of left- and right- continuity/differentiability of a function of
a real variable. Consider a function f : [a, b] → R and a point c ∈ (a, b). We define the
left-limit (respectively right-limit) of f at c (if they exist) to be
 
f (c− ) = lim f (x) respectively f (c+ ) = lim f (x) .
x→c− x→c+

Similarly, we say that f is left-differentiable (respectively right-differentiable) at c if the

following limit exist
f (x) − f (c− )  f (x) − f (c+ ) 
f 0 (c− ) = lim respectively f 0 (c+ ) = lim .
x→c− x−a x→c+ x−c
The values f (c− ) and f (c+ ) are the left- and right-values of f at c, respectively. Similarly,
f 0 (c− ) and f 0 (c+ ) are called the left- and right-derivatives of f at c, respectively. When
c = a we can talk about f (c+ ) and f 0 (c+ ), while for c = b, we can talk about f (c− ) and
f 0 (c− ).
Definition 3.1. [Piecewise smooth functions]
A function f : [a, b] → R is said to be piecewise C 1 on [a, b] if
(1) There exists a finite subset {x0 , x1 , · · · , xn } ⊂ [a, b] such that x0 = a, xn = b and
x0 < x1 < · · · < xn ;
(2) For all i ∈ {0, 1, · · · , n − 1}, f is differenitable on ]xi , xi+1 [ ;
(3) For all i ∈ {0, 1, · · · , n − 1}, f is left-differentiable at xi and right-differentiable
at xi+1 .
Remark 3.2. Suppose that f is a piecewise C 1 function as in Definition 3.1. We
would like to mention the following useful properties satisfied by f .
(1) According to Definition 3.1, the right-value and the right-derivative of f exist at
every xi . Also, the left-value and the left-derivative of f exist at every xi+1 .
(2) On every interval [xi , xi+1 ], the function f admits an extension fi : [xi , xi+1 ] → R
defined as follows

f (x)

 if xi < x < xi+1 ,
fi (x) = f (xi )+ if x = xi ,
f (x )
if x = x .

i+1 − i+1

Furthermore, the function fi is differentiable on the interval [xi , xi+1 ].

(3) We denote by f 0 the function on [a, b] defined as follows (we may call f 0 the
generalized derivative of f ):

df
 dx (x)
if x ∈ [a, b] \ {x0 , x1 , · · · , xn−1 } ,

f 0 (x) = f 0 (xi )+ if x ∈ {x0 , x1 , · · · , xn−1 } ,
f 0 (x )
if x = x = b .

n − n
74 4. FOURIER SERIES

(4) The functions f and f 0 are Riemann integrable on [a, b] (hence square integrable,
i.e. in L2 ([a, b])) and
Z b n−1 Z xi+1
X Z b n−1 Z xi+1
X
0
f (x) dx = fi (x) dx , f (x) dx = fi0 (x) dx .
a i=0 xi a i=0 xi

Next, we will write a criterion for the convergence of a series relative to the norm of
uniform convergence.
Lemma 3.3. [Weirstrass M -test] Let (un ) be a sequence in C([0, 2π]). Suppose that
• There exists a sequence (an ) ⊂ R+ such that, ∀ n ≥ 1, kun k∞ ≤ an ;
X∞
• The series an is convergent .
n=1
∞
X
Then, the series un is convergent in (C([0, 2π]), k · k∞ ), i.e. the sequence of partial
n=1
sums
n
X
sn = uk
k=1
is a convergent sequence in (C([0, 2π]), k · k∞ ).
The proof of the lemma is left as an exercise. All we have to do is prove that the
sequence of partial sums (sn ) is a Cauchy sequence. The convergence then follows by the
completeness of the space (C([0, 2π]), k · k∞ ) (see Chapter 1 - Theorem 7.4).

Now, we can prove the main theorem of this section (compare with Theorem 2.4):
Theorem 3.4. [Uniform convergence of Fourier series]
Let f : [0, 2π] → R be a continuous function such that
• f (0) = f (2π) ;
• f is piecewise C 1 on [0, 2π] .
X∞  
Then, the Fourier series of f , a0 + an cos(nx) + bn sin(nx) , uniformly converges to
n=1
f on [0, 2π] (i.e. it is a convergent series with limit f in (C([0, 2π]), k · k∞ )).
Proof. The proof is quite simple. We will denote by (an (g), bn (g)) the Fourier coef-
ficients of a Riemann integrable function g. With this notation in hand, we check by a
straightforward calculation, using in particular integration by parts and the assumption
that f (0) = f (2π), that, for all n ≥ 1,
1 1
an (f ) = − bn (f 0 ) and bn (f ) = an (f 0 ) .
n n
Next we use Cauchy’s inequality, for all s, t ∈ R, |st| ≤ 21 (s2 + t2 ), which follows by
expanding the square in (s ± t)2 ≥ 0. That way we obtain,
   
1 1 0 2 1 1 0 2
|an (f )| ≤ + |bn (f )| and |bn (f )| ≤ + |an (f )| .
2 n2 2 n2
Thus, the general term of the Fourier series of f satisfies
1 1
|an (f 0 )|2 + |bn (f 0 )|2 ,


|an cos(nx) + bn sin(nx)| ≤ 2 +
n 2
 4. THE DIRICHELT THEOREM 75

1
P
where n≥1 n2 is a convergent sequence in R, and the same holds for the series
0 2 0 2
P
n≥1 (|an (f )| + |bn (f )| ), in light of Theorem
2.4 (Parseval’s identity). Now, Lemma 3.3
yields that the Fourier series of f is a uniformly convergent series.
That the limit of the Fourier series is equal to f follows from Theorem 2.4 (the limit
is equal to f relative to the L2 -norm) and the inequality in (1.2). 

4. The Dirichelt theorem

In this section we discuss the pointwise convergence of Fourier series, i.e. for a function
f (x), we wish to find the value of the Fourier series FS(f (x)) at every point x. The
answer to this question is the content of the Dirichlet theorem and is valid for periodic
functions, a notion that we will define next.
Definition 4.1. Let T > 0. A function f : R → R is said to be periodic with period
T (or simply, T -periodic) if
∀ x ∈ R, f (x + T ) = f (x) .
Remark 4.2. Suppose that f is a T -periodic function. The function f inherits many
properties from its restriction on the interval [−T /2, T /2]. In particular:
(1) If f is (Riemann) integrable on [−T /2, T /2], then f is Riemann integrable on
every interval [a, b] ;
(2) If f is piecewise C 1 on [−T /2, T /2], then for all x ∈ R, the limits f (x± ) and
f 0 (x± ) exist in R.
(3) If f is continuous on [−T /2, T /2], then f is continuous on R\{± T2 , ±3 T2 , ±5 T2 , · · · } ;
(4) If f is continuous on [−T /2, T /2] and f (−T /2) = f (T /2), then f is continuous
on R.
Now we can state the Dirichlet theorem for 2π-periodic functions:
Theorem 4.3. [Dirichlet Theorem]
Suppose that f is a 2π-periodic function. If furthermore f is piecewise C 1 on [−π, π],
then, the Fourier series of f ,
∞ 
X 
a0 + an cos(nx) + bn sin(nx)
n=1
is convergent for all x ∈ R and its value is equal to
f (x+ ) + f (x− )
.
2
Remark 4.4. If the function f in Theorem 4.3 is furthermore continuous, we get
that the Fourier series of f is equal to f pointwise, i.e.
X∞  
∀ x ∈ R , a0 + an cos(nx) + bn sin(nx) = f (x) .
n=1
This conclusion agrees with that of Theorem 3.4, which in fact states that the above
equality holds uniformly, i.e. with respect to the norm of uniform convergence.
The proof of Theorem 4.3 is quite lengthy. We will split it into several lemmas.
Lemma 4.5. [Jordan’s inequality]
π 2θ
If 0 ≤ θ ≤ , then ≤ sin θ ≤ θ.
2 π
76 4. FOURIER SERIES

Proof. Define the function h(θ) = sin θ − θ. Clearly, h0 (θ) = cos θ − 1 ≤ 0, hence h
is decreasing. Consequently, for all θ ∈ [0, π/2], h(0) ≤ h(θ) ≤ h(π/2) with h(0) = 0. In
particular, this yields that sin θ ≤ θ.
Consider the function g(θ) = sin θ − 2θ
π
. It is easy to check that max0≤θ≤ π2 g(θ) = g(θ0 )
where θ0 = arccos π , and min0≤θ≤ π2 g(θ) = 0. Thus, for all θ ∈ [0, π2 ], we have g(θ) ≥ 0,
2

i.e. sin θ ≥ π2 θ. 
Lemma 4.6. If g is a 2π-periodic function and Riemann integrable on [−π, π], then,
for all a ∈ R,
Z a+2π Z π
g(x) dx = g(x) dx .
a −π


Proof. Note that, for all u ∈ R, g(u) = g (u − 2π) + 2π = g(u − 2π), since g is
2π-periodic. Thus, the change of variable u = x + 2π yields
Z −π Z π Z a+2π Z a+2π
g(x) dx = g(u − 2π) du = − g(u) du = − g(x) dx .
a a+2π π π

Thus, we get
Z a+2π Z −π Z π Z a+2π Z π
g(x) dx = g(x) dx + g(x) dx + g(x) dx = g(x) dx .
a a −π π −π

Lemma 4.7. If g : [a, b] → R is a piecewise C 1 function, then
Z b  
1
lim g(x) sin n + x dx = 0 .
n→∞ a 2
Proof. There exist points x0 < x1 < · · · < xm such that x0 = a, xm = b and
for all i ∈ {1, · · · , m}, g can be viewed as a C 1 function on the interval [xi−1 , xi ] (see
Remark 3.2). Moreover,
Z b   m Z xi  
1 X 1
g(x) sin n + x dx = g(x) sin n + x dx .
a 2 k=1 x i−1
2
Z xi  
1
Let i ∈ {1, · · · , m} be fixed. We will prove that lim g(x) sin n + x dx = 0.
n→∞ x
i−1
2
Indeed, performing an integration by parts, we have
     
Z xi 
1   Z xi cos n + 1
2
x g(x) cos n + 1
2
x xi
0
g(x) sin n + x dx = g (x) dx − .
xi−1 2 xi−1 n + 12 n + 12
xi−1

Clearly, the terms on the right hand side tend to 0 as n → ∞. 

Lemma 4.8. Let f be a 2π-periodic function and x ∈ [−π, π] be a fixed point. If f is
piecewise C 1 on [−π, π], then there exist two constants δ > 0 and M > 0 such that
∀ t ∈ (0, δ), |f (x + t) − f (x+ )| ≤ M t and |f (x − t) − f (x− )| ≤ M t .
Proof. Let x ∈ [−π, π] be fixed. The following two limits exist (see Remark 4.2)
f (x + t) − f (x) f (x − t) − f (x)
f 0 (x+ ) = lim and f 0 (x− ) = lim .
t→0+ t t→0+ −t
 4. THE DIRICHELT THEOREM 77

The definition of the limit yields the existence of δ > 0 such that,

f (x + t) − f (x) f (x − t) − f (x)
∀ t ∈ (0, δ) , − f 0 (x+ ) ≤ 1 and − f 0 (x− ) ≤ 1 .
t −t

Consequently, we get the conclusion of Lemma 4.8 with M = 1 + max |f 0 (x+ )|, |f 0 (x− )| .




Lemma 4.9. [Dirichlet Kernel]

For all n ≥ 1 and t ∈ R, define the Dirichlet kernel of order n:

n
1 X
Dn (t) = + cos(kt) .
2 k=1

Then, it holds the following:

Z π
2
(1) Dn (u) du = 1 ;
π 0
2n + 1
(2) If t ∈ {0, ±2π, ±4π, ±6π · · · }, then Dn (t) = ;
2
sin 2n+1


2
t
(3) If t 6∈ {0, ±2π, ±4π, ±6π · · · }, then Dn (t) = ;
2 sin 2t
Z δ
π
(4) For all δ ∈ ]0, 2 [ and n ≥ 1, |tDn (t)| dt ≤ πδ .
0
Proof.
Proof of (1):Z
π
For all k ∈ Z, cos(kt) dt = 0. Thus,
0

Z π Z π n Z π
1 X π
Dn (t) dt = dt + cos(kt) dt = .
0 0 2 k=1 0
2

Proof of (2):
n
1 X 1
For t ∈ {0, ±2π, ±4π, ±6π · · · }, Dn (t) = + 1 = + n.
2 k=1 2
Proof of (3):
We will use the Euler formula: For all θ ∈ R, eiθ = cos θ + i sin θ (consequently,
cos θ = Re(eiθ )).
78 4. FOURIER SERIES

Thus, we may write

n
!
1 X ikt
Dn (t) = Re + e
2 k=1
1 1 − ei(n+1)t
 
= Re − +
2 1 − eit


1 1 − cos t − cos (n + 1)t + cos(nt)
=− +
2 2 − 2 cos t
− cos n + 12 t + 2θ + cos n + 21 t − 2t





=
2 − 2 cos 2 2t

2 sin n + 21 t sin 2t sin n + 12 t

= = .
4 sin2 2t 2 sin 2t

Proof of (4):
Let δ ∈ ]0, π/2[ be fixed. Since the function t 7→ tDn (t) is continuous, the Fundamental
Theorem of Calculus yields
Z δ Z δ
|tDn (t)| = lim |tDn (t)| dt .
0 →0+ 
We use the Jordan inequality (Lemma 4.5) and the identity in (3) above to write,
  
1
∀  ∈ ]0, π/2[ , ∀ t ∈ ], π/2[ , |tDn (t)| ≤ π sin n+ t ≤ π.
2
Consequently,
Z δ Z δ
lim |tDn (t)| dt ≤ lim π dt = πδ .
→0+  →0+ 

Lemma 4.10. Let f be a 2π-periodic function. Assume that f is Riemann integrable
on the interval [−π, π]. Then, for all x ∈ R,
Z π Z π  
Dn (x − u)f (u) du = Dn (t) f (x − t) + f (x + t) dt ,
−π 0

where Dn (·) is the Dirichlet kernel of order n (see Lemma 4.9).

Proof. The change of variable t = x − u yields
Z π Z x+π
Dn (x − u)f (u) du = Dn (t)f (x − t) dt .
−π x−π
The last integral can be written in the form
Z x+π Z a+2π
Dn (t)f (x − t) dt = Dn (t)f (x − t) dt where a = x − π
x−π a
Z π
= Dn (t)f (x − t) dt by Lemma 4.6,
−π

since the function t 7→ Dn (t)f (x − t) is 2π-periodic. Therefore, we have the identity

Z π Z π
Dn (x − u)f (u) du = Dn (t)f (x − t) dt . (4.1)
−π −π
 4. THE DIRICHELT THEOREM 79

In a similar fashion, the change of variable t = x + u yields

Z π Z π
Dn (x − u)f (u) du = Dn (t)f (x + t) dt . (4.2)
−π −π

Adding the two identities in (4.1) and (4.2), we get

Z π
1 π 1 π
Z   Z
Dn (x − u)f (u) du = Dn (t) f (x + t) + f (x − t) dt = g(t) dt
−π 2 −π 2 −π
  Z π
where g(t) = Dn (t) f (x + t) + f (x − t) is an even function, hence g(t) dt =
Z π −π

2 g(t) dt. That way we get the identity in Lemma 4.10. 

0

Lemma 4.11. Let f be a 2π-periodic function and assume that f is piecewise C 1 on

the interval [−π, π]. Let {a0 , an , bn } be the Fourier coefficients of f . Then
X n   1Z π  
a0 + ak cos(kx) + bk sin(kx) = Dn (t) f (x + t) + f (x − t) dt ,
k=1
π 0
where Dn (·) is the Dirichlet kernel of order n (see Lemma 4.9).
Z π
1
Proof. We have a0 = f (u) du and, for all k ≥ 1,
2π −π
1 π
Z 
ak cos(kx) + bk sin(kx) = cos(ku) cos(kx) + sin(ku) sin(kx) f (u) du
π −π
1 π
Z


= cos k(x − u) f (u) du .
π −π
Consequently, using the definition of the Dirichlet kernel in Lemma 4.9, we get the fol-
lowing identity
Xn   1Z π
a0 + ak cos(kx) + bk sin(kx) = Dn (x − u)f (u) dt .
k=1
π −π

Now, applying Lemma 4.10 finishes the proof of Lemma 4.11. 

Proof of Theorem 4.3.

Let x ∈ [−π, π] be fixed. In light of Lemma 4.11, it is sufficient to prove that
 Z π 
1   f (x+ ) + f (x− )
lim Dn (t) f (x + t) + f (x − t) dt − = 0, (4.3)
n→∞ π 0 2
where Dn is the Dirichlet kernel of order n (see Lemma 4.10).
In light of Lemma 4.9, we write, for all n ≥ 1,
1 π
Z   f (x+ ) + f (x− )
Dn (t) f (x + t) + f (x − t) dt −
π 0 2
Z π
1  
= Dn (t) f (x + t) + f (x − t) − f (x+ ) − f (x− ) dt
π 0
Z π
1 π
Z
1    
= Dn (t) f (x + t) − f (x+ ) dt + Dn (t) f (x − t) − f (x− ) dt .
π 0 π 0
80 4. FOURIER SERIES
Z π  
We will prove that lim Dn (t) f (x ± t) − f (x± ) dt = 0. Let  > 0 be fixed. Choose
n→∞ 0
δ > 0 such that Lemma 4.8 holds. Choose δ0 > 0 such that δ0 < min(δ, π2 ) and M πδ0 < 2 ,
where M is the constant in Lemma 4.8.
By Lemma 4.7, we may choose n0 ≥ 1 such that, for all n ≥ n0 ,
Z π   
f (x + t) − f (x+ ) 1 
t sin n+ t dt < . (4.4)
δ0 2 sin 2 2 2
Now we will prove that, for all n ≥ n0 ,
Z π  
Dn (t) f (x + t) − f (x+ ) dt <  .
0
Indeed, for n ≥ n0 ,
• the expression for Dn (t) in Lemma 4.9 and the inequality in (4.4) yield
Z π   
Dn (t) f (x + t) − f (x+ ) dt < ;
δ0 2
• Lemma 4.8, our choice of δ0 and Lemma 4.9 yield
Z δ0 Z δ0
  
Dn (t) f (x + t) − f (x+ ) dt ≤ M |tDn (t)| dt ≤ M πδ0 < ;
0 0 2
• the triangle inequality now yields
Z π  
Dn (t) f (x + t) − f (x+ ) dt
0
Z δ0   Z π  
= Dn (t) f (x + t) − f (x+ ) dt + Dn (t) f (x + t) − f (x+ ) dt
0 δ0
Z δ0   Z π  
≤ Dn (t) f (x + t) − f (x+ ) dt + Dn (t) f (x + t) − f (x+ ) dt
0 δ0
 
< + = .
2 2
Z π  
This proves that lim Dn (t) f (x + t) − f (x+ ) dt = 0. In a similar fashion we prove
n→∞ 0
Z π  
that lim Dn (t) f (x − t) − f (x− ) dt = 0 
n→∞ 0