Notes on Elementary Linear

Analysis
Erik Bédos

Institute of Mathematics, University of Oslo

Version 1.1

May 23, 2019

 Acknowledgements

The first draft of these notes were written during the spring of 2018 and tested
simultaneously as a part of the curriculum for the course MAT3400/4400
at the University of Oslo. Hearthy thanks to Adam Sørensen, who was in
charge of the lectures, for his patience and for his comments.
I have tested myself the notes during the spring of 2019 and used this
opportunity to perform some minor revisions. I am grateful to Ulrik Enstad
and to some of the students following the course for pointing out some
misprints in the text. If you find others, or if you have some suggestions,
I’ll be happy if you send me a message.

E.B.
 Contents

Acknowledgements i

Contents ii

1 Preliminaries 1
1.1 Normed spaces . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Inner product spaces . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Linear operators . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 On Lp -spaces 9
2.1 The case 1 ≤ p < ∞ . . . . . . . . . . . . . . . . . . . . . . 9
2.2 The case p = ∞ . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

3 More on normed spaces and linear operators 21

3.1 Aspects of finite dimensionality . . . . . . . . . . . . . . . . 21
3.2 Direct sums and projections . . . . . . . . . . . . . . . . . . 25
3.3 Extension by density and continuity . . . . . . . . . . . . . . 32
3.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

4 More on Hilbert spaces 41

4.1 Geometry in Hilbert spaces . . . . . . . . . . . . . . . . . . 41
4.2 Orthonormal bases in Hilbert spaces . . . . . . . . . . . . . 47
4.3 Adjoint operators . . . . . . . . . . . . . . . . . . . . . . . . 55
4.4 Self-adjoint operators . . . . . . . . . . . . . . . . . . . . . . 62
4.5 Unitary operators . . . . . . . . . . . . . . . . . . . . . . . . 66
4.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

5 On compact operators 79
 Contents

5.1 Introduction to compact operators between normed spaces . 79

5.2 On compact operators on Hilbert spaces . . . . . . . . . . . 83
5.3 The spectral theorem for a compact self-adjoint operator . . 90
5.4 Application: The Fredholm Alternative . . . . . . . . . . . . 96
5.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

iii
 CHAPTER 1

Preliminaries

In this chapter we fix some notation and give a review of some of the concepts
and results that we will need. These are usually covered in undergraduate
courses in real analysis, and the reader may consult the book of T. Lindstrøm,
Spaces: an introduction to real analysis (AMS 2017), or any other standard
book in real analysis, for details and proofs.

1.1 Normed spaces

Throughout these notes F will denote either R (the real numbers) or C (the
complex numbers). If X, Y are sets, we let X × Y denote their Cartesian
product, i.e., n o
X × Y = (x, y) : x ∈ X, y ∈ Y .
A metric space (X, d) is called complete when every Cauchy sequence in
(X, d) is convergent.
Definition 1.1.1. A normed space (X, k · k) over F is a vector space X over
F which is equipped with a norm k · k. We recall that X is then a metric
space with respect to the metric given by d(x, y) = kx − yk for x, y ∈ X.
We will only consider normed spaces over F in these notes, and we will often
just write X to denote such a normed space, assuming tacitly that some
norm on X is given.
When x ∈ X and r > 0, we let BrX (x) denote the closed ball in X with
center in x and radius r, that is,
n o
BrX (x) := y ∈ X : kx − yk ≤ r .
When there is no danger of confusion, we just write Br (x) instead of BrX (x).
We also set
n o
X1 := B1X (0), i.e., X1 = x ∈ X : kxk ≤ 1 .
1. Preliminaries

Definition 1.1.2. If (X, k · k) is a normed space, and k · k0 is also a norm

on X, we say that k · k and k · k0 are equivalent when there exist positive
real numbers K and L such that

kxk0 ≤ K kxk and kxk ≤ L kxk0 for all x ∈ X.

When k · k and k · k0 are equivalent, it is clear that a sequence {xn }∞

n=1
in X converges to x ∈ X w.r.t. k · k if and only if it converges to x ∈ X
w.r.t. k · k0 . The following proposition implies that for many purposes the
choice of a norm in a finite-dimensional space can be made arbitrarily.

Proposition 1.1.3. If X is a finite-dimensional vector space over F, then

all norms on X are equivalent.

Definition 1.1.4. Assume {xn }∞ n=1 is a sequence in a normed space (X, k·k).
We say that the series ∞ n=1 n is convergent in X if there is some x∈X
P
x
such that kx − n=1 xn k → 0 as N → ∞, in which case we say that n=1 xn
PN P∞

converges to x (w.r.t. k · k), and also write x = ∞ n=1 xn .

P

Definition 1.1.5. When a normed space (X, k · k) is complete with respect

to the associated metric given by

d(x, y) = kx − yk

for all x, y ∈ X, we say that X is a Banach space (over F).

To check that a normed space is a Banach space, the following result is

often useful:

Theorem 1.1.6. Let (X, k · k) be a normed space. Then X is a Banach

space if and only if every absolutely convergent series in X is convergent in
X, that is, if and only if the following condition holds :
Whenever ∞ n=1 xn is a series in X such that kxn k < ∞, then
P P∞
n=1
n=1 xn is convergent in X.
P∞

Remark 1.1.7. It is good to know that if X is a normed space, then we

can always form its completion; this means that whenever needed, we can
assume that X sits as a dense subspace of a Banach space X f where the
norm of Xf extends the norm on X. An elegant way to construct X f (as
an application of the so-called Hahn-Banach theorem) is covered in more
advanced courses on linear analysis.

2
 1.2. Inner product spaces

1.2 Inner product spaces

Definition 1.2.1. An inner product space over F is a vector space X over
F which is equipped with an inner product h·, ·i : X × X → F. This means
that for x, y, z ∈ X and λ ∈ F we have:
i) hx + y, zi = hx, zi + hy, zi,
ii) hλx, yi = λhx, yi,
iii) hy, xi = hx, yi,
iv) hx, xi ≥ 0,
v) hx, xi = 0 if and only if x = 0.
Remark 1.2.2. a) Properties i) and ii) say that the inner product is linear
in the first variable.
b) When F = R, property iii) says that the inner product is symmetric,
i.e., hy, xi = hx, yi; combining i) and ii) with iii), we then get that the inner
product is also linear in the second variable.
c) When F = C, we get that the inner product is conjugate-linear in the
second variable; this means that we have
hx, y + zi = hx, yi + hx, zi and hx, λyi = λhx, yi.
Some authors prefer to use inner products that are linear in the second
variable and conjugate-linear in the first variable. This is common in
textbooks related to physics or mathematical physics. As one can go from
one type to the other by setting hx, yi0 := hy, xi, it is mainly a matter of
taste which convention one chooses to use.
In the sequel, by an inner product space, we will always mean an inner
product space over F. An inequality of fundamental importance is:
Theorem 1.2.3 (The Cauchy-Schwarz inequality). Let X be an inner prod-
uct space. For x ∈ X set kxk := hx, xi1/2 . Then we have
|hx, yi| ≤ kxk kyk (1.2.1)
for all x, y ∈ X, with equality if and only if x and y are linearly dependent.
If X is an inner product space, then using the Cauchy-Schwarz inequality,
one deduces that kxk = hx, xi1/2 gives a norm on X. Thus, X is then a
normed space, and its norm is easily seen to satisfy the parallellogram law,
that is, for all x, y ∈ X we have
kx + yk2 + kx − yk2 = 2 kxk2 + 2 kyk2 . (1.2.2)

3
1. Preliminaries

Definition 1.2.4. Let X be an inner product space. If x, y ∈ X, then x

and y are said to be orthogonal (to each other) when hx, yi = 0. A subset
S ⊂ X is called orthogonal if x and y are orthogonal for all x, y ∈ S such
that x =6 y. Moreover, S is called orthonormal if S is orthogonal and kxk = 1
for all x ∈ S.

Proposition 1.2.5 (Pythagoras). Assume {x1 , . . . , xn } is a finite orthogo-

nal subset of an inner product space X. Then we have

kx1 + · · · + xn k2 = kx1 k2 + · · · + kxn k2 .

Proposition 1.2.6. Assume S = {u1 , . . . , un } is a finite orthonormal subset

of an inner product space X. Then S is linearly independent. Moreover, if
u ∈ Span{u1 , . . . , un }, i.e., if u is a linear combination of the vectors in S,
then
n n
u= and kuk2 = |hu, uj i|2 .
X X
hu, uj i uj
j=1 j=1

Proposition 1.2.7 (Bessel’s inequality). Assume S = {uj : j ∈ J} is a

countable orthonormal subset of an inner product space X. Then for any
x ∈ X we have
|hx, uj i|2 ≤ kxk2 .
X

j∈J

Definition 1.2.8. An inner product space X (over F) is called an Hilbert

space (over F) when X is complete with respect to the norm associated with
its inner product.

Remark 1.2.9. Assume X is an inner product space. Considering X as a

normed space, we may form its completion X f (cf. Remark 1.1.7), and extend
the inner product on X to an inner product on X f as follows: if y, y 0 ∈ X,
f
then we can pick sequences {xn }∞ n=1 , {xn }n=1 in X converging respectively
0 ∞

to y and y 0 ; after checking that {hxn , x0n i}∞

n=1 is a Cauchy sequence in F,
hence is convergent, we may set

hy, y 0 i := lim hxn , x0n i.

n→∞

It is then a somewhat tedious exercise to verify that this gives a well-defined

inner product on Xf which extends the one on X. This means that whenever
needed, we may assume that X sits as a dense subspace of a Hilbert space
f (called the completion of X) where the inner product on X
X f extends the
inner product on X.

4
 1.3. Linear operators

1.3 Linear operators

Definition 1.3.1. Assume that X and Y are both vectors spaces over F.
Then a map T : X → Y is called a linear operator if we have

T (λ1 x1 + λ2 x2 ) = λ1 T (x1 ) + λ2 T (x2 )

for all λ1 , λ2 ∈ F and all x1 , x2 ∈ X.

We denote by L(X, Y ) the set of all linear operators from X to Y . One

readily checks that L(X, Y ) is a vector space over F with respect to the
operations defined by

(S + T )(x) = S(x) + T (x), (λT )(x) = λ T (x)

for S, T ∈ L(X, Y ), λ ∈ F and x ∈ X. We also set L(X) := L(X, X). We

let IX ∈ L(X) denote the identity map from X into itself, that is, IX (x) = x
for all x ∈ X. We just I instead of IX if no confusion is possible.

Definition 1.3.2. Assume that X and Y are both normed spaces over F.
Then a linear operator T : X → Y is called bounded if there exists some
real number M > 0 such that

kT (x)k ≤ M kxk ∀ x ∈ X,

or, equivalently, such that kT (x)k ≤ M for all x ∈ X1 .

Proposition 1.3.3. Assume that X and Y are both normed spaces over F
and let T ∈ L(X, Y ). Then the following conditions are equivalent:

(a) T is bounded.

(b) T is uniformly continuous on X.

(c) T is continuous on X.

(d) T is continuous at x = 0.

We denote the set of all bounded linear operators from X to Y by B(X, Y ).

We follow tradition here and use the qualifying adjective “bounded” , although
we could equally well have used “continuous” instead. One readily checks
that B(X, Y ) is a subspace of L(X, Y ). We also set B(X) = B(X, X).

5
1. Preliminaries

Proposition 1.3.4. Assume that X and Y are both normed spaces over F.
For T ∈ B(X, Y ), set
n o
kT k := sup kT (x)k : x ∈ X1 < ∞.
Then the map T → kT k is a norm on B(X, Y ), called the operator norm.
Moreover, we have
n o
kT k = sup kT (x)k : x ∈ X, kxk = 1 (when X 6= {0}),
and
kT (x)k ≤ kT k kxk ∀ x ∈ X.
Theorem 1.3.5. Assume that X is a normed space over F, while Y is a
Banach space. Then B(X, Y ) is Banach space. In particular, B(X) is a
Banach space whenever X is a Banach space.
An immediate consequence of this theorem is that B(X, F) is a Banach
space whenever X is normed space over F. Elements of L(X, F) are called
linear functionals. Thus B(X, F) consists of the bounded linear functionals
on X; it is usually called the dual space of X and denoted by X ∗ in many
books, or by X ] in others.
Definition 1.3.6. A map T : X → Y between two vector spaces over F is
called a (vector space) isomorphism if T ∈ L(X, Y ) and T is bijective (that
is, T is both one-to-one and onto). It is then easy to check that the inverse
map of T , T −1 : Y → X, is linear, i.e., T −1 ∈ L(Y, X).
Definition 1.3.7. Assume that X and Y are normed spaces over F. A map
T : X → Y is called an isomorphism of normed spaces if T is a (vector
space) isomorphism such that both T and T −1 are bounded.
Definition 1.3.8. Assume that X is a normed space and T ∈ B(X). Then
we say that T is invertible in B(X) if T is an isomorphism of normed spaces.
In other words, an operator T ∈ B(X) is invertible in B(X) if T is bijective
and T −1 ∈ B(X).
Proposition 1.3.9. Let X, Y, Z be normed spaces over F, and let T ∈
B(X, Y ), S ∈ B(Y, Z). Set ST := S ◦ T : X → Z. Then ST ∈ B(X, Z)
and
kST k ≤ kSk kT k .
Corollary 1.3.10. Assume that X is a normed space and S ∈ B(X). For
each n ∈ N, let S n := S · · · S denote the product of S with itself n times,
then S n ∈ B(X) and kS n k ≤ kSkn . Note that by setting S 0 = IX , this
formula also holds when n = 0.

6
 1.3. Linear operators

Theorem 1.3.11. Assume that X is a Banach space and S ∈ B(X) is such

that kSk < 1. Then I − S is invertible in B(X) and
∞
(I − S)−1 = Sn (convergence w.r.t. operator norm).
X

n=0

Moreover, k(I − S)−1 k ≤ 1

1−kSk
.

7
 CHAPTER 2

On Lp-spaces

An important class of Banach spaces over F associated with measure spaces

are the so-called Lp -spaces, where 1 ≤ p ≤ ∞. We will assume that F = C,
and just mention that the case where F = R may be handled in a similar
way. Our presentation is somewhat more detailed than the one given in
section 7.7 and 7.9 of Lindstrøm’s book.

2.1 The case 1 ≤ p < ∞

Let (X, A, µ) be a measure space and set
n o
M = M(X, A) := f : X → C : f is A-measurable ,

which we know is a vector space (with its natural operations). We will

be interested in subspaces of M associated with any p ∈ [1, ∞]. We first
consider the case 1 ≤ p < ∞. For each f ∈ M we note that the function |f |p
is non-negative and belongs to M (since the function z → |z|p is continuous
on C), so we can set
Z 1/p
kf kp := p
|f | dµ ∈ [0, ∞]
X

(using the convention that ∞1/p = ∞). Moreover, we set

Lp (X, A, µ) := {f ∈ M : kf kp < ∞}.

We will just write Lp when there is no danger of confusion, and note that
some authors write Lp (µ). It is then clear that L1 consists of all the complex
functions on X which are integrable (w.r.t. µ). When A = P(X) and
µ is the counting measure on A, it is common to write `p (X) instead of
Lp (X, A, µ).
2. On Lp -spaces

It is not difficult to see that Lp is a subspace of M. For example,

closedness under addition follows readily from the inequality |z + w|p ≤
2p (|z|p + |w|p ), which is easily seen to hold for all z, w ∈ C. On the other
hand, it is not true in general that k · kp is a norm on Lp . The reason is that
for f ∈ Lp , we have
Z
kf kp = 0 ⇔ |f |p dµ = 0 ⇔ |f |p = 0 µ-a.e. ⇔ f = 0 µ-a.e.
X

As we will soon see, k · kp is a seminorm on Lp in the following sense:

Definition 2.1.1. A seminorm on a vector space V (over F) is a function

v → kvk from V into [0, ∞) satisfying kλ vk = |λ| kvk and the triangle
inequality kv + wk ≤ kvk + kwk for all v, w ∈ V and λ ∈ F.

We note that a seminorm k · k is a norm if it also satisfies that kvk = 0

only if v = 0. Using the triangle inequality for | · | on C, one readily deduces
that k · k1 gives a seminorm on L1 . To handle the case p > 1 we will need:

Theorem 2.1.2 (Hölder’s inequality). Assume p ∈ (1, ∞) and let q ∈ (1, ∞)

denote p’s conjugate exponent given by q = p−1
p
, so that p1 + 1q = 1.
Let f ∈ Lp and g ∈ Lq . Then f g ∈ L1 and
Z
kf gk1 = |f g| dµ ≤ kf kp kgkq . (2.1.1)
X

Proof. We first note that if a, b are nonnegative real numbers, then we have

ap bq
ab ≤ + . (2.1.2)
p q

A geometric way to prove this inequality (called Young’s inequality) is to

ap q
observe that p is the area given 0 x dx, while bq is the area given by
R a p−1

dy. As q − 1 = 1/(p − 1), we have y = xp−1 ⇔ x = y q−1 when

R b q−1
0 y
x, y ≥ 0. By considering the graph of y = x p−1 and the rectangle [0, a]×[0, b]
in the xy-plane, one realizes that (2.1.2) must be true.
Next, we note that we may assume that kf kp = kgkq = 1. Indeed,
assume that (2.1.1) holds whenever kf kp = kgkq = 1, and consider f ∈ Lp
and g ∈ Lq . If kf kp = 0 or kgkq = 0, then both sides of (2.1.1) are equal to
zero. On the other hand, if kf kp and kgkq are both nonzero, then we may
use that (2.1.1) holds for the functions f /kf kp and g/kgkq , and deduce that
it holds in the general case.

10
 2.1. The case 1 ≤ p < ∞

Hence, assume that kf kp = kgkq = 1. Then, using (2.1.2) with a = |f (x)|

and b = |g(x)| for each x ∈ X, and linearity of the integral, we get
Z Z
|f g| dµ = |f (x)| |g(x)| dµ(x)
X X
1 Z 1 Z
≤ |f (x)|p dµ(x) + |g(x)|q dµ(x)
p X q X
1 1
= kf kpp + kgkqq
p q
1 1
= + =1
p q
= kf kp kgkq ,

as desired. 

Corollary 2.1.3. Let p ∈ [1, ∞). Then k · kp is a seminorm on Lp . In

particular, for all f, g ∈ Lp , we have

kf + gkp ≤ kf kp + kgkp (Minkowski’s inequality) (2.1.3)

Proof. As already mentioned, the case p = 1 is straightforward. So assume

p ∈ (1, ∞). The reader should have no problem to see that we have
kλ f kp = |λ| kf kp for all λ ∈ C and all f ∈ Lp . Next, let f, g ∈ Lp , and let
q be p’s conjugate exponent. As (p − 1)q = p and p/q = p − 1, we have
Z 1/q Z 1/q
k |f + g|p−1 kq = |f + g|(p−1)q dµ = |f + g|p dµ
X X
= kf + gkp/q
p = kf + gkp−1
p .

Since f + g ∈ Lp , this shows that |f + g|p−1 ∈ Lq ; moreover, using Hölder’s

inequality (at the 4th step), we get
Z Z
kf + gkpp = |f + g| dµ =
p
|f + g| |f + g|p−1 dµ
ZX X Z
≤ |f | |f + g|p−1 dµ + |g| |f + g|p−1 dµ
X X
≤ kf kp k |f + g|p−1 kq + kgkp k |f + g|p−1 kq
= (kf kp + kgkp ) k|f + g|p−1 kq
= (kf kp + kgkp ) kf + gkp−1
p ,

and Minkowski’s inequality clearly follows. 

11
2. On Lp -spaces

Let {fn } be a sequence in Lp and f ∈ Lp . We note that it may happen

that fn → f pointwise on X while kfn − f kp 6→ 0 as n → ∞. For example
one may let X = R, A = BR , µ = Lebesgue measure on BR , and consider
the sequence given by fn = χ[n,n+1] for each n ∈ N: it converges pointwise
to 0 on R as n → ∞, and satisfies kfn kp = 1 for all n ∈ N.
The following Lp -version of Lebesgue’s Dominated Convergence Theorem
gives conditions ensuring that a pointwise limit is also convergent w.r.t. k · kp .
Proposition 2.1.4. Let p ∈ [1, ∞) and {fn }n∈N ⊆ Lp . Assume that there
exist some g ∈ Lp such that |fn | ≤ g µ-a.e. for all n ∈ N, and some f ∈ M
such that fn → f pointwise µ-a.e. on X.
Then f ∈ Lp and kfn − f kp → 0 as n → ∞.
Proof. The assumptions imply that |fn |p ≤ g p µ-a.e. for all n ∈ N and that
|fn |p → |f |p pointwise µ-a.e. on X. It follows that we |f |p ≤ g p µ-a.e., so
Z Z
|f |p dµ ≤ g p dµ < ∞ ,
X X

hence f ∈ Lp . Further, we get

 p
|fn − f |p ≤ |fn | + |f | ≤ (2 g)p = 2p g p µ-a.e.,

and |fn − f |p → 0 pointwise µ-a.e. on X. Since 2p g p ∈ L1 , we can apply

Lebesgue’s Dominated Convergence Theorem and get
Z Z
lim |fn − f |p dµ = 0 dµ = 0,
n→∞ X X

which gives that kfn − f kp → 0 as n → ∞, as desired. 

Let p ∈ [1, ∞). It follows from Corollary 2.1.3 that we obtain a normed
space Lp by identifying functions in Lp that agree µ-a.e. To achieve this in
a formal way, we first define a relation ∼ on Lp by setting

f ∼g ⇔ f =g µ-a.e.

for f, g ∈ Lp . In other words, f ∼ g ⇔ kf −gkp = 0. It is almost immediate

that ∼ is an equivalence relation, and we will denote the equivalence class
of f ∈ Lp by [f ], that is, we set
n o
[f ] := g ∈ Lp : f ∼ g .

It is then a routine matter to check that

n o
Lp = Lp (X, A, µ) := [f ] : f ∈ Lp

12
 2.1. The case 1 ≤ p < ∞

becomes a normed space w.r.t.

[f ] + [g] := [f + g] , λ [f ] := [λ f ] , k [f ] kp := kf kp
where f, g ∈ Lp and λ ∈ C. (The reader may consult Exercise 2.1 for a more
general statement.) Moreover, we have:
Theorem 2.1.5. Let p ∈ [1, ∞). Then (Lp , k · kp ) is a Banach space.
Proof. Let {[fn ]}n∈N ⊆ Lp be such that ∞ n=1 k[fn ]kp < ∞, i.e., such that
P

S := n=1 kfn kp < ∞. According to Theorem 1.1.6 we have to show that

P∞

the series ∞n=1 [fn ] is convergent in Lp . It suffices to show that there exists
P

some F ∈ L such that limN →∞ k n=1 fn − F kp = 0, because this will give

p PN

that
N N
hX i N
lim k [fn ] − [F ]kp = lim k fn − F kp = lim k fn − F kp = 0 ,
X X
N →∞ N →∞ N →∞
n=1 n=1 n=1

thus showing that ∞n=1 [fn ] converges to [F ] in L .

p
P

For each N ∈ N, set gN := N n=1 |fn |. Also, let g : X → [0, ∞] be given

P

by
∞
g(x) := |fn (x)| for all x ∈ X.
X

n=1
Clearly, the sequence {gN
p
} of A-measurable nonnegative functions is nonde-
creasing, and it converges pointwise to the A-measurable function g p on X.
Further, using Minkowski’s inequality, we get
N N
k |fn | kp =
X X
kgN kp ≤ kfn kp ≤ S
n=1 n=1

for all N ∈ N. Hence, using the Monotone Convergence Theorem, we get

Z Z
g dµ = lim
p p
gN dµ = lim kgN kpp ≤ S p < ∞ .
X N →∞ X N →∞

Since g ≥ 0, it follows from [L; Exercise 7.5.6] that g p is finite µ-a.e., hence
p

that g is finite µ-a.e. This means that the series ∞ n=1 fn (x) is absolutely
P

convergent for every x belonging to some E ∈ A such that µ(E c ) = 0. We

may therefore define F ∈ M by
P
 ∞ fn (x) if x ∈ E,
F (x) =  n=1
0 if x ∈ E c .

With FN := N n=1 fn we then have |FN | ≤ gN ≤ g ∈ L for every N ∈ N,

p
P

and FN → F pointwise µ-a.e. on X as N → ∞. Proposition 2.1.4 gives

now that F ∈ Lp and limN →∞ kFN − F kp = 0, as we wanted to show. 

13
2. On Lp -spaces

2.2 The case p = ∞

We now consider the case p = ∞. Let F(X) denote the vector space
consisting of all complex functions on X (with its natural operations).
By an algebra of complex functions on X, we will mean a subspace of
F(X) which is also closed under pointwise multiplication. For example,
M = M(X, A) is an algebra of complex functions on X. Another natural
algebra is the one consisting of those functions in M which are bounded.
We will actually be interested in a slightly larger algebra.

Definition 2.2.1. A function f ∈ M is said to be essentially bounded

(w.r.t. µ) if there exists some real number M > 0 such that

|f | ≤ M µ-a.e.,
n o
in which case we set kf k∞ := inf M > 0 : |f | ≤ M µ-a.e. .

Example 2.2.2. a) Asume g ∈ M is bounded and set kgku := supx∈X |g(x)|.

Then g is essentially bounded (w.r.t. µ), and we have

kgk∞ ≤ kgku .
 
Indeed, we have µ {x ∈ X : |g(x)| > kgku } = µ(∅) = 0. This gives that
|g| ≤ kgku µ-a.e., and both assertions follow readily.
We note that it may happen that kgk∞ < kgku . For example, consider
the Borel function g on X = R given by g = χ{0} ; letting µ be the Lebsgue
measure on BR , we get

kgk∞ = 0 < 1 = kgku .

b) Consider X = [0, ∞), A = the Borel subsets of X and µ = the Lebesgue

measure on A. Let f ∈ M be given by
∞
f (x) = e +
ix
n χ{2nπ} (x), x ≥ 0.
X

n=1

Then f (2kπ) = k + 1 for every k ∈ N, so f is unbounded. On the other hand,

f is essentially bounded (w.r.t. µ), with kf k∞ = 1, since µ |f |−1 ((M, ∞))
is equal to 0 if M ≥ 1 and to ∞ if 0 < M < 1.

The following useful observation may seem obvious, but it requires a

proof.

14
 2.2. The case p = ∞

Lemma 2.2.3. Let f ∈ M be essentially bounded (w.r.t. µ). Then we have

|f | ≤ kf k∞ µ-a.e. (2.2.1)

Proof. Set B := {x ∈ X : |f (x)| > kf k∞ } ∈ A and assume (for contradic-

tion) that µ(B) > 0. For each n ∈ N, set

1
 
Bn := x ∈ X : |f (x)| > kf k∞ + ∈ A.
n

Clearly, Bn ⊆ Bn+1 for every n, and B = Bn , so we have

S∞
n=1

lim µ(Bn ) = µ(B) > 0 .

n→∞

Hence there must exist at least one N ∈ N such that µ(BN ) > 0. Now, by
definition of kf k∞ , we can find M > 0 such that kf k∞ ≤ M < kf k∞ + N1 and
|f | ≤ M µ-a.e. But this implies that |f | ≤ kf k∞ + N1 µ-a.e., i.e., µ(BN ) = 0,
and we have reached a contradiction. 

Using Lemma 2.2.3, it is straightforward to verify that the set L∞ =

L (X, A, µ) consisting of all functions in M that are essentially bounded
∞

(w.r.t. µ) is an algebra of complex functions on X (cf. Exercise 2.9). Another

application is the following Hölder-type inequality:

Proposition 2.2.4. Let q ∈ [1, ∞), f ∈ L∞ and g ∈ Lq . Then f g ∈ Lq

and
kf gkq ≤ kf k∞ kgkq .

Proof. Using Lemma 2.2.3 we get that |f g|q = |f |q |g|q ≤ kf kq∞ |g|q µ-a.e.
It follows that Z Z
q q
|f g| dµ ≤ kf k∞ |g|q dµ < ∞ .
X X

Hence f g ∈ L . Moreover, taking the q-th root, we obtain the desired

q

inequality. 

Convergence in L∞ with respect to k · k∞ is closely related to uniform

convergence:

Proposition 2.2.5. Let {fn }n∈N ⊆ L∞ and f ∈ L∞ . Then we have that

kfn − f k∞ → 0 as n → ∞ if and only if there exists some E ∈ A such that
µ(E c ) = 0 and fn → f uniformly on E.

15
2. On Lp -spaces

Proof. Assume kfn − f k∞ → 0 as n → ∞. For each n ∈ N, set

Fn := {x ∈ X : |fn (x) − f (x)| > kfn − f k∞ } ∈ A.

Since fn − f ∈ L∞ , we have µ(Fn ) = 0 for each n. Hence, F := ∪n∈N Fn ∈ A

and µ(F ) = 0.
Set now E := F c ∈ A. Then µ(E c ) = 0 and

E = {x ∈ X : |fn (x) − f (x)| ≤ kfn − f k∞ for all n ∈ N} .

It is then obvious that fn → f uniformly on E. The proof of the reverse im-

plication goes along the same lines, and we leave it as an exercise (cf. Exercise
2.11). 

As with the Lp -spaces for 1 ≤ p < ∞, an annoying fact is that in general

k·k∞ is only a seminorm on L∞ . To get a norm we have to identify functions
that agree µ-a.e. Thus, for each f ∈ L∞ we set [f ] = {g ∈ L∞ : g = f µ-a.e.}.
Then n o
L∞ = L∞ (X, A, µ) := [f ] : f ∈ L∞
becomes a vector space w.r.t. the operations given by [f ] + [g] := [f + g],
λ[f ] := [λf ] (where f, g ∈ L∞ and λ ∈ C), and k[f ]k∞ := kf k∞ gives a
norm on L∞ (cf. Exercise 2.1).

Theorem 2.2.6. (L∞ , k · k∞ ) is a Banach space.

Proof. We have to show that L∞ is complete w.r.t. the metric associated

with k · k∞ .
Let {[fn ]}n∈N be a Cauchy sequence in L∞ . So each fn belongs to L∞
and for any given ε > 0, there exists some N ∈ N such that

m, n ≥ N ⇒ k [fm ] − [fn ] k∞ < ε ,

that is,
m, n ≥ N ⇒ k fm − fn k∞ < ε . (2.2.2)
For each m, n ∈ N, set
n o
Fm,n := x ∈ X : |fm (x) − fn (x)| > kfm − fn k∞ .

Then Fm,n ∈ A and µ(Fm,n ) = 0 for all m, n ∈ N (because fm − fn ∈ L∞ ).

Next, set F := Fm,n ∈ A and E := F c ∈ A.

S
m,n∈N

16
 2.2. The case p = ∞

Note that µ(E c ) = µ(F ) = 0 (since 0 ≤ µ(F ) ≤ µ(Fm,n ) = 0).

P
m,n∈N
Moreover,
n o
E= (Fm,n )c = x ∈ X : |fm (x) − fn (x)| ≤ kfm − fn k∞
\ \

m,n∈N m,n∈N
n o
= x ∈ X : |fm (x) − fn (x)| ≤ kfm − fn k∞ for all m, n ∈ N .

Let now ε > 0 be given, and choose N ∈ N such that (2.2.2) holds.
Then for all x ∈ E and all m, n ≥ N , we have
|fm (x) − fn (x)| ≤ k fn − fm k∞ < ε . (2.2.3)
It follows that {fn (x)}n∈N is a Cauchy sequence in C for each x ∈ E. Since
C is complete, this implies that {fn (x)}n∈N is convergent for each x ∈ E,
hence that limn→∞ fn (x) = g(x) for some g(x) ∈ C for each x ∈ E. Thereby
we obtain a function g : E → C, which is AE -measurable since g is the
pointwise limit of the restriction of the fn ’s to E. (Here, AE denotes the
σ-algebra of all sets in A which are contained in E).
We can now extend g to an A-measurable function f : X → C by setting
f (x) = g(x) if x ∈ E, and f (x) = 0 otherwise.
Again, let ε > 0 be given and choose N as above. Then, for all x ∈ E
and all m ∈ N such that m ≥ N , we get from (2.2.3) that
|fm (x) − f (x)| = |fm (x) − g(x)| = lim |fm (x) − fn (x)| ≤ ε .
n→∞

This implies that {fm }m∈N converges uniformly to f on E.

Moreover, set D := E ∩ {x ∈ X : |fN (x)| ≤ kfN k∞ } ∈ A. Then we
have
|f (x)| = |f (x) − fN (x) + fN (x)| ≤ |f (x) − fN (x)| + |fN (x)| ≤ ε + kfN k∞
for all x ∈ D. As
 
0 ≤ µ(Dc ) ≤ µ(F ) + µ {x ∈ X : |fN (x)| > kfN k∞ } = 0 ,
we have µ(Dc ) = 0, so
|f | ≤ ε + kfN k∞ µ-a.e.
This shows that f ∈ L∞ . Using Proposition 2.2.5, we can now conclude
that kfm − f k∞ → 0 as m → ∞. Thus
k [fm ] − [f ] k∞ = kfm − f k∞ → 0 as m → ∞ .
This means that {[fm ]}j∈N converges to [f ] in L∞ . We have thereby shown
that every Cauchy sequence in L∞ is convergent and the proof is finished.


17
2. On Lp -spaces

2.3 Exercises
Exercise 2.1. Let V be a vector space (over F) and let k · k denote a
seminorm on V . Define a relation ∼ on V by setting

v ∼ w ⇔ kv − wk = 0

for v, w ∈ V .
a) Check that ∼ is an equivalence relation.
Denote the equivalence class of v ∈ V by [v], that is,
n o
[v] := w ∈ V : v ∼ w ,
n o
and set Ve := [v] : v ∈ V . Moreover, for v, w ∈ V , and λ ∈ F, set

[v] + [w] := [v + w] , λ [v] := [λ v] , k [v] k := kvk .

b) Show that these operations on Ve are well-defined, that is, show

that if v, v 0 , w, w0 ∈ V are such that v 0 ∼ v, w0 ∼ w, and λ ∈ C, then
(v 0 + w0 ) ∼ (v + w), λv 0 ∼ λv and kv 0 k = kvk.
c) Verify that (Ve , k · k) is a normed space. (Check at least three of the
axioms.)

In the following exercises, unless otherwise specified, (X, A, µ) denotes a

measure space and M denotes the space of A-measurable complex functions
on X.
Exercise 2.2. Assume that X = [1, ∞), A = the Borel subsets of X and µ
is the Lebesgue measure on A. Let f ∈ M be given by
1
f (x) = for all x ≥ 1,
x
and let 1 ≤ p < ∞. Show that f ∈ Lp (X, A, µ) if and only if p > 1, and
compute kf kp in this case.
Exercise 2.3. Assume that X = R, A = the Borel subsets of X and µ is
the Lebesgue measure on A. Let f ∈ M be given by
2
f (x) = e−x for all x ∈ R.

Show that f ∈ Lp (X, A, µ) for all p ∈ [1, ∞)

√ and compute kf kp . (You are
allowed to use that limN →∞ −N e dt = π without proof.)
R N −t2

18
 2.3. Exercises

Exercise 2.4. Assume that X = (0, 1], A = the Borel subsets of X and µ
is the Lebesgue measure on A. Let f ∈ M be given by
1
f (x) = √ for all x ∈ (0, 1],
x
and let 1 ≤ p < ∞.
a) Show that f ∈ Lp (X, A, µ) if and only if p < 2, and compute kf kp in
this case.
b) Let ν be the measure on A given by
Z
ν(A) = x dµ(x) for all A ∈ A .
A

Show that f ∈ Lp (X, A, ν) if and only if p < 4, and compute kf kp in

this case.
Exercise 2.5. Assume that X = [1, ∞), A = the Borel subsets of X and µ
is the Lebesgue measure on A. For each n ∈ N, define fn ∈ M by
n
fn (x) = for all x ≥ 1.
nx + 1
1/3

a) Show that fn ∈ Lp for all n ∈ N whenever 3 < p < ∞.

b) Assume that 3 < p < ∞. Decide whether the sequence {[fn ]}n∈N is
convergent in Lp and find its limit if it converges.
Exercise 2.6. Let p ∈ [1, ∞). Let E denote the space of simple functions in
M and E 0 denote the subspace of E spanned by {χA : A ∈ A, µ(A) < ∞}.
a) Show that E 0 = E ∩ Lp .
b) Let f ∈ Lp . Show that there exists a sequence {gn } in E 0 such that
kf − gn kp → 0 as n → ∞. Deduce that the space
n o
[E 0 ] := [g] : g ∈ E 0
is dense in Lp with respect to k · kp .
Exercise 2.7. Let a, b ∈ R, a < b, A denote the Lebesgue measurable
subsets of X = [a, b] and µ denote the Lebesgue measure on A. Finally, let
C([a, b]) denote the space of all continuous complex functions on [a, b]. Let
p ∈ [1, ∞).
a) Let A ∈ A and δ > 0. Show that there exists some k ∈ C([a, b]) such
that kχA − kkp < δ.
n o
b) Use a) and Exercise 2.6 to show that the space [f ] : f ∈ C([a, b]) is
dense in Lp ([a, b], A, µ) with respect to k · kp .

19
2. On Lp -spaces

Exercise 2.8. Assume that X = R, A = the Lebesgue measurable subsets

of R and µ is the Lebesgue measure on A. Say that a function f : R → C
has compact support if f = 0 outside some closed, bounded interval. Let
Cc (R) denote the space of all continuous complex functions on R which have
compact support. Let p ∈ [1, ∞).
Show that the space {[f ] : f ∈ Cc (R)} is dense in Lp with respect to
k · kp .

Exercise 2.9. Check that k · k∞ is a seminorm on L∞ (so that k · k∞ gives

a norm on L∞ ). Check also that L∞ is an algebra of functions on X and
that we have kf gk∞ ≤ kf k∞ kgk∞ for all f, g ∈ L∞ .

Exercise 2.10. Let f ∈ M. Show that f ∈ L∞ if and only if there exists

a bounded function g ∈ M such that f = g µ-a.e., in which case we have

kf k∞ = inf{ kgku : g ∈ M is bounded and g = f µ-a.e.}.

Exercise 2.11. Finish the proof of Proposition 2.2.5.

Exercise 2.12. Let 1 ≤ p ≤ r < ∞ and X be a nonempty set. Show that

` p (X) ⊆ ` r (X) ⊆ `∞ (X) .

Exercise 2.13. Let p ∈ [1, ∞) and assume that (X, A, µ) is finite, that is,
µ(X) < ∞.
a) Show that L∞ ⊆ Lp .
b) Consider 1 ≤ p ≤ r < ∞ and let f ∈ Lr . Show that f ∈ Lp and
1 1
kf kp ≤ µ(X) p − r kf kr .

Hint: Use Hölder’s inequality in a suitable way.

Note that this shows that Lr ⊆ Lp . In particular, we have L∞ ⊆ L2 ⊆ L1 .
c) Consider the Lebesgue measure on the Borel subsets of R. Give an
example of a function which is in L2 , but not in L1 Give also an example of
a function which is in L∞ , but not in L2 .

Exercise 2.14. Let E denote the space of simple functions in M and let f ∈
L∞ . Show that there exists a sequence {hn } in E such that kf − hn k∞ → 0
as n → ∞. Deduce that the space [E] := {[h] : h ∈ E} is dense in L∞ with
respect to k · k∞ .

20
 CHAPTER 3

More on normed spaces and

linear operators

3.1 Aspects of finite dimensionality

Unless otherwise specified, we always assume that the space Fn , n ∈ N, is
equipped with the Euclidean norm k · k2 given by
 1/2
kxk2 = |x1 |2 + · · · + |xn |2 for x = (x1 , . . . , xn ) ∈ Fn ,

and with the metric induced by this norm. As we recalled in Section 1.1,
all norms on a finite-dimensional vector space are equivalent. The usual
way to prove this is to consider first Fn and show that any other norm on
Fn is equivalent to k · k2 . A crucial fact in the proof is that a subset of
Fn is compact (w.r.t. the metric associated with k · k2 ) if and only if it is
closed and bounded. It will be useful for us to know that this property,
sometimes called the Heine-Borel property, holds in any finite-dimensional
normed space. We will need the following lemma.
Lemma 3.1.1. Let X and Y be finite-dimensional normed spaces. Assume
that X and Y are isomorphic as vector spaces and let T ∈ L(X, Y ) be an
isomorphism. Then T is an isomorphism of normed spaces.
Proof. We have to show that T and T −1 are bounded. To avoid confusion,
we let k · k and k · k0 denote the respective norms on X and Y . For x ∈ X
set
kxkT := kT (x)k0 .
Clearly, the map x → kxkT is a seminorm on X; in fact, it is a norm since

kxkT = 0 ⇔ kT (x)k0 = 0 ⇔ T (x) = 0 ⇔ x = 0,

3. More on normed spaces and linear operators

the last equivalence being a consequence of the injectivity of T . Since X is

finite-dimensional, k · kT is equivalent to k · k. In particular, this means that
there exists some C > 0 such that

kT (x)k0 = kxkT ≤ C kxk for all x ∈ X,

which shows that T is bounded. Similarly, by considering the norm on

Y given by kykT −1 := kT −1 (y)k for y ∈ Y , one deduces that T −1 is also
bounded. 
Proposition 3.1.2. Let X be a finite-dimensional normed space. Then a
subset K of X is compact (w.r.t. the metric induced by the given norm) if
and only if K is closed and bounded.
Proof. Since a compact subset of a metric space is always closed and bounded,
we only have to show the reverse implication. So let K ⊆ X be closed and
bounded. We must show that K is compact. If X = {0}, this is obviously
true, so we may assume that m := dim(X) ≥ 1. Let then T : X → Fm
denote the coordinate map w.r.t. some basis for X. Lemma 3.1.1 gives
that T is an isomorphism of normed spaces. Set K 0 := T (K) ⊆ Fm . Then
K 0 is bounded (since T is bounded). Moreover, K 0 is closed. Indeed, as
K 0 = (T −1 )−1 (K), this follows from the continuity of T −1 . By the Heine-
Borel property of Fm , we can conclude that K 0 is compact. As K = T −1 (K 0 )
and T −1 is continuous, this implies that K is compact, as desired. 
Since the unit ball X1 of a normed space is closed and bounded we get:
Corollary 3.1.3. The unit ball X1 of a finite-dimensional normed space X
is compact.
We note that if X is an infinite-dimensional normed space, then X1 is
not compact. (See Exercises 3.1 and 3.2.) In particular, this implies that an
infinite-dimensional normed space never has the Heine-Borel property.
Another property which is automatically satisfied for a finite-dimensional
normed space is completeness:
Proposition 3.1.4. Let X be a finite-dimensional normed space. Then X
is a Banach space.
Proof. We may clearly assume that X 6= {0}. To show that X is complete,
we let {xn }n∈N be a Cauchy sequence in X and have to prove that it is
convergent. As in the proof of Proposition 3.1.2, we can pick an isomorphism
of normed spaces T : X → Fm , where m = dim(X). For each n ∈ N, set
yn := T (xn ). Since kyn − yk k2 = kT (xn − xk )k2 ≤ kT k kxn − xk k for all

22
 3.1. Aspects of finite dimensionality

k, n ∈ N, we see that {yn }n∈N is a Cauchy sequence in Fm . Since Fm is

complete, there exists y ∈ Fm such that kyn − yk2 → 0 as n → ∞. Set
x := T −1 (y) ∈ X. Then we get

kxn − xk = kT −1 (yn − y)k ≤ kT −1 k kyn − yk2 → 0 as n → ∞.

Thus, {xn }n∈N is convergent, as desired. 

Corollary 3.1.5. Assume M is a finite-dimensional subspace of a normed

space X. Then M is closed in X.

Proof. Assume {xn }n∈N ⊆ M converges to x ∈ X. We have to show that

x ∈ M . As M is complete by Proposition 3.1.4, and {xn }n∈N is a Cauchy
sequence in M , it follows that {xn }n∈N converges to some y ∈ M . Thus we
get that x = limn→∞ xn = y ∈ M . 

Finite dimensionality has also some impact on linear operators.

Example 3.1.6. Let m, n ∈ N and let T ∈ L(Fn , Fm ). Then T is bounded.

Indeed, let A = [ai,j ] denote the standard matrix of T . Then we have
T (x) = (F1 (x), . . . , Fn (x)), where Fi (x) := nj=1 ai,j xj for each i = 1, . . . , m
P

and x = (x1 , . . . , xn ) ∈ Fn . Since each component Fi is clearly a continuous

function from Fn to F, we get that T is continuous, and therefore bounded.

More generally, we have:

Proposition 3.1.7. Let X and Y be normed spaces and let T ∈ L(X, Y ).

Assume that X is finite-dimensional. Then T is bounded.

Proof. By replacing Y with T (X) if necessary, we may assume that Y is

finite-dimensional. Moreover, we may also assume that both X and Y are
different from {0}. Set n = dim(X), m = dim(Y ), and let C : X → Fn ,
D : Y → Fm be isomomorphims, which are then necessarily isomorphisms of
normed spaces by Lemma 3.1.1. The composition T 0 := D ◦ T ◦ C −1 is then
a linear map from Fn to Fm , hence it is bounded by the previous example.
It follows that T = D−1 ◦ T 0 ◦ C, being the composition of bounded maps,
is bounded. 

Note that the above result is not true in general if we instead assume that
Y is finite-dimensional, even in the case where Y = F : a linear functional
T : X → F may be unbounded when X is an infinite-dimensional normed
space. For an example, see Exercise 3.3.

23
3. More on normed spaces and linear operators

Definition 3.1.8. A linear operator T : X → Y between two vector spaces

X and Y is said to have finite-rank if the range of T is finite-dimensional,
i.e., if dim(T (X)) < ∞.

It is obvious that a linear functional on a normed space has always

finite-rank. As such a linear functional can be unbounded, we get that
a finite-rank linear operator T between normed spaces is not necessarily
bounded; in fact, it can be shown that such an operator T is bounded if and
only if ker(T ) is closed. Bounded finite-rank operators have the following
interesting property:

Proposition 3.1.9. Let X and Y be normed spaces over F, and assume that
T ∈ B(X, Y ) has finite-rank. Then, for any given bounded sequence {xn }n∈N
in X, we have that the sequence {T (xn )}n∈N has a convergent subsequence
in Y .

Proof. Assume {xn }n∈N ⊆ X satisfies kxn k ≤ M for all n ∈ N for some
M > 0. Then we have

kT (xn )k ≤ kT k kxn k ≤ kT k M

for all n ∈ N. Now, the ball B := {y ∈ Y : kyk ≤ kT k M } is closed in Y .

Considering T (X) as a normed space w.r.t. to the norm it inherits from
Y , we get that the set K := T (X) ∩ B is a closed and bounded subset of
T (X). Since T (X) is finite-dimensional (by assumption), it follows from
Proposition 3.1.2 that K is compact in T (X). As {T (xn )}n∈N is a sequence
in K, we can therefore conclude that it has a convergent subsequence. 
An operator T ∈ L(X, Y ) satisfying the property described in the
conclusion of Proposition 3.1.9 is said to be compact. We will have a closer
look at this important class of operators in Chapter 5.

24
 3.2. Direct sums and projections

3.2 Direct sums and projections

We first discuss the concepts of direct sums and projections in a purely
linear algebraic setting. Let X be a vector space over F, and let M1 and M2
be subspaces of X. We define the sum of M1 and M2 as the subset of X
given by n o
M1 + M2 := x1 + x2 : x1 ∈ M1 , x2 ∈ M2 .
It is straightforward to verify that it the least subspace of X containing
both M1 and M2 .

Definition 3.2.1. We will say that X is the (internal) algebraic direct sum
of M1 and M2 , and write X = M1 +̇ M2 , when

X = M1 + M2 and M1 ∩ M2 = {0} .

Obviously, we have X = M1 +̇ M2 if and only if X = M2 +̇ M1 .

We first make a simple, but fundamental, observation:

Lemma 3.2.2. The following two conditions are equivalent:

(i) X = M1 +̇ M2 ;

(ii) every x ∈ X can be written in a unique way as x = x1 + x2 with

x1 ∈ M1 and x2 ∈ M2 .

Proof. Assume (i) holds and let x ∈ X. Then we have x = x1 + x2 for some
x1 ∈ M1 , x2 ∈ M2 . If we also have x = x01 + x02 for some x01 ∈ M1 , x02 ∈ M2 ,
then we get
x1 − x01 = x02 − x2 ∈ M1 ∩ M2 .
Since M1 ∩ M2 = {0}, this implies that x01 = x1 and x02 = x2 . Thus (ii)
holds.
Conversely, assume (ii) holds. It then obvious that X = M1 + M2 .
Consider y ∈ M1 ∩ M2 . Then we have y = y + 0 with y ∈ M1 , 0 ∈ M2 ,
and y = 0 + y with 0 ∈ M1 , y ∈ M2 . By uniqueness, we get y = 0. Thus,
M1 ∩ M2 = {0}, so (i) holds. 

Remark 3.2.3. If V1 and V2 are vector spaces over F, then one may form
their direct product V1 × V2 , which is often called the (external) algebraic
direct sum of V1 and V2 . (This concept is presumably well-known; the
definition is recalled in Exercise 3.5). In the case of an (internal) algebraic
direct sum X = M1 +̇ M2 , it can easily be verified that X is isomorphic to
M1 × M2 .

25
3. More on normed spaces and linear operators

Example 3.2.4. a) Let X be the space of all n × n matrices over F, and let
M1 (resp. M2 ) denote the subspace of X consisting of all upper (resp. lower)
triangular matrices in X. Then it is obvious that we have X = M1 + M2 ;
but X is not the algebraic direct sum of M1 and M2 , since M1 ∩ M2 consists
of all the diagonal matrices.
b) Let X be the space of all n × n matrices over R, and let M1 (resp. M2 )
denote the subspace of symmetric (resp. skew-symmetric) matrices in X.
(We recall that A ∈ X is called skew-symmetric when At = −A.) Then we
have X = M1 +̇ M2 . Indeed, if A ∈ X, then A = A1 + A2 , where
1 1
A1 := (A + At ) ∈ M1 and A2 := (A − At ) ∈ M2 .
2 2
Moreover, if A ∈ M1 ∩ M2 , then we have A = At = −A, so A = 0.

There is a tight connection between projection operators and directs

sums.

Definition 3.2.5. Let X be a vector space. An operator P ∈ L(X) is

called a projection when P is an idempotent map, that is, when it satisfies
P 2 = P.

One readily checks that P ∈ L(X) is a projection if and only if I − P is

a projection. We leave it as an exercise to check the following:

Proposition 3.2.6. Assume X = M1 +̇ M2 , and define P1 , P2 : X → X by

P1 (x) := x1 , P2 (x) := x2 ,

whenever x = x1 + x2 with x1 ∈ M1 and x2 ∈ M2 .

Then P1 , P2 are projections in L(X) such that

P1 + P2 = I, P1 P2 = P2 P1 = 0,

P1 (X) = M1 = ker(P2 ) and P2 (X) = M2 = ker(P1 ).

The map P1 is called the projection (from X) on M1 along M2 , while the
map P2 is called the projection (from X) on M2 along M1 .

Example 3.2.7. Consider X = R2 . The most familiar direct sum decom-

position of R2 is of course

R2 = M1 +̇ M2

26
 3.2. Direct sums and projections

where M1 = {(s, 0) : s ∈ R} and M2 = {(0, t) : t ∈ R}, in which case P1 and

P2 are the usual coordinate maps, i.e.,

P1 ((s, t)) = (s, 0) and P2 ((s, t)) = (0, t) .

However, there are infinitely ways of writing R2 as a direct sum, even if

we fix M1 to be the first axis: indeed, we can then let M2 be any line
through the origin which is different from M1 . For example, if we choose
M2 = {(t, t) : t ∈ R}, then M1 ∩ M2 = {(0, 0)}, and for any (u, v) ∈ R2 we
have

(u, v) = (u − v, 0) + (v, v) , with (u − v, 0) ∈ M1 and (v, v) ∈ M2 .

Thus, in this case, we get that the projection maps P1 , P2 : R2 → R2 are

given by P1 ((u, v)) = (u − v, 0) and P2 ((u, v)) = (v, v) for all (u, v) ∈ R2 .

A converse to Proposition 3.2.6 is the following:

Proposition 3.2.8. Assume P ∈ L(X) is a projection. Then we have

X = P (X) +̇ ker(P ).

Moreover, we have P (X) = ker(I − P ), ker(P ) = (I − P )(X), and P is the

projection from X on P (X) along ker(P ).

Proof. Let x ∈ X. Note that

x = P (x) + (x − P (x)) . (3.2.1)

Since  
P x − P (x) = P (x) − P 2 (x) = 0 ,

that is, (x − P (x)) ∈ ker(P ), this shows that X = P (X) + ker(P ).

Next, assume that x ∈ P (X) ∩ ker(P ). Thus we have x = P (y) for some
y ∈ X and P (x) = 0. This gives that

x = P (y) = P 2 (y) = P (P (y)) = P (x) = 0 .

Hence, P (X) ∩ ker(P ) = {0}, so X = P (X) +̇ ker(P ).

If we now set M1 := P (X) and M2 := ker(P ), then, using the notation
from Proposition 3.2.6, we get from equation (3.2.1) that P1 = P and
P2 = I − P , so the last assertions follow readily from this proposition. 

27
3. More on normed spaces and linear operators

Remark 3.2.9. If M1 is a subspace of a vector space X, then it can be

shown that M1 can be algebraically complemented, i.e., that there exists
a subspace M2 of X such that X = M1 +̇ M2 . (In fact, if {0} = 6 M1 6= X,
then there exist infinitely many such subspaces, which are all isomorphic to
each other.) When X is infinite-dimensional, the proof requires the axiom
of choice, in the form of Zorn’s lemma, as explained in more advanced
textbooks.

We now turn our attention to normed spaces.

Definition 3.2.10. Assume that X is a normed space over F, and let M1

and M2 be subspaces of X. We will say that X is the (internal) direct sum
of M1 and M2 , and write
X = M1 ⊕ M2 ,
when X = M1 +̇ M2 and both M1 and M2 are closed in X.

Proposition 3.2.11. Let X be a normed space and assume P ∈ L(X) is a

projection which is bounded (so P ∈ B(X)). Then we have

X = P (X) ⊕ ker(P ) .

Proof. We know from Proposition 3.2.8 that X = P (X) +̇ ker(P ), so it

remains only to check that P (X) and ker(P ) are closed in X. Since ker(P ) =
P −1 ({0}) and P is continuous, ker(P ) is closed. Moreover, since P (X) =
ker(I − P ) and I − P is continuous, we also get that P (X) is closed. 

Example 3.2.12. Let V1 , V2 be normed spaces over F. As is readily verified

(if not already known), the direct product V := V1 × V2 becomes a normed
space with respect to the norm given by

k(v1 , v2 )k := kv1 k + kv2 k .

Moreover, with Ve1 := {(v1 , 0) : v1 ∈ V1 } and Ve2 := {(0, v2 ) : v2 ∈ V2 }, we

have V = Ve1 +̇ Ve2 (cf. Exercise 3.5). Let P1 ∈ L(V ) denote the projection
from V on Ve1 along Ve2 . Then

kP1 ((v1 , v2 ))k = k(v1 , 0)k = kv1 k + k0k = kv1 k ≤ k(v1 , v2 )k

for all (v1 , v2 ) ∈ V , so P1 is bounded. Thus, Proposition 3.2.11 gives that

V = P1 (V ) ⊕ ker(P1 ). As P1 (V ) = Ve1 and ker(P1 ) = Ve2 , we get that

V1 × V2 = Ve1 ⊕ Ve2 .

28
 3.2. Direct sums and projections

Remark 3.2.13. Somewhat surprisingly, if X is a normed space and X =

M1 ⊕ M2 for some closed subspaces M1 , M2 , then it may happen that the
projection P1 from X on M1 along M2 is unbounded (cf. Exercise 3.7), in
which case the projection P2 on M2 along M1 is also unbounded (since
P1 + P2 = I). However this peculiarity does not arise if X is a Banach
space, but the proof of this fact is beyond the scope of these notes. (One
may for example invoke the so-called closed graph theorem, proven in more
advanced courses).
Remark 3.2.14. It is common to say that a closed subspace M of a normed
space X can be complemented when there exists a closed subspace N of X
such that X = M ⊕ N . It is not true that a closed subspace can always be
complemented, even if X is a Banach space; for example, it is known that
the closed subspace
c0 (N) = {f ∈ `∞ (N) : lim f (n) = 0}
n→∞

can not be complemented in `∞ (N) (with uniform norm), but we don’t have
yet the tools necessary to prove this. Proposition 3.2.11 tells us that if a
closed subspace M of a normed space X is the range of a projection P in
B(X), then M can be complemented. The previous remark implies that the
converse holds when X is a Banach space. It is also known that a finite
dimensional subspace of a normed space can always be complemented. We
will see in the next chapter that any closed subspace of a Hilbert space can
be complemented (by its orthogonal complement).
Direct sums and projections are useful in connection with the study of
linear operators.
Proposition 3.2.15. Assume X is a vector space such that X = M1 +̇ M2
for some subspaces M1 , M2 . To each S1 ∈ L(M1 ) and S2 ∈ L(M2 ), we may
associate an operator S = S1 +̇S2 ∈ L(X) given by
(S1 +̇S2 )(x) := S1 (x1 ) + S2 (x2 )
for x = x1 + x2 ∈ X with x1 ∈ M1 and x2 ∈ M2 .
If P1 (resp. P2 ) denote the projection from X on M1 along M2 (resp. on
M2 along M1 ), then S = S1 +̇S2 commutes with each Pj , that is, we have
SPj = Pj S for j = 1, 2.
Moreover, if X is a normed space, M1 and M2 are closed in X, and P1
is bounded (or, equivalently, P2 is bounded), then S1 +̇S2 is bounded if and
only if S1 and S2 are bounded.
(Note that if X is a Banach space, then P1 and P2 are automatically
bounded, as mentioned in Remark 3.2.13).

29
3. More on normed spaces and linear operators

Proof. The reader should have no difficulty to provide the necessary details,
so we leave this as an exercise. 

Definition 3.2.16. Let notation be as in Proposition 3.2.15. When an

operator S ∈ L(X) can be written as S = S1 +̇S2 for some S1 ∈ L(M1 ) and
S2 ∈ L(M2 ), then we say that S is decomposable w.r.t. X = M1 +̇ M2 .

When an operator is decomposable w.r.t. a direct sum decomposition, we

may study it by studying each of its components. It is therefore of interest
to know when this happens. The following notion will be useful.

Definition 3.2.17. Let X be a vector space and T ∈ L(X). A subset M

of X is said to be invariant under T when T (M ) ⊆ M .

Example 3.2.18. Let notation be as in Proposition 3.2.15, and set

S := S1 +̇S2 ∈ L(X). Then M1 and M2 are both invariant under S. Indeed,
if x1 ∈ M1 , then S(x1 ) = S1 (x1 ) ∈ M1 . Similarly, S(x2 ) ∈ M2 for all
x2 ∈ M2 .

Example 3.2.19. Assume X is a vector space over F and T ∈ L(X). The

range of T is then a subspace of X which is invariant under T : indeed, with
M = T (X), we have T (M ) ⊆ T (X) = M .
Moreover, for λ ∈ F, set

EλT := ker(T − λI) .

Then EλT is also a subspace of X, which is invariant under T : indeed, for

every x ∈ EλT , we have T (x) = λx ∈ EλT . Of course, when EλT =
6 {0}, then
λ is an eigenvalue of T , and Eλ is the associated eigenspace.
T

Proposition 3.2.20. Assume X is a vector space over F such that X =

M1 +̇ M2 for some subspaces M1 , M2 . Let P1 (resp. P2 ) denote the projection
on M1 along M2 (resp. on M2 along M1 ) and consider S ∈ L(X). Then the
following conditions are equivalent :

(a) S is decomposable w.r.t. X = M1 +̇ M2 ;

(b) both M1 and M2 are invariant under S ;

(c) S commutes with P1 ;

(d) S commutes with P2 .

30
 3.2. Direct sums and projections

Proof. If (a) holds, then it follows from Proposition 3.2.15 that (c) and (d)
hold. Since P2 = I − P1 , it is elementary that (c) is equivalent to (d).
Assume that (c) holds. Let x1 ∈ M1 . Then we have

S(x1 ) = S(P1 (x1 )) = P1 (S(x1 )) ∈ P1 (X) = M1 .

Thus, M1 is invariant under S. Moreover, as (d) also holds, we get in a

similar way that M2 is invariant under S. Hence, (b) holds.
Finally, assume (b) holds. Then, for j = 1, 2, we may define Sj ∈ L(Mj )
by
Sj (xj ) := S(xj ) for all xj ∈ Mj .
Let x ∈ X. Then x = x1 + x2 for x1 ∈ M1 and x2 ∈ M2 , so we get

S(x) = S(x1 + x2 ) = S(x1 ) + S(x2 ) = S1 (x1 ) + S2 (x2 ) = (S1 +̇S2 )(x) .

This shows that S = S1 +̇S2 , hence that (a) holds. 

Remark 3.2.21. Assume that X is a vector space over F and T ∈ L(X)

has an eigenvalue λ ∈ F. (For example, if X is finite dimensional and F = C,
then every T ∈ L(X) has an eigenvalue). A natural question is then whether
the eigenspace M1 = EλT , which is invariant under T , can be complemented
in X by some subspace M2 which is also invariant under T . This may not be
the case (see Exercise 3.9), but if it happens, then we have T = λIM1 +̇ T2
where T2 = T|M2 ∈ L(M2 ), and we can focus on T2 . Moreover, in good cases,
one can proceed further in an inductive way. This is basically the main idea
used in the proof of the spectral theorem for symmetric real matrices. The
same idea can also be used for compact self-adjoint operators on Hilbert
spaces.

Finally, we mention for completeness that one can also consider direct
sums decompositions of a vector space with more than two summands.
Let X is a vector space over F, and assume that M1 , M2 , . . . , Mn are
subspaces of X. Then X is said to be the (internal) algebraic direct sum of
M1 , M2 , . . . , Mn if X = M1 + M2 + · · · + Mn and the following independence
condition holds: if x1 ∈ M1 , x2 ∈ M2 , . . . , xn ∈ Mn and

x1 + x2 + · · · + xn = 0 ,

then x1 = x2 = · · · = xn = 0. We leave it as an easy exercise to check that

these two conditions are equivalent to requiring that every x ∈ X can be
written in a unique way as x = x1 + · · · + xn with x1 ∈ M1 , . . . , xn ∈ Mn .

31
3. More on normed spaces and linear operators

3.3 Extension by density and continuity

This short section is devoted to a very useful principle in linear analysis,
often called the principle of extension by density and continuity. We will
need the following elementary lemma, which is probably well-known.
Lemma 3.3.1. Assume that X and Y are metric spaces and f, g are con-
tinuous maps from X to Y which agree on a dense subset X0 of X. Then
f = g.
Proof. Let x ∈ X. Since X0 is dense in X, there exists a sequence {xn }n∈N
in X0 which converges to x. By continuity of f and g, we get

f (x) = lim f (xn ) = lim g(xn ) = g(x) .

n n

Theorem 3.3.2. Assume that X is a normed space and Y is a Banach
space (both over F). Assume also that X0 is a dense subspace of X, while Y0
is a subspace of Y . Let T0 ∈ B(X0 , Y0 ). Then T0 extends in a unique way to
an operator T ∈ B(X, Y ). It satisfies that kT k = kT0 k.
Proof. Let x ∈ X. Since X0 is dense in X, there exists a sequence {xn }n∈N
in X0 such that kx − xn k → 0 as n → ∞. In particular, {xn }n∈N is a Cauchy
sequence in X0 . We claim that {T0 (xn )}n∈N is a Cauchy sequence in Y .
Indeed, let ε > 0, and choose N ∈ N such that

kxm − xn k < ε/kT0 k for all m, n ≥ N .

Then, for all m, n ∈ N , we get

kT0 (xm ) − T0 (xn )k = kT0 (xm − xn )k = kT0 k kxm − xn k < ε ,

as desired.
Since Y is complete, we can conclude that there exists some y ∈ Y
such that limn T0 (xn ) = y. Note that y only depends on x. Indeed, as-
sume {x0n }n∈N is another sequence in X0 converging to x. Then the se-
quence x1 , x01 , x2 , x02 , . . . , xn , x0n , . . . in X0 also converges to x, so, arguing
as above, we get that there exists some z ∈ Y such that the sequence
T0 (x1 ), T0 (x01 ), T0 (x2 ), T0 (x02 ), . . . , T0 (xn ), T0 (x0n ), . . . converges to z. This
implies that
lim
n
T0 (x0n ) = z = lim n
T0 (xn ) = y .
Hence it makes sense to define T (x) := y. Doing this for every x ∈ X, we
get a map T : X → Y , and it is easy to check that T is linear, so we leave
this as an exercise.

32
 3.3. Extension by density and continuity

Next, we show that T is bounded. Let x ∈ X and pick {xn }n∈N in X0

converging to x. As T (x) = limn T0 (xn ) and kT0 (xn )k ≤ kT0 k kxn k for all
n ∈ N, we get

kT (x)k = lim
n
kT0 (xn )k ≤ kT0 k lim
n
kxn k = kT0 k kxk .

It follows that T ∈ B(X, Y ) with kT k ≤ kT0 k.

Further, T is an extension of T0 . Indeed, let x ∈ X0 . Then set xn := x
for all n ∈ N. Since {xn }n∈N is a sequence in X0 converging to x, we get
that
T (x) = lim
n
T0 (xn ) = T0 (x) .
The uniqueness of T as an extension of T0 is immediate from Lemma 3.3.1.
Finally, we have

kT0 k = sup{kT0 (x)k : x ∈ X0 , kxk ≤ 1}

= sup{kT (x)k : x ∈ X0 , kxk ≤ 1}
≤ sup{kT (x)k : x ∈ X, kxk ≤ 1} = kT k ≤ kT0 k .

Thus, kT k = kT0 k, as desired. 

Remark 3.3.3. The conclusion of Theorem 3.3.2 is not necessarily true if

Y is a normed space which is not complete (cf. Exercise 3.15).

An interesting special case of Theorem 3.3.2 is when T0 is an isometry.

We recall that a linear map between normed spaces is an isometry when it
is norm-preserving. A linear isometry is clearly bounded.

Corollary 3.3.4. Assume that X is a normed space, Y is a Banach space,

X0 is a dense subspace of X, Y0 is a subspace of Y , and U0 ∈ L(X0 , Y0 ) is
an isometry. Then the unique extension of U0 to an operator U in B(X, Y )
is also an isometry.

Proof. Theorem 3.3.2 guarantees that U0 extends in a unique way to U ∈

B(X, Y ). Let x ∈ X and pick {xn }n∈N in X0 converging to x. We then have
U (x) = limn U0 (xn ), so we get

kU (x)k = lim kU0 (xn )k = lim kxn k = kxk .

n n


Using Corollary 3.3.4, it can be shown that the completion of a (non-

complete) normed space is unique up to isometric isomorphism (cf. Exercise
3.17). We also record an important particular case of Theorem 3.3.2.

33
3. More on normed spaces and linear operators

Corollary 3.3.5. Assume that X is a Banach space and X0 is a dense

subspace of X. Then every T0 ∈ B(X0 ) extends in a unique way to an
operator T ∈ B(X), which satisfies that kT k = kT0 k.

Example 3.3.6. Let a, b ∈ R, a < b, and equip the space C([a, b]) of all con-
tinuous complex functions on [a, b] with the norm kf k2 = ( ab |f (s)|2 ds)1/2 .
R

Considering the square [a, b] × [a, b] as a metric space w.r.t. the Euclidean
metric inherited from R2 , let K : [a, b] × [a, b] → C be a continuous function.
One can then associate to K an integral operator TK on C([a, b]) as follows.
Let f ∈ C([a, b]). Since the function t → K(s, t) f (t) is continuous on
[a, b] for each s ∈ [a, b], we may define a function TK (f ) : [a, b] → C by
Z b
[TK (f )](s) = K(s, t) f (t) dt for all s ∈ [a, b] .
a

We leave it as an exercise to verify, using basic knowledge from elementary

analysis, that TK (f ) is continuous on [a, b] and satisfies
Z bZ b 1/2
kTK (f )k2 ≤ |K(s, t)|2 dsdt kf k2 .
a a

As the map f → TK (f ) is then clearly linear, it follows that TK is a bounded

linear operator from C([a, b]) into itself.
Let now L2 ([a, b]) denote the L2 -space associated with the measure space
([a, b], A, µ), where µ is the Lebesgue measure on the σ-algebra A of all
Lebesgue measurable subsets of [a, b].
As we may identify C([a, b]) with a dense closed subspace of L2 ([a, b])
(cf. Exercise 2.7), we get from Corollary 3.3.5 that TK has a unique extension
to a bounded operator on L2 ([a, b]), also denoted by TK . The function K is
usually called the kernel of the integral operator TK . We will come back to
such integral operators later.
We note that more generally, one can define integral operators associated
with kernels K which are L2 -functions on [a, b] × [a, b] (with respect to
the Lebesgue product measure), but this requires a thorough knowledge of
integration theory on product spaces.

34
 3.4. Exercises

3.4 Exercises
Exercise 3.1. Let H be a Hilbert space which is infinite-dimensional (as a
vector space). Argue first that there exists an orthonormal sequence {xn }n∈N
in H. Then use this sequence to show that the unit ball H1 is not compact.
Exercise 3.2. Let X be a normed space. Let M denote a finite-dimensional
subspace of X, and assume M 6= X.
a) Let x ∈ X \ M . Show that d := inf m∈M kx − mk > 0.
b) Show that there exists y ∈ X such that kyk = 1 and
1
≤ ky − mk for all m ∈ M.
2
c) Assume that X is infinite-dimensional (as a vector space). Show that
the unit ball X1 is not compact.
(Hint : Use b) to construct inductively a sequence {yn }n∈N in X1 such
that 1/2 ≤ kyn − yk k for all 1 ≤ k < n.)
Exercise 3.3. Let X be the subspace of `∞ (N) given by
X = {f : N → C : f (n) = 0 for all but finitely many n}.

a) Show that X is infinite-dimensional.

b) Consider X as a normed space w.r.t. kf ku = supn∈N |f (n)| and let
L : X → C be defined by
∞
L(f ) = f (n)
X

n=1

for all f ∈ X. Clearly, L ∈ L(X, C). Show that L is unbounded. Check

also that ker(L) is not closed in X.
Exercise 3.4. Let PR denote the real vector space consisting of all polyno-
mials in one real variable with real coefficients. For p ∈ PR , set
kpk := sup |p(t)| .
t∈[0,1]

a) Explain why p → kpk gives a well-defined norm on PR .

b) Define a linear operator D : PR → PR by
D(p) = p0 (the derivative of p).
Show that D is unbounded. Conclude that PR is infinite-dimensional.

35
3. More on normed spaces and linear operators

Exercise 3.5. Let V1 , V2 be vector spaces over F. We recall that their

(external) algebraic direct sum (also called their algebraic direct product) is
the vector space
V1 × V2 = {(v1 , v2 ) : v1 ∈ V1 , v2 ∈ V2 },
with operations given by
(v1 , v2 ) + (v10 , v20 ) := (v1 + v10 , v2 + v20 ) ,
λ (v1 , v2 ) := (λv1 , λv2 )
for v1 , v10 ∈ V1 , v2 , v20 ∈ V2 and λ ∈ F.
a) Set Ve1 = {(v1 , 0) : v1 ∈ V1 } and Ve2 = {(0, v2 ) : v2 ∈ V2 }. Check that
Vei is a subspace of V1 × V2 which is isomorphic to Vi for i = 1, 2, and that
V1 × V2 = Ve1 +̇ Ve2 .
b) Assume X is a vector space, M1 and M2 are subspaces of X and
X = M1 +̇ M2 . Show that X is isomorphic to M1 × M2 .
Exercise 3.6. Let V1 , V2 be normed spaces over F and set V := V1 × V2 .
For p ∈ [1, ∞) and (v1 , v2 ) ∈ V , set
 1/p
k(v1 , v2 )kp := kv1 kp + kv2 kp .
Set also k(v1 , v2 )k∞ := max{kv1 k, kv2 k}.
a) Check that k · kp gives a norm, called the p-norm on V , for each
p ∈ [1, ∞]. Then show that all these p-norms on V are equivalent.
b) Set Ve1 := {(v1 , 0) : v1 ∈ V1 } and Ve2 := {(0, v2 ) : v2 ∈ V2 }, so
V = Ve1 +̇ Ve2 (cf. Exercise 3.5). Let P1 ∈ L(V ) denote the projection from V
on Ve1 along Ve2 , and consider the normed space (V, k · kp ) for some p ∈ [1, ∞].
Show that P1 is bounded.
Exercise 3.7. Let X be the subspace of `1 (N) given by
X = {f : N → C : f (n) = 0 for all but finitely many n}
and consider X as a normed space w.r.t. the 1-norm kf k := |f (n)].
P
n∈N
Let M1 be the subspace of X given by
M1 = {f ∈ X : f (2n) = n f (2n − 1) for all n ∈ N} ,
and let M2 be the subspace of X given by
M2 = {f ∈ X : f (2n − 1) = 0 for all n ∈ N} .
Show that X = M1 ⊕ M2 , and that the projection from X on M1 along M2
is unbounded.

36
 3.4. Exercises

Exercise 3.8. Prove Proposition 3.2.15.

Exercise 3.9. Set X = C2 and let {e1 , e2 } denote the standard basis
of X. Let T ∈ L(X) be the linear operator satisfying T (e1 ) = e1 and
T (e2 ) = i e1 + e2 . Clearly, 1 is an eigenvalue of T . Set M1 = E1T , so M1 is a
subspace of X which is invariant under T .
Show that there is no subspace M2 of X which is invariant under T and
satisfies that X = M1 +̇ M2 .

Exercise 3.10. Let X be a vector space over F.

a) Assume X = M1 +̇ M2 for some subspaces M1 and M2 of X, and let
P1 , P2 denote the associated projection maps. Define S ∈ L(X) by

S(x) := P1 (x) − P2 (x) = 2 P1 (x) − x .

Check that S 2 = I. Moreover, check that

M1 = ker(I − S) = {x ∈ X : S(x) = x},

M2 = ker(I + S) = {x ∈ X : S(x) = −x}.

The map S is called the symmetry through M1 along M2 .
b) Assume S ∈ L(X) satisfies S 2 = I. Show that (I +S)(X) = ker(I −S)
and (I − S)(X) = ker(I + S). Moreover, show that

X = ker(I − S) +̇ ker(I + S)

and that S is the symmetry through ker(I − S) along ker(I + S). Finally,
check that S is decomposable with respect to this direct sum decomposition.
c) Assume now that X is a normed space and that S ∈ B(X) satisfies
S = I. Deduce that X = ker(I − S) ⊕ ker(I + S).
2

d) Let a > 0 and consider the space X = C([−a, a]) with the uniform
norm. Define S : X → X by

[S(f )](t) = f (−t) for all f ∈ X and t ∈ [−a, a] .

Check that S is bounded and S 2 = I. Deduce that X = Xeven ⊕ Xodd , where

Xeven := {g ∈ X : g(−t) = g(t) for all t ∈ [−a, a]} and

Xodd := {h ∈ X : h(−t) = −h(t) for all t ∈ [−a, a]}.

37
3. More on normed spaces and linear operators

Exercise 3.11. Let X = R3 and consider X as an inner product space

w.r.t. the Euclidean inner product. Let R ∈ L(X), R 6= IX . Assume that the
standard matrix U of R is orthogonal (i.e., U t U = I) and has determinant
equal to 1.
a) Show that 1 is an eigenvalue of R, and that the associated eigenspace
M := E1R is 1-dimensional.
b) Let N = M ⊥ denote the orthogonal complement of M . As should be
well-known, we have X = M +̇ N .
Show that R is decomposable w.r.t. X = M +̇ N , so we may write
R = IM +̇R0 with R0 ∈ L(N ).
c) Let B 0 be an orthonormal basis for N . Show that the matrix of R0
w.r.t. B 0 is a 2 × 2 rotation matrix.
d) Describe how R acts in geometrical terms.

Exercise 3.12. Let X be a vector space over F and let M be a subspace

of X. Define a relation ∼M on X by x ∼M y if and only if y − x ∈ M .
a) Check that ∼M is an equivalence relation on X.
The equivalence class of x ∈ X w.r.t. ∼M is the set {x + m : m ∈ M },
which we will denote x + M . The set consisting of all these equivalence
classes is called the quotient space (of X by M ), and is denoted by X/M .
b) Check that X/M becomes a vector space over F with respect to the
operations given by

(x + M ) + (x0 + M ) := (x + x0 ) + M , λ (x + M ) := (λx) + M

for all x, x0 ∈ X and λ ∈ F. You should first argue that these operations
are well-defined.
The map Q : X → X/M given by Q(x) = x + M is called the quotient
map. It is evident that Q is linear.
c) Assume now that X = M +̇ N for some subspaces M and N of X.
Show that X/M is isomorphic to N . (Similarly, X/N is isomorphic to M ).
Hint : Consider the map π : N → X/M given by π := Q|N : N → X/M ,
i.e.,
π(y) := y + M for all y ∈ N , (3.4.1)
and show that π is an isomorphism.

38
 3.4. Exercises

Exercise 3.13. Assume that X is a normed space and M is a closed

subspace of X. For each element x + M in the quotient space X/M (defined
in the previous exercise), set

kx + M k := inf kx + mk (= inf kx − mk) .

m∈M m∈M

a) Show that the map x + M → kx + M k gives a norm on X/M , called

the quotient norm. Then check that the quotient map Q : X → X/M
is contractive, i.e., kQ(x)k ≤ kxk for all x ∈ X. In particular, we have
Q ∈ B(X, X/M ).
b) Let now N be a subspace of X such that X = M +̇ N .
Let π : N → X/M be defined by (3.4.1), and let P 0 : X → X denote
the projection from X on N along M . Consider X/M as a normed space
w.r.t. the quotient norm, and M × N as a normed space w.r.t. any choice of
p -norm, cf. Exercise 3.6. Show that the following assertions are equivalent:
(i) N is closed in X (so X = M ⊕N ) and π : N → X/M is an isomorphism
of normed spaces;

(ii) The map (m, n) → m + n from M × N to X is an isomorphism of

normed spaces;

(iii) P 0 is bounded.
Exercise 3.14. Consider X = R2 . Find three subspaces M1 , M2 , M3 of X
such that
• X = M1 + M2 + M3 ;

• M1 ∩ M2 ∩ M3 = {(0, 0)};

• X is not the algebraic direct sum of M1 , M2 and M3 .

This illustrates why the definition of an algebraic direct sum of more than
two subspaces must be formulated in a different way than the one you
possibly had guessed.
Exercise 3.15. Let X be a Banach space having a dense subspace X0 which
is not complete. Consider the identity map I0 : X0 → X0 . Show that I0
does not have an extension to a bounded linear map I0 : X → X0 .
Exercise 3.16. Assume that X is a normed space and Y is a Banach space
(both over F), and let {Tk }k∈N be a sequence in B(X, Y ) which is uniformly
bounded in the sense that M := supk∈N kTk k < ∞ .

39
3. More on normed spaces and linear operators

Moreover, assume that there exists a dense subset S of X such that

{Tk (x)}k∈N converges in Y for every x ∈ S.
Show that there exists T ∈ B(X, Y ) such that
T (x) = lim Tk (x) for all x ∈ X .
k

Exercise 3.17. Assume X0 is a normed space and let (X, i) denote a

completion of X0 , that is, X is a Banach space and i : X0 → X is a linear
isometry such that i(X0 ) is dense in X. (As mentioned in Remark 1.1.7,
such a completion always exists.)
Show that (X, i) is unique up to isometric isomorphism, meaning that
the following holds: if (X 0 , i0 ) is another completion of X0 , then there exists
an isometric isomorphism U : X → X 0 such that i0 = U ◦ i.
Exercise 3.18. a) Provide the details missing in Example 3.3.6.
We outline below how one may define more directly integral operators on
L ([a, b]). Let µ denote the Lebesgue measure on the Lebesgue measurable
2

subsets of [a, b] and let K be a continuous complex function on [a, b] × [a, b].
For each s ∈ [a, b], let ks : [a, b] → C denote the continuous function defined
by ks (t) := K(s, t) for all t ∈ [a, b].
b) Let f ∈ L2 ([a, b]) and s ∈ [a, b]. Show the function ks f is Lebesgue
integrable on [a, b] and satisfies
Z
ks f dµ | ≤ kks k2 kf k2 .
[a,b]

c) Let f ∈ L2 ([a, b]) and define g : [a, b] → K by

Z Z
g(s) = ks f dµ = K(s, t) f (t) dµ(t) for each s ∈ [a, b] .
[a,b] [a,b]

Show that g is continuous and check that

 Z bZ b 1/2
kgk2 ≤ M kf k2 , where M := |K(s, t)|2 dsdt .
a a

Deduce that we obtain a linear map TK0 : L2 ([a, b]) → L2 ([a, b]) by setting
  Z
TK0 (f ) (s) := ks f dµ for each f ∈ L2 ([a, b]) and all s ∈ [a, b],
[a,b]

which satisfies that kTK0 (f )k2 ≤ M kf k2 for all f ∈ L2 ([a, b]).

d) Check that the operator TK : L2 ([a, b]) → L2 ([a, b]) defined by
TK ([f ]) = [TK0 (f )] for all [f ] ∈ L2 ([a, b])
is well-defined, linear and bounded, with kTK k ≤ M .

40
 CHAPTER 4

More on Hilbert spaces

By a Hilbert space we always mean a Hilbert space over F, unless otherwise

specified.

4.1 Geometry in Hilbert spaces

In courses in elementary linear algebra, one learns that if M is finite-
dimensional subspace of an inner product space H, then every vector in
H can be written in a unique way as the sum of a vector in M and a
vector in the orthogonal complement M ⊥ . Since M and M ⊥ are both closed
subspaces of H, this means that H = M ⊕ M ⊥ . The projection of H on M
along M ⊥ is then called the orthogonal projection of H on M . As we are
going to establish, such a decomposition of H also holds when H is a Hilbert
space and M is closed subspace of H, not necessarily finite-dimensional.
We recall first that if (X, d) is a metric space, x ∈ X and A is a nonempty
subset of X, then the distance from x to A is defined by
d(x, A) = inf{d(x, y) : y ∈ A} .
If for example A is compact, then the function y → d(x, y), being continuous,
will attain its minimum on A; hence, in this case, there exists some (not
necessarily unique) xA ∈ A such that d(x, A) = d(x, xA ). However, if A is
only closed, then such an xA may not exist (cf. Exercise 4.1).
If we now consider a Hilbert space H with the metric dH associated to its
norm, a vector x ∈ H and a closed subspace M of H, then M is not compact
and the result above does not apply. However, if M is finite-dimensional,
then we know from previous courses that there exists a unique xM ∈ M
which gives the best approximation to x in M , i.e., which satisfies that
kx − xM k ≤ kx − yk for all y ∈ M,
4. More on Hilbert spaces

that is, we have dH (x, xM ) = dH (x, M ). Moreover, we also know that xM

is given as the orthogonal projection of x on M . When M is not finite-
dimensional, but still closed, we will reverse this prosess by showing first
that there exists a unique best approximation xM to x in M , and then use
this to define the orthogonal projection of x on M .
We will actually prove a more general result, valid for any closed convex
subset of H. We recall that a subset C of some vector space V (over F) is
called convex if C contains the line segment between any two elements of C,
i.e., if we have (1 − t)x + ty ∈ C whenever x, y ∈ C and t ∈ [0, 1].
Clearly, any subspace of a vector space is convex, as is any ball in
a normed space. Using that the norm in a Hilbert space satisfies the
parallellogram law, we will prove the following result, which the reader is
advised to illustrate geometrically by looking at various examples in R2 .
Theorem 4.1.1. Let C be a nonempty closed convex subset of a Hilbert
space H and let x ∈ H. Then there is a unique vector xC ∈ C such that
dH (x, xC ) = dH (x, C), that is, such that
kx − xC k ≤ kx − yk for all y ∈ C.
The vector xC is called the best approximation to x in C.
Proof. We first consider the case where x = 0. We then have to show that
there is a unique vector 0C ∈ C of minimal norm, i.e, which satisfies that
n o
k0C k = inf kyk : y ∈ C .
n o
Set s := inf kyk2 : y ∈ C . For each n ∈ N we can find yn ∈ C such that

1
s ≤ kyn k2 < s + . (4.1.1)
2n
Then the sequence {yn }n∈N is Cauchy in H. Indeed, consider m, n ∈ N.
Then, using the parallellogram law and (4.1.1), we get that
1 1
kyn + ym k2 + kyn − ym k2 = 2 kyn k2 + 2 kym k2 < 4 s + + .
n m
Now, since C is convex, we have c := 12 yn + 21 ym ∈ C. Hence,
kyn + ym k2 = 4 kck2 ≥ 4 s ,
so we get
1 1 1 1
kyn − ym k2 < 4 s + + − kyn + ym k2 ≤ + .
n m n m

42
 4.1. Geometry in Hilbert spaces

Thus, given ε > 0, we can choose N ∈ N such that N ≥ (2ε2 )−1 , and obtain
that kyn − ym k < ε for all n, m ≥ N , as desired.
As H is complete, there exists y0 ∈ H such that limn yn = y0 . Since C
is closed, y0 ∈ C. Letting n → ∞ in (4.1.1), we get that
√
ky0 k = s = inf{kyk : y ∈ C}.

If y00 ∈ C also satisfies that ky00 k = inf{kyk : y ∈ C}, then we can consider
the sequence {zn }n∈N in C given by zn = y00 if n is odd and zn = y0 if n is
even. Since zn satisfies (4.1.1) (with yn = zn ) for each n, we can conclude as
above that {zn }n∈N is convergent. This clearly implies that y00 = y0 . Thus,
y0 is the unique vector in C satisfying ky0 k = inf{kyk : y ∈ C}, and we can
set 0C := y0 .
In the general case where x ∈ H, we note that the set

D := {x − y : y ∈ C}

is closed and convex. Using the first part, we get that there exists a unique
vector 0D ∈ D such that k0D k = inf{kzk : z ∈ D} = d(x, C). Then
xC := x − 0D ∈ C has the desired properties. 
One important application is when C is a closed subspace M of H.
Theorem 4.1.2. Let M be a closed subspace of a Hilbert space H. Then
we have
H = M ⊕ M⊥ .
The associated projection PM of H on M along M ⊥ is given by

PM (x) = xM for all x ∈ H ,

where xM ∈ M is the best approximation to x in M (cf. Theorem 4.1.1). We

call PM the orthogonal projection of H on M and write sometimes ProjM
instead of PM . The linear map PM is bounded, with kPM k = 1 if M =6 {0}.
Moreover, we have

(M ⊥ )⊥ = M and PM ⊥ = IH − PM .

Proof. Let x ∈ H and set x⊥ := x − xM . We claim that x⊥ belongs to M ⊥ .

To show this, let y ∈ M and ε > 0. Since (xM + ε y) ∈ M , we get from
Theorem 4.1.1 that

kx⊥ k2 = kx − xM k2 ≤ kx − (xM + ε y)k2 = kx⊥ − ε yk2

= kx⊥ k2 − 2 ε Re(hx⊥ , yi) + ε2 kyk2 ,

43
4. More on Hilbert spaces

which gives that

2 Re(hx⊥ , yi) ≤ ε kyk2 .
As this holds for every ε > 0, we obtain that Re(hx⊥ , yi) ≤ 0. Applying this
to −y ∈ M , we also get that − Re(hx⊥ , yi) ≤ 0, i.e., Re(hx⊥ , yi) ≥ 0. Thus,
it follows that Re(hx⊥ , yi) = 0. If F = R, this means that hx⊥ , yi = 0. If
F = C, we also have that iy ∈ M , and this gives that

Im(hx⊥ , yi) = Re(−i hx⊥ , yi) = Re(hx⊥ , i yi) = 0.

Thus, hx⊥ , yi = 0 in this case too. As this holds for every y ∈ M , the claim
is proven.
Since x = xM + x⊥ , by definition of x⊥ , we get that

H = M + M⊥ .

Now, we also have that M ∩ M ⊥ = {0} (because if y ∈ M ∩ M ⊥ , then

hy, yi = 0, so y = 0), while M and M ⊥ are both closed in H. Thus,
H = M ⊕ M ⊥ , as we wanted to show.
The projection map PM : H → H on M along M ⊥ is then clearly given
by PM (x) = xM for x ∈ H. Using Pythagoras’ identity, we get that

kPM (x)k2 = kxM k2 ≤ kxM k2 + kx⊥ k2 = kxM + x⊥ k2 = kxk2

for all x ∈ H, showing that kPM k ≤ 1. Since PM (y) = y whenever y ∈ M ,

we have that kPM (y)k = 1 if y ∈ M and kyk = 1. It follows that kPM k = 1
if M 6= {0}, as asserted.
Consider now y ∈ M . Then for all z ∈ M ⊥ , we have hy, zi = 0. This
implies that y ∈ (M ⊥ )⊥ . Hence we have M ⊆ (M ⊥ )⊥ .
To show the reverse inclusion, that is (M ⊥ )⊥ ⊆ M , we first observe that
by applying the first part of the theorem to M ⊥ , we get that

H = M ⊥ ⊕ (M ⊥ )⊥ .

Now let x ∈ (M ⊥ )⊥ , and set again x⊥ := x − xM . Since x⊥ ∈ M ⊥ and

xM ∈ M ⊆ (M ⊥ )⊥ , we can write

x = x⊥ + xM , where x⊥ ∈ M ⊥ and xM ∈ (M ⊥ )⊥ , and

x = 0 + x, where 0 ∈ M ⊥ and x ∈ (M ⊥ )⊥ .

By the uniqueness of decomposition in a direct sum, we get that x = xM ,

so x ∈ M . Thus, we have shown that (M ⊥ )⊥ ⊆ M , and we can conclude
that (M ⊥ )⊥ = M .

44
 4.1. Geometry in Hilbert spaces

Finally, for x ∈ H, we have

x = (x − xM ) + xM ,
where (x − xM ) ∈ M ⊥ and xM ∈ M = (M ⊥ )⊥ . This gives that
PM ⊥ (x) = x − xM = (IH − PM )(x) .
Hence, PM ⊥ = IH − PM .

Remark 4.1.3. Assume that M is finite-dimensional subspace of a Hilbert
space H and that B = {u1 , . . . , un } is an orthonormal basis for M . Then
we know that the orthogonal projection PM of H on M is given by
n
PM (x) = for all x ∈ H .
X
hx, uj i uj
j=1

A similar formula holds when M is only assumed to be a closed subspace

of H, as we will see in the next section after having discussed orthonormal
bases in Hilbert spaces.
Corollary 4.1.4. Let M be closed subspace of a Hilbert space H. Then
M = H if and only if M ⊥ = {0}.
In connection with the next corollary, we recall that if S is a nonempty
subset of a vector space V , then Span (S) denote the subspace of V consisting
of all possible finite linear combinations of vectors in S.
Corollary 4.1.5. Let S denote a nonempty subset of a Hilbert space H.
Then Span (S) is dense in H if and only if S ⊥ = {0}.
Proof. Set M := Span (S), which is a closed subspace of H. Then Span (S)
is dense in H if and only if M = H. As one easily verifies that S ⊥ = M ⊥
(cf. Exercise 4.3), the result follows from Corollary 4.1.4. 
A nonempty subset S of a normed space X is sometimes called total in
X when Span (S) is dense in X. So the corollary above says that S is total
in H if and only if S ⊥ = {0}.
Example 4.1.6. Let (X, A, µ) be a measure space and set L2 := L2 (X, A, µ).
We can organize L2 as a Hilbert space (over C) as follows.
Let f, g ∈R L2 . Then g is measurable (since g = Re(g) − i Im(g)) and
X |g| dµ = X |g| dµ = kgk2 < ∞, so g ∈ L . Hence, f g ∈ L , and we
2 2 2 2 1
R

can set D E Z
[f ], [g] := f g dµ .
X

45
4. More on Hilbert spaces

We leave it as an exercise to check that this gives a well-defined inner product

on L2 . As the associated norm obviously coincides with k · k2 , L2 is complete
w.r.t. this norm and we can conclude that L2 is a Hilbert space.
Now, let E ∈ A and set F := E c ∈ A. If g : X → C is measurable,
 let
us say that g lives essentially on E when µ {x ∈ F : g(x) 6= 0} = 0. Then
let ME be the subset of L2 given by
n o
ME := [g] : g ∈ L2 and g lives essentially on E .

Similarly, we can define MF . We claim that

MF = (ME )⊥ and ME = (MF )⊥ . (4.1.2)

To prove this, assume first that [g] ∈ ME and [h] ∈ MF . Then one easily
sees that g = g 1E µ-a.e. and h = h 1F µ-a.e., so, as E ∩ F = ∅, we get
D E Z Z
[g], [h] = g 1E h 1F dµ = g h 1E∩F dµ = 0 .
X X

Since this is true for all [g] ∈ ME , this implies that [h] ∈ (ME )⊥ . As this
holds for all [h] ∈ MF , we get that MF ⊆ (ME )⊥ .
To show the reverse inclusion, let [h] ∈ (ME )⊥ . Then we have
Z
g h dµ = 0 whenever [g] ∈ ME .
X

In particular, since [h 1E ] ∈ ME , we get

Z Z
|h|2 1E dµ = (h 1E ) h dµ = 0 .
X X

Since |h|2 1E is nonnegative on X, this implies that

 
µ {x ∈ X : |h(x)|2 1E (x) 6= 0} = 0 .

As {x ∈ E : h(x) 6=  0} = {x ∈ X : |h(x)|2 1E (x) 6= 0}, we get that

µ {x ∈ E : h(x) 6= 0} = 0, hence that h lives essentially on F . Thus,
[h] ∈ MF . This shows that (ME )⊥ ⊆ MF .
Altogether, we have proved that MF = (ME )⊥ . Interchanging E and F ,
we get that ME = (MF )⊥ , and the proof of (4.1.2) is finished.
Since the orthogonal complement of any subset is a closed subspace, we
can conclude that ME and MF are closed subspaces of L2 . Theorem 4.1.2
now gives that
L2 = ME ⊕ (ME )⊥ = ME ⊕ MF .

46
 4.2. Orthonormal bases in Hilbert spaces

We note that the fact that L2 = ME +̇ MF is a simple consequence of the

equation

[f ] = [f 1E ] + [f 1F ] , where [f 1E ] ∈ ME , [f 1F ] ∈ MF ,

which holds for all [f ] ∈ L2 . From this equation, we now see that the
orthogonal projection of L2 on ME (resp. MF ) is given by

PME ([f ]) = [f 1E ] (resp. PMF ([f ]) = [f 1F ] ).

4.2 Orthonormal bases in Hilbert spaces

The notion of an orthonormal basis for a finite-dimensional inner product
space, which is well-known from elementary linear algebra, have a natural
generalization to Hilbert spaces.

Definition 4.2.1. A nonempty subset B of a Hilbert space H is called an

orthonormal basis for H when B is orthonormal and Span (B) is dense in H.

Suppose a Hilbert space H is finite-dimensional (and nonzero). Then an

orthonormal set B in H has to be finite, so Span (B), being finite-dimensional,
is closed in H; hence, Span (B) is dense in H if and only if Span (B) = H.
Thus we see that Definition 4.2.1 agrees with the usual one when H is
finite-dimensional. We also mention that some authors like to define the
empty set to be an orthonormal basis for the trivial Hilbert space H = {0}.
Our first example is of great importance in Fourier analysis.

Example 4.2.2. Let H = L2 ([−π, π], A, µ), where A denotes the σ-algebra
of all Lebesgue measurable subsets of [−π, π], and µ is the normalized
Lebesgue measure on A, that is,
1
µ(A) := λ(A) for all A ∈ A ,
2π
where λ denotes the Lebesgue measure on R. In particular, we have
µ([−π, π]) = 1. For each n ∈ Z, let en : [−π, π] → C denote the con-
tinuous function given by

en (t) := e int for all t ∈ [−π, π] .

As is probably well-known (and easy to check), the set

B := {[en ] : n ∈ Z}

47
4. More on Hilbert spaces

is an orthonormal subset of H. Moreover, Span (B) is dense in H.

To show this, let T denote the space of all (complex) trigonometrical
polynomials, i.e., T := Span ({en : n ∈ Z}). Clearly, we have
Span (B) = {[h] : h ∈ T } .
Further, let Cper ([−π, π]) = {k ∈ C([−π, π]) : k(−π) = k(π)}. We will
use the fact (shown for example in Lindstrøm’s book) that T is dense in
Cper ([−π, π]) w.r.t. the uniform norm k · ku .
Let [f ] ∈ H and ε > 0. Using Exercise 2.7 we can find g ∈ C([−π, π]) such
that
k [f ] − [g] k2 < ε/3 . (4.2.1)
Moreover, it is easy to see that we can pick k ∈ Cper ([−π, π]) such that
k [g] − [k] k2 = k g − k k2 < ε/3 . (4.2.2)
Now, as mentioned above, we can find h ∈ T such that kk − hku < ε/3.
Since Z
k [k] − [h] k2 =
2
|k − h|2 dµ
[−π,π]
Z
≤ kk − hk2u dµ
[−π,π]

= kk − hk2u µ([−π, π])

= kk − hk2u ,
we get
k [k] − [h] k2 ≤ kk − hku < ε/3 . (4.2.3)
Using the triangle inequality, (4.2.1), (4.2.2) and (4.2.3), we obtain that
k [f ] − [h] k2 = k [f ] − [g] + [g] − [k] + [k] − [h] k2
≤ k [f ] − [g] k2 + k [g] − [k] k2 + k [k] − [h] k2
< ε/3 + ε/3 + ε/3 = ε .

This shows that [f ] ∈ Span (B). Hence, Span (B) = H, as asserted.

We can now conclude that B = {[en ] : n ∈ Z} is an orthonormal basis
for H.
More generally, one may consider the L2 -space associated to an interval
[a, b] and the normalized Lebesgue measure µ := b−a 1
λ. Then, letting e0n be
defined for each n ∈ Z by
e0n (t) = e int 2π/(b−a) for all t ∈ [a, b],
one may argue in a similar way as above, and conclude that B 0 = {e0n : n ∈ Z}
is an orthonormal basis for this L2 -space.

48
 4.2. Orthonormal bases in Hilbert spaces

An immediate consequence of Corollary 4.1.5 is the following useful

characterization of orthonormal bases:

Proposition 4.2.3. Assume that B is an orthonormal subset of a Hilbert

space H. Then B is an orthonormal basis for H if and only if B ⊥ = {0}.

Example 4.2.4. Let X be a nonempty set. Then `2 (X) has a natural

orthonormal basis E which is the analogue of the standard basis {e1 , . . . , en }
for Fn (which may be identified with `2 ({1, . . . , n})).
Indeed, for each x ∈ X, let ex ∈ `2 (X) be defined by ex = 1{x} , and set

E := {ex : x ∈ X} .

Then E is clearly orthonormal. Moreover, let f ∈ `2 (X), f ∈ E ⊥ . Thus, for

each x ∈ X, we have hf, ex i = 0. As

hf, ex i = f (y)ex (y) = f (y) = f (x) ,

X X

y ∈X y ∈{x}

we get that f (x) = 0 for all x ∈ X, i.e., f = 0. This shows that E ⊥ = {0},
and Proposition 4.2.3 gives that E is an orthonormal basis for `2 (X).

It will be shown in more advanced courses that every Hilbert space

(which is non-zero) has an orthonormal basis. The proof is nonconstructive
as it relies on Zorn’s lemma, i.e., on the axiom of choice. We will take this
fact as granted. Of course, in applications, it is better to have at hand a
concrete orthonormal basis whenever possible.

Example 4.2.5. The Gram-Schmidt orthonormalization prosess, of great

importance in the finite-dimensional case, can be generalized to cover the
following situation:
Let H be a Hilbert space, H =6 {0}. Let {xj }j∈N be a sequence of vectors in
H \ {0} and set S := {xj : j ∈ N}. Assume that Span (S) is dense in H.
We remark that such a sequence exists whenever H is finite-dimensional
(since repetitions are allowed in a sequence). More generally, it exists
whenever H is separable, i.e., whenever H contains a countable dense subset,
cf. Exercise 4.9.
For each n ∈ N, set Mn := Span ({x1 , . . . , xn }) . We note that for each n
we have Mn ⊆ Mn+1 . Moreover, Span (S) = n∈N Mn .
S

Proceeding inductively, we can construct an orthonormal basis Bn for each

Mn as follows:

49
4. More on Hilbert spaces

n o
i) We set B1 := 1
kx1 k
x1 . Clearly, B1 is an orthonormal basis for M1 .

ii) Let n ∈ N and assume that we have constructed an orthonormal basis

Bn for Mn .
If xn+1 ∈ Mn , then set Bn+1 := Bn . Otherwise, set
n 1 o
yn+1 := xn+1 − ProjMn (xn+1 ) and Bn+1 := Bn ∪ yn+1 .
kyn+1 k
It follows readily that Bn+1 is an orthonormal basis for Mn+1 .
Set now B := Bn . Then B is orthonormal, and Span (B) = Span (S),
S
n∈N
so
Span (B) = Span (S) = H .
Hence, B is an orthonormal basis for H.
We observe that since each Bn is finite, B is countable. Conversely, if H
has a countable orthonormal basis, then it can be shown that H is separable
(cf. Exercise 4.9).

When H is a nontrivial finite-dimensional inner product space, and

B = {u1 , . . . , un } is an orthonormal basis for H, we know that every x ∈ H
has a Fourier expansion w.r.t. B, i.e., we have
n
x=
X
hx, uj i uj .
j=1

As we will soon see, a similar expansion also holds in any infinite dimensional
Hilbert space.
We will use the following notation. If j → tj is a function from a
nonempty set J into [0, ∞), then we set
n X o
tj := sup tj : F ⊆ J, F is finite and nonempty ∈ [0, ∞] .
X

j ∈J j ∈F

Equivalently, j ∈ J tj is the integral of the nonnegative function j → tj

P

w.r.t. the counting measure on P(J) (= the σ-algebra of all subsets of J).
We first note that Bessel’s inequality holds for any orthonormal set:
Lemma 4.2.6. Assume that B is an orthonormal set in an inner product
space H, and let x ∈ H. Then
2
≤ kxk2 ,
X
hx, ui
u∈B
n o
and the set Bx := u ∈ B : hx, ui =
6 0 is countable.

50
 4.2. Orthonormal bases in Hilbert spaces

Proof. Let F be a nonempty finite subset of B. As F is orthonormal, Bessel’s

inequality for F gives that
2
≤ kxk2 .
X
hx, ui
u∈F
Thus we get that
n X 2 o
sup hx, ui : F ⊆ B, F is finite and nonempty ≤ kxk2 ,
u∈F

which proves the first assertion.

Further, this implies that the set Bx,n := {u ∈ B : |hx, ui|2 ≥ 1/n} is
finite for every n ∈ N. Hence, Bx = n∈N Bx,n is countable.
S


The next lemma will be useful at several occasions.
Lemma 4.2.7. Assume {uj : j ∈ N} is a countably infinite orthonormal
set of distinct vectors in a Hilbert space H and let {cj }j∈N be any sequence
in F satisfying that
∞
|cj |2 < ∞ .
X

j=1

Then the series uj converges to some y ∈ H, and we have that

P∞
j=1 cj

hy, uk i = ck for every k ∈ N.

Proof. This result is essentially shown in Lindstrøm’s book, but we sketch
the argument for the ease of the reader. For each n ∈ N, set yn = nj=1 cj uj .
P

Then, for any m > n, Pythagoras’ identity gives that

m m
kym − yn k2 = kcj uj k2 = |cj |2 .
X X

j=n+1 j=n+1

Using the assumption, the sum above can be made as small as we want by
choosing m and n large enough. Thus the sequence {yn }n∈N is Cauchy in
H, so it converges to some y ∈ H, i.e., we have
∞
y=
X
cj uj .
j=1

For each k ∈ N, continuity and linearity of the inner product in the first
variable gives then that
∞
hy, uk i = cj huj , uk i = ck .
X

j=1

51
4. More on Hilbert spaces

Theorem 4.2.8. Let H be a Hilbert space, H = 6 {0}, and let B be an

orthonormal subset of H. Then the following conditions are equivalent:
(a) B is an orthonormal basis for H.
(b) Every x ∈ H \ {0} has a Fourier expansion

x= (4.2.4)
X
hx, ui u
u ∈ Bx

where Bx = {u ∈ B : hx, ui =
6 0} is countable (cf. Lemma 4.2.6) and
nonempty.
By (4.2.4) we mean that if Bx is not finite, and Bx = {uj : j ∈ N} is
any enumeration of the distinct elements of Bx , then we have
n ∞
lim x − hx, uj i uj = 0 , i.e., x=
X X
n→∞
hx, uj i uj .
j=1 j=1
2
(c) For every x ∈ H we have kxk2 = hx, ui .
P
u∈B

The formula in (c) is called Parseval’s identity.

Proof. (a) ⇒ (b): Assume that B is an orthonormal basis for H and let
x ∈ H \ {0}.
We first observe that Bx 6= ∅. Indeed, suppose that Bx = ∅. This means
that x ∈ B ⊥ . But B ⊥ = {0} by Proposition 4.2.3, so x = 0, a contradiction.
We now consider the case where Bx is countably infinite. (The case
where Bx is finite is much easier and left to the reader). Let {uj : j ∈ N}
be an enumeration of the distinct elements of Bx . Since Bx is orthonormal,
Bessel’s inequality gives that
∞
hx, uj i|2 ≤ kxk2 .
X

j=1

Applying Lemma 4.2.7 with cj = hx, uj i for every j ∈ N, we get that the
series ∞j=1 hx, uj i uj converges to some y ∈ H, which satisfies that
P

hy, uk i = ck = hx, uk i for every k ∈ N.

Moreover, if u ∈ B \ Bx , we get that

∞
hy, ui = hx, uj i huj , ui = 0 = hx, ui .
X

j=1
It follows that x − y ∈ B = {0}, hence that x = y. This shows that the
⊥

assertion in (b) holds in this case.

52
 4.2. Orthonormal bases in Hilbert spaces

(b) ⇒ (c): Assume (b) holds, and let x ∈ H \ {0}. Again we consider
the more difficult case where Bx is countably infinite, so Bx = {uj : j ∈ N}
as above. By continuity of the norm and Pythagoras’ identity, we get
∞
2
kxk2 =
X
hx, uj i .
j=1
2
Hence, given ε > 0, we can find n ∈ N such that kxk2 −
Pn
j=1 hx, uj i < ε,
giving
n
2 2
kxk2 − ε <
X X
hx, uj i ≤ hx, ui .
j=1 u∈B
2
Since this holds for every ε > 0, we get that kxk2 ≤ u ∈ B hx, ui . Com-
P

bining this inequality with Lemma 4.2.6, we see that (c) holds.
2
(c) ⇒ (a): Assume kxk2 = u ∈ B hx, ui for every x ∈ H. If x ∈ B ⊥ ,
P

i.e., hx, ui = 0 for every u ∈ B, then we get kxk2 = 0, so x = 0. Hence,

B ⊥ = {0}, and Proposition 4.2.3 gives that (a) holds.


Remark 4.2.9. The Fourier expansion of x in Theorem 4.2.8 (b) can be

written in the form
x= (4.2.5)
X
hx, ui u
u∈B

if one takes care of giving a meaning to convergence of generalized sums in

normed spaces. We discuss this in Exercise 4.12. In these notes, we will
sometimes use (4.2.5) as a short form of the Fourier expansion of x given by
(4.2.4).

Example 4.2.10. Let M be a closed subspace of a Hilbert space H and

assume that we have found an orthonormal basis C for M . Then we can use
it to compute the orthogonal projection PM of H on M :
Let x ∈ H. If x ∈ M ⊥ , then PM (x) = 0, so we can assume x ∈ H \ M ⊥ .
Since C is orthonormal in H, we know that Cx := {v ∈ C : hx, vi 6= 0} is
countable. Set xM := PM (x) ∈ M and x⊥ := x − xM ∈ M ⊥ , and note that
xM 6= 0. Now, for each v ∈ C, we have

hx, vi = hxM , vi + hx⊥ , vi = hxM , vi .

Hence, Cx = CxM . Moreover, applying Theorem 4.2.8 to M , xM ∈ M \ {0}

and C, we get that
xM =
X
hxM , vi v .
v ∈ C xM

53
4. More on Hilbert spaces

Using our previous observations, this formula can be rewritten as

PM (x) =
X
hx, vi v ,
v ∈ Cx

generalizing the usual formula for PM (x) when M is finite-dimensional.

A straightforward consequence of Theorem 4.2.8 is the following:
Corollary 4.2.11. Assume a Hilbert space H contains a countably infinite
orthonormal subset B, enumerated as B = {vk : k ∈ N}. Then B is an
orthonormal basis for H if and only if
∞
x=
X
hx, vk i vk
k=1

for all x ∈ H, if and only if

∞
2
kxk =2
X
hx, vk i
k=1

for all x ∈ H.
Example 4.2.12. Let B = {[en ] : n ∈ Z} denote the orthonormal basis
for H = L2 ([−π, π], A, λ/2π) described in Example 4.2.2. For [f ] ∈ H and
n ∈ Z it is common to set
1 Z
[f ] (n)
d := h [f ], [en ] i = f (t)e−int dλ(t) ,
2π [−π,π]
which is called the Fourier coefficient of [f ] at n.
In fact, it is usual to write f instead of [f ], having in mind that one
identifies functions which agree µ-a.e. Hence, the Fourier coefficient of f
at n is denoted by fb(n), and the Fourier expansion of f w.r.t. B is then
written as
f= fb(n) en ,
X

n∈Z

meaning that
m
f = lim fb(n) en (w.r.t. k · k2 ).
X
m→∞
n=−m

This follows from Corollary 4.2.11 by enumerating B as e0 , e−1 , e1 , e−2 , e2 ,

etc. Similarly, we have
2
kf k22 = fb(n) .
X

n∈Z

54
 4.3. Adjoint operators

4.3 Adjoint operators

Let X be a normed space (over F). We recall that the dual space X ∗ consists
of the bounded linear functionals on X (with values in F), and that X ∗ is
a Banach space w.r.t. the norm kϕk = sup{|ϕ(x)| : x ∈ X1 }. The goal of
functional analysis is to gain new insight by exploiting the interplay between
a space and its dual. This is particularly successful when X is a Hilbert
space because the dual space may then be identified in a natural way with
the space itself.

Theorem 4.3.1. Let H be a Hilbert space (over F). For each y ∈ H, define
ϕy : H → F by
ϕy (x) := hx, yi for all x ∈ H .
Then ϕy ∈ H ∗ for all y ∈ H.
Moreover, the map y → ϕy is a bijection from H onto H ∗ , which is
isometric, and conjugate-linear in the sense that

ϕλ1 y1 +λ2 y2 = λ1 ϕy1 + λ2 ϕy2

for all λ1 , λ2 ∈ F and all y1 , y2 ∈ H.

Proof. Let y ∈ H. Then the map ϕy is clearly linear. Moreover, for all
x ∈ H, we have
|ϕy (x)| = |hx, yi| ≤ kxk kyk .
Hence, ϕy is bounded, with kϕy k ≤ kyk. If y 6= 0, then
 1  1
ϕy y = hy, yi = kyk,
kyk kyk

so kϕy k ≥ kyk. Thus, kϕy k = kyk.

This shows that the map y → ϕy is an isometry from H into H ∗ . In
particular, it is injective. To show that it is surjective, let ϕ ∈ H ∗ . If ϕ = 0,
then we have ϕ = ϕ0 . So assume ϕ = 6 0 and set M := ker ϕ. Then M is a
closed subspace of H such that M 6= H. By Corollary 4.1.4, M ⊥ 6= {0}, so
we can pick z ∈ M ⊥ such that kzk = 1, and set

y := ϕ(z) z ∈ H .
We claim that ϕ = ϕy . Indeed, let x ∈ H and set m := ϕ(x) z − ϕ(z) x ∈
H. Then we have

ϕ(m) = ϕ(x) ϕ(z) − ϕ(z) ϕ(x) = 0 ,

55
4. More on Hilbert spaces

so m ∈ M . As z ∈ M ⊥ , we get hm, zi = 0, i.e.,

D E D E
ϕ(x) z, z = ϕ(z) x, z .
Hence
ϕ(x) = ϕ(x) kzk2 = hϕ(x) z, zi = hϕ(z) x, zi = ϕ(z) hx, zi

= hx, ϕ(z) zi = hx, yi = ϕy (x) ,

which shows the claim above, hence that the map y → ϕy is surjective.
Altogether, we have shown that this map is an isometric bijection from
H onto H ∗ , as desired.
The final assertion is an obvious consequence of the conjugate-linearity
of the inner product in the second variable. 
This theorem, which is one among a diversity of results being called
the Riesz representation theorem, has several useful consequences that will
be covered in later courses. Our main application here will be to use it to
associate an adjoint operator to every bounded operator on a Hilbert space.
Some people like to think of the adjoint as a kind of twin (or as a kind
of shadow), which happens to coincide with the original operator in many
cases of interest.
Theorem 4.3.2. Let H be a Hilbert space (over F). For each T ∈ B(H),
there is a unique operator T ∗ ∈ B(H), called the adjoint of T , satisfying
D E D E
T (x), y = x, T ∗ (y) (4.3.1)

for all x, y ∈ H .
The ∗-operation on B(H), T → T ∗ , enjoys the following properties:
For all S, T ∈ B(H) and all α, β ∈ F, we have
• i) (α S + β T )∗ = α S ∗ + β T ∗ ; ii) (ST )∗ = T ∗ S ∗ ; iii) (T ∗ )∗ = T ;
• iv) kT ∗ k = kT k ; v) kT ∗ T k = kT k2 .
Remark 4.3.3. If H and K are Hilbert spaces (over the same F), then one
may associate to each T ∈ B(H, K) a unique adjoint operator T ∗ ∈ B(K, H)
satisfying (4.3.1) for all x ∈ H and all y ∈ K, and enjoying similar properties.
We leave this as an exercise.
Proof of Theorem 4.3.2. Let T ∈ B(H) and consider y ∈ H. Using the
linearity of T and the linearity of the inner product in the first variable, we
get that the map ϕ : H → F defined by
D E
ϕ(x) := T (x), y for all x ∈ H ,

56
 4.3. Adjoint operators

is a linear functional on H. Moreover, as we have

D E
T (x), y ≤ kT (x)k kyk ≤ kT k kxk kyk

for all x ∈ H, ϕ is bounded with kϕk ≤ kT k kyk. Hence, ϕ ∈ H ∗ , and

Theorem 4.3.1 gives that there exists a unique vector in H, that we denote
by T ∗ (y), such that ϕ = ϕT ∗ (y) , i.e., such that
D E D E
T (x), y = x, T ∗ (y) (4.3.2)
for all x ∈ H. This theorem also gives that
kT ∗ (y)k = kϕT ∗ (y) k = kϕk ≤ kT k kyk . (4.3.3)
As what we have done above holds for every y ∈ H, we obtain a map
T ∗ : H → H which sends each y ∈ H to T ∗ (y) ∈ H. In view of (4.3.2), it is
clear that (4.3.1) holds for all x, y ∈ H.
To show that T ∗ is linear, let y, y 0 ∈ H and α ∈ F. Then, for all x ∈ H,
we have
D E D E
x, T ∗ (α y + y 0 ) = T (x), α y + y 0
D E D E
= α T (x), y + T (x), y 0
D E D E
= α x, T ∗ (y) + x, T ∗ (y 0 )
D E
= x, α T ∗ (y) + T ∗ (y 0 ) .

This implies that T ∗ (α y + y 0 ) = α T ∗ (y) + T ∗ (y 0 ), as desired.

Next, from (4.3.3), we see that T ∗ is bounded with kT ∗ k ≤ kT k. To show
the asserted uniqueness property of T ∗ , assume that S ∈ B(H) satisfies the
same property as T ∗ , i.e.,
D E D E
T (x), y = x, S(y) for all x, y ∈ H .
Let y ∈ H. Then, for all x ∈ H, we get
D E D E D E
x, S(y) = T (x), y = x, T ∗ (y) .

This implies that S(y) = T ∗ (y). Thus, S = T ∗ .

We leave the proof of properties i) and ii) as an exercise. To show the
other properties, let T ∈ B(H). Then, for each y ∈ H, using equation (4.3.1)
for T ∗ instead of T , we get that, for all x ∈ H, we have
D E D E D E
x, (T ∗ )∗ (y) = T ∗ (x), y = y, T ∗ (x)
D E D E
= T (y), x = x, T (y)

57
4. More on Hilbert spaces

This implies that (T ∗ )∗ (y) = T (y). Thus, (T ∗ )∗ = T , i.e., iii) holds

Now, we have seen that kT ∗ k ≤ kT k holds for all T ∈ B(H). Thus we
get
kT k = k(T ∗ )∗ k ≤ kT ∗ k ≤ kT k .
Hence kT ∗ k = kT k, i.e., iv) holds.
Further, using iv), we get kT ∗ T k ≤ kT ∗ kkT k = kT k2 . On the other
hand, for every x ∈ H, we have
D E D E
kT (x)k2 = T (x), T (x) = x, T ∗ (T (x))
D E
= x, (T ∗ T )(x) ≤ kxk k(T ∗ T )(x)k
≤ kT ∗ T k kxk2 .
This implies that kT k2 ≤ kT ∗ T k. Hence we get kT k2 = kT ∗ T k, i.e., v)
holds. 
Example 4.3.4. Consider H = Fn for some n ∈ N with its usual inner
product, and T ∈ B(H). Let A denote the standard matrix of T . Then the
t
standard matrix of T ∗ is A∗ := A .
Here, A denotes the matrix obtained by conjugating every coefficient of
A, while B t denotes the transpose of a matrix B. Of course, if F = R, then
we just get A∗ = At .
Alternatively, we can formulate our assertion above by saying that if
TA ∈ B(H) denotes the operator given by multiplication with a matrix
A ∈ Mn (F), then we have
(TA )∗ = TA∗ .
To prove this, let x, y ∈ H. Recall that hx, yi = xt y. So we get
hTA (x), yi = (Ax)t y = xt At y = xt A∗ y = hx, TA∗ (y)i.
Since this holds for all x, y ∈ H, this implies that (TA )∗ = TA∗ , as asserted.
More generally, if H is a nontrivial finite-dimensional Hilbert space, B is
an orthonormal basis for H, and [T ]B is the matrix of T relative to B, then
we have  ∗
[T ∗ ]B = [T ]B .
The verification of this fact is an easy exercise. (One may argue in a similar
way as in the next example).
Example 4.3.5. Assume a Hilbert space H has a countably infinite or-
thonormal basis, enumerated as B = {uj }j∈N . Let T ∈ B(H). For each
(j, k) ∈ N × N, set D E
A(j, k) := T (uk ), uj ∈ F.

58
 4.3. Adjoint operators

We may think of the map A sending each (j, k) to A(j, k) as the (infinite)
matrix of T (w.r.t. B) since, for each k ∈ N, we have
∞ D E ∞
T (uk ) = T (uk ), uj uj = (4.3.4)
X X
A(j, k) uj .
j=1 j=1

Now, as every x ∈ H has a Fourier expansion w.r.t. B, it is clear that T is

uniquely determined as a bounded operator on H by its values on B. Thus
we see from (4.3.4) that T is uniquely determined by its matrix A.
As T ∗ ∈ B(H) and
D E D E D E
T ∗ (uk ), uj = uk , T (uj ) = T (uj ), uk = A(k, j),

we can conclude that the matrix of T ∗ w.r.t. B is A∗ , where

A∗ (j, k) := A(k, j) ,

so that, for all k ∈ N, we have

∞ ∞
T ∗ (uk ) = A∗ (j, k) uj =
X X
A(k, j) uj .
j=1 j=1

From (4.3.4) and Parseval’s identity, we also get that

∞
|A(j, k)|2 = kT (uk )k2 ≤ kT k2 < ∞
X

j=1

for each k ∈ N, so the `2 -norms of the column vectors of A are uniformly

bounded. However, such a condition on the column vectors of an infinite
matrix A is not sufficient in general to ensure that A is the matrix of some
operator in B(H). There are some known conditions guaranteeing this, but
we will only look at two cases below where one can argue directly.
a) Let {λj }j∈N be a bounded sequence in F, so that

M := sup{|λj | : j ∈ N} < ∞.

In other words, the function j → λj belongs to `∞ (N, F).

It is not difficult to see that there exists an operator D ∈ B(H) satisfying
that
D(uk ) = λk uk for each k ∈ N. (4.3.5)
Indeed, consider x ∈ H. Then Parseval’s identity gives that
∞ ∞
2 2
≤ M2 hx, uj i = M 2 kxk2 < ∞.
X X
λj hx, uj i
j=1 j=1

59
4. More on Hilbert spaces

Hence, Lemma 4.2.7 gives that the vector

∞
D(x) :=
X
λj hx, uj i uj
j=1

satisfies that hD(x), uj i = λj hx, uj i for each j ∈ N. Thus, using Parseval’s

identity again, we get that
∞
2
kD(x)k2 = λj hx, uj i ≤ M 2 kxk2 .
X

j=1

It follows now readily that the map x → D(x) gives an operator D ∈ B(H)
such that kDk ≤ M and satisfying (4.3.5). Since kDk ≥ kD(uk )k = |λk |
for all k ∈ N, we also have that kDk ≥ M . Hence, kDk = M .
It is now obvious that the matrix of D (w.r.t. B) is the diagonal (infinite)
matrix Λ defined for each (j, k) ∈ N by

λ
j if j = k,
Λ(j, k) =
0 otherwise.

The operator D is often called the diagonal operator associated to {λj }j∈N
(w.r.t. B).
From our discussion in the first part, we get that the matrix of D∗ is Λ∗ .
Thus we have D∗ (uk ) = λk uk for all k ∈ N, so D∗ is the diagonal operator
associated to { λj }j∈N (w.r.t. B).
b) We may also easily argue that there exists an operator S ∈ B(H)
satisfying that
S(uk ) = uk+1 for all k ∈ N. (4.3.6)
Indeed, since ∞
n=2 |hx, un−1 i| =
2
j=1 |hx, uj i| = kxk < ∞ for all x ∈ H,
∞ 2 2
P P

we may use Lemma 4.2.7 to define a map S : H → H by

∞
S(x) =
X
hx, un−1 i un
n=2

which is then a linear isometry satisfying that S(un−1 ) = un for all n ≥ 2,

i.e., such that (4.3.6) holds. The map S is called the right shift operator on
H(w.r.t. B). The matrix of S (w.r.t. B) is the (infinite) matrix σ given by

1 if j = k + 1,
σ(j, k) =
0 otherwise.

60
 4.3. Adjoint operators

for each (j, k) ∈ N. Thus, the matrix of S ∗ (w.r.t. B) is the matrix σ ∗ given
by 
1 if k = j + 1,
σ ∗ (j, k) = σ(k, j) = 
0 otherwise.
for all j, k ∈ N, so we get that

∞ 0 if k = 1,
S ∗ (uk ) = σ ∗ (j, k) uj = 
X

j=1 uk−1 if k ≥ 2.

The operator S ∗ is called the left shift operator on H (w.r.t. B). We note
that S ∗ is not isometric, in fact not even injective, since S ∗ (u1 ) = 0.

Example 4.3.6. Let (X, A, µ) be a measure space. Set L∞ := L∞ (X, A, µ)

and H := L2 (X, A, µ). For each f ∈ L∞ , we may define a "multiplication"
operator Mf ∈ B(H) by
 
Mf [g] = [f g] for all [g] ∈ H .

Indeed, this follows readily from Proposition 2.2.4 (with q = 2). Now, for
all [g], [h] ∈ H, we have
D E Z Z D E
Mf ([g]), [h] = f g h dµ = g f h dµ = [g], Mf ([h]) .
X X

This implies that (Mf )∗ = Mf .

Example 4.3.7. Let K : [a, b] × [a, b] → C be a continuous function and let

TK denote the associated integral operator on H = L2 ([a, b]), cf. Example
3.3.6 and Exercise 3.18. Then we leave it as an exercise to check that
(TK )∗ = TK ∗ , where K ∗ (s, t) := K(t, s) for all s, t ∈ [a, b] .

As an illustration that the adjoint operator contains valuable information

about the original operator, we include a proposition showing the connection
between the fundamental subspaces associated to these operators.

Proposition 4.3.8. Let H be a Hilbert space (over F) and let T ∈ B(H).

Then we have:

(a) ker(T ) = T ∗ (H)⊥ and ker(T ∗ ) = T (H)⊥ .

(b) T (H) = ker(T ∗ )⊥ and T ∗ (H) = ker(T )⊥ .

61
4. More on Hilbert spaces

Proof. Both equalities in (a) are immediate consequences of (4.3.1). Using

Exercise 4.3 with N = T (H), we then get T (H) = (T (H)⊥ )⊥ = ker(T ∗ )⊥ .
The second equality in (b) is shown similarly (or by replacing T with T ∗ in
the first one).

Corollary 4.3.9. Let H be a Hilbert space (over F) and let T ∈ B(H).
Then the image of T is dense in H if and only if T ∗ is injective (i.e., is
one-to one).
Proof. Using Proposition 4.3.8 and Corollary 4.1.4, we get

T (H) = H ⇔ ker(T ∗ )⊥ = H ⇔ ker(T ∗ ) = {0} .

As T ∗ is linear, we also have ker(T ∗ ) = {0} ⇔ T ∗ is injective. 

As another illustration, we also mention:
Proposition 4.3.10. Let H be a Hilbert space (over F) and let T ∈ B(H).
Then T is invertible in B(H) if and only if T ∗ is invertible in B(H), in
which case we have (T ∗ )−1 = (T −1 )∗ .
Proof. Left to the reader as Exercise 4.17. 

4.4 Self-adjoint operators

In this section, we introduce one of the most important classes of bounded
operators on a Hilbert space and discuss some of their properties.
Definition 4.4.1. Let H be a Hilbert space (over F). An operator T ∈ B(H)
is called self-adjoint when T ∗ = T , that is, we have
D E D
T (x), y = x, T (y)i for all x, y ∈ H .

If F = C, a self-adjoint operator in B(H) is also called Hermitian, while

it is often called symmetric if F = R.
We note that if T, T 0 ∈ B(H) are self-adjoint, and λ ∈ R, then it is
obvious that λT + T 0 is also self-adjoint.
Example 4.4.2. Let A = [aj,k ] ∈ Mn (F) and let TA ∈ B(Fn ) denote the
operator given by multiplication with A (cf. Example 4.3.4). Then TA is
self-adjoint if and only if A∗ = A, i.e., ak,j = aj,k for all j, k ∈ {1, . . . , n}. In
particular, when F = R, TA is self-adjoint if and only if A is symmetric.

62
 4.4. Self-adjoint operators

Example 4.4.3. Assume H is a Hilbert space with a countably infinite

orthonormal basis B. Let D denote a diagonal operator associated to a
bounded sequence {λj }j∈N in F (w.r.t. B), as in Example 4.3.5.
Then D is self-adjoint if and only if λj = λj for all j ∈ N, i.e., λj ∈ R
for all j ∈ N. In particular, D is always self-adjoint when F = R.
Let S denote the right shift operator on H (w.r.t. B). Then S ∗ is the
left shift operator and it is obvious that S ∗ 6= S. So S is not self-adjoint.

Example 4.4.4. Let (X, A, µ) be a measure space and set H := L2 (X, A, µ).
If f ∈ L∞ , then the multiplication operator Mf ∈ B(H) defined in Example
4.3.6 is self-adjoint if and only if Mf = Mf .
Thus, Mf is self-adjoint whenever f is real-valued (µ-a.e.). It can be
shown that the converse statement holds whenever (X, A, µ) satisfies the
mild assumption that it is semifinite (cf. Exercise 4.22).

Example 4.4.5. Let K : [a, b] × [a, b] → C be a continuous function

and let TK denote the associated integral operator on H = L2 ([a, b]), cf.
Example 4.3.7. Then TK is self-adjoint if and only if TK ∗ = TK (where
K ∗ (s, t) = K(t, s)). Hence it is clear that TK is self-adjoint whenever K is
real-valued. We leave it as an exercise to check that the converse statement
also holds.

Example 4.4.6. Let M be a closed subspace of a Hilbert space H and let

PM denote the ortogonal projection of H on M . Then PM is self-adjoint.
Indeed, let x, y ∈ H. As PM (x) ∈ M and y − PM (y) ∈ M ⊥ , we have
D E
PM (x), y − PM (y) = 0 .
D E
Similarly, we have x − PM (x), PM (y) = 0 . Hence we get
D E D E
PM (x), y = PM (x), PM (y) + y − PM (y)
D E D E
= PM (x), PM (y) + PM (x), y − PM (y)
D E
= PM (x), PM (y)
D E D E
= PM (x), PM (y) + x − PM (x), PM (y)
D E
= PM (x) + x − PM (x), PM (y)
D E
= x, PM (y)

It is easy to create self-adjoint operators.

63
4. More on Hilbert spaces

Proposition 4.4.7. Let H be a Hilbert space (over F) and T ∈ B(H).

Then T + T ∗ , T ∗ T and T T ∗ are all self-adjoint. Moreover, if F = C,
then −i (T − T ∗ ) is also self-adjoint.
Proof. The reader should have no difficulty to verify these assertions by
using the properties of the ∗-operation on B(H) listed in Theorem 4.3.2. 
A noteworthy consequence is that bounded self-adjoint operators on a
complex Hilbert space have a canonical decomposition similar to the one
enjoyed by complex numbers.
Corollary 4.4.8. Let H be a Hilbert space over C and let T ∈ B(H). Set
1  1 
Re(T ) := T + T∗ , Im(T ) := T − T∗ .
2 2i
Then Re(T ) and Im(T ) are both self-adjoint, and we have

T = Re(T ) + i Im(T ) .

Proof. The first assertion follows readily from Proposition 4.4.7. The second
one is elementary. 
Consider a bounded operator T on a Hilbert space H 6= {0}. The
numerical range of T is defined as the subset of F given by
nD E o
WT := T (x), x : x ∈ H, kxk = 1 .

Some properties of T are reflected in the geometric properties of WT and of

its closure, see Exercise 4.31 for some facts illustrating this. We will mainly
be interested in the numerical radius of T , given by

NT := sup{ |λ| : λ ∈ WT } = sup{ |hT (x), xi| : x ∈ H, kxk = 1} .

We note that the Cauchy-Schwarz inequality implies that NT ≤ kT k.

As we are going to prove below, a remarkable fact is that NT agrees with
kT k when T is self-adjoint. We observe first that if T is self-adjoint, then
WT ⊆ R. Indeed, if T ∗ = T , then for every x ∈ H, we have

hT (x), xi = hx, T (x)i = hT (x), xi ,

and the claim clearly follows.

Theorem 4.4.9. Let H be a Hilbert space, H 6= {0}, and let T ∈ B(H) be
self-adjoint. Then we have
kT k = NT .

64
 4.4. Self-adjoint operators

Proof. It suffices to prove that kT k ≤ NT , hence that

kT (x)k ≤ NT for all x ∈ H1 . (4.4.1)

We first note that if v ∈ H, then |hT (v), vi| ≤ NT kvk2 .

Indeed, if v = 0, the claim is trivial. Otherwise, if v 6= 0 and u := 1
kvk
v,
so v = kvk u, then
D E D E
T (v), v = kvk2 T (u), u ≤ NT kvk2 .

Let now x ∈ H1 . If T (x) = 0, then the inequality in (4.4.1) is trivially

satisfied, so we can assume that T (x) 6= 0 and set y := kT (x)k
1
T (x) ∈ H1 .
Then we have
1 D E D E
kT (x)k = T (x), T (x) = T (x), y . (4.4.2)
kT (x)k
D E
Similarly, kT (x)k = y, T (x) . As T is self-adjoint, we get
D E
kT (x)k = T (y), x . (4.4.3)

Combining (4.4.2) and (4.4.3), and using our previous observations, as well
as the parallellogram law and the fact that kxk ≤ 1, kyk = 1, we get

1 D E D E
kT (x)k = T (x), y + T (y), x
2
1 D E D E
= T (x + y), x + y − T (x − y), x − y
4
1  
≤ NT kx + yk2 + kx − yk2
4
1  
= NT kxk2 + kyk2
2
≤ NT .

This shows that (4.4.1) is satisfied, as desired. 

Having in mind the spectral theorem for symmetric real matrices, it is

legitimate to wonder whether it could be true that every self-adjoint operator
T ∈ B(H) is diagonalizable in the sense that there exists an orthonormal
basis for H consisting of eigenvectors for T . One quickly realizes that this
can not be the case, as a self-adjoint operator may not have any eigenvalue
at all!

65
4. More on Hilbert spaces

Example 4.4.10. Let H = L2 ([0, 1]) (with usual Lebesgue measure) and
let T = Mf be the self-adjoint operator in B(H) given by multiplication with
the bounded continuous function f (t) = t on [0, 1], cf. Example 4.4.4. Then
the reader should have no trouble in checking that T has no eigenvalues.
We will see in the next chapter that every compact self-adjoint operator
can be diagonalized in the sense mentioned above. Theorem 4.4.9 will help us
to make the first step in proving this, by showing that a compact self-adjoint
operator T has at least one an eigenvalue, namely kT k or −kT k.

4.5 Unitary operators

In this section, we look at another important class of operators on Hilbert
spaces. As a warm-up, we first characterize the linear operators which are
isometric. We recall that if H is a Hilbert space, then a map T : H → H is
said to preserve the inner product when it satisfies
D E
T (x), T (y) = hx, yi for all x, y ∈ H.

Proposition 4.5.1. Let H 6= {0} be a Hilbert space (over F) and let

S : H → H. Then the following conditions are equivalent:
(i) S ∈ B(H) and S ∗ S = IH ;

(ii) S is linear and preserves the inner product ;

(iii) S is a linear isometry.

Proof. (i) ⇒ (ii): Assume S ∈ B(H) satisfies S ∗ S = IH . Then S is linear
and for all x, y ∈ H, we have
D E D E
S(x), S(y) = x, (S ∗ S)(y) = hx, yi ,

so (ii) holds.
(ii) ⇒ (iii): Any map preserving the inner product is isometric, so this is
evident.
(iii) ⇒ (i): Assume S is a linear isometry. Then S ∈ B(H) and T :=
S ∗ S − IH ∈ B(H) is self-adjoint. Then for any x ∈ H, we have
D E D E D E
T (x), x = (S ∗ S −I)(x), x = S(x), S(x) −hx, xi = kS(x)k2 −kxk2 = 0

Thus, WT = {0}, so, using Theorem 4.4.9, we get that kT k = NT = 0.

Hence, T = 0, i.e., S ∗ S = IH , so (i) holds. 

66
 4.5. Unitary operators

Example 4.5.2. Assume H is finite-dimensional and S : H → H is a

linear isometry, so S ∗ S = IH , cf. Proposition 4.5.1. As S is injective, it is
also surjective (since dim(S(H)) = dim(H) − dim(ker(S)) = dim(H), so
S(H) = H). Thus, S is bijective, so it has an inverse S −1 (which is also a
linear isometry). Since S ∗ S = IH , we get that S −1 = S ∗ . In particular, we
also have SS ∗ = IH .

Remark 4.5.3. When H is infinite-dimensional, then a linear isometry S

is not necessarily surjective. A typical example is the right shift operator S
considered in Example 4.3.5, whose range does not contain the first basis
vector; in this case, we have S ∗ S = IH , while SS ∗ 6= IH (cf. Exercise 4.20).

Definition 4.5.4. Let H be a Hilbert space (over F). An operator U ∈ B(H)

is called unitary when it satisfies

U ∗ U = U U ∗ = IH .

Thus, U ∈ B(H) is unitary if and only if U is bijective and U −1 = U ∗ .

When F = R, some authors use the word orthogonal instead of unitary.

Proposition 4.5.5. Let H be a Hilbert space (over F) and let U : H → H.

Then the following conditions are equivalent :

(i) U ∈ B(H) and U is unitary ;

(ii) U is bijective, linear and preserves the inner product ;

(iii) U is a surjective linear isometry.

Proof. (i) ⇒ (ii): If U ∈ B(H) is unitary, then U is bijective and linear,

and Proposition 4.5.1 gives that it preserves the inner product. Hence, (ii)
holds.
(ii) ⇒ (iii): This implication is evident.
(iii) ⇒ (i): Suppose U is a surjective linear isometry. As a linear isometry
is injective, U is bijective. Moreover, Proposition 4.5.1 gives that U ∗ U = IH .
So we get that U −1 = U ∗ , i.e., U is unitary, and (i) holds. 

Example 4.5.6. Assume H has a countably infinite orthonormal basis B

and D is the diagonal operator associated to a bounded sequence {λj }j∈N
in F (w.r.t. B).
Then it is straightforward to check that D is unitary if and only if
λj λj = 1, i.e., |λj | = 1, for all j ∈ N.

67
4. More on Hilbert spaces

Example 4.5.7. Let (X, A, µ) be a measure space and set H := L2 (X, A, µ).
For f ∈ L∞ , consider the multiplication operator Mf ∈ B(H). Then we
clearly have
(Mf )∗ Mf = M|f |2 = Mf (Mf )∗ ,
so we see that Mf is unitary whenever |f | = 1 µ-a.e. The converse holds if
µ is semifinite, cf. Exercise 4.29.

Example 4.5.8. Let H = `2 (Z). We may then define the bilateral forward
shift operator U : H → H by

[U (ξ)](j) = ξ(j − 1) for all ξ ∈ H and all j ∈ Z.

Indeed, since the counting measure on Z is obviously translation-invariant,

we have j∈Z |ξ(j − 1)|2 = j∈Z |ξ(j)|2 < ∞, so we see that U (ξ) ∈ H and
P P

kU (ξ)k2 = kξk2 . Thus U is isometric.

Clearly, U is also linear. Moreover, it is surjective: if η ∈ H, then we
have U (ξ) = η, where ξ is defined by ξ(j) := η(j + 1) for every j ∈ Z; one
may here argue as above to see that ξ ∈ H.
We may now conclude from Proposition 4.5.5 that U is unitary. Its adjoint
U = U −1 is called the bilateral backward shift operator (on H = `2 (Z)). We
∗

note that if B = {en }n∈Z denotes the canonical basis of H = `2 (Z) as in

Example 4.2.4, then we have

U (en ) = en+1 and U ∗ (en ) = en−1 for all n ∈ Z.

Let now H, K be Hilbert spaces (over F). A bijective, linear map U from
H onto K which preserves the inner product is often called an isomorphism
of Hilbert spaces. As in Proposition 4.5.5, one shows that it is equivalent to
require that U is a surjective linear isometry, or that U ∈ B(H, K) is unitary
in the sense that we have U ∗ U = IH and U U ∗ = IK . (Here, U ∗ ∈ B(K, H)
denotes the adjoint of U , cf. Remark 4.3.3). We will say therefore say that
H and K are isomorphic as Hilbert spaces when such a map U : H → K
exists.

Theorem 4.5.9. Let H 6= {0} be a Hilbert space over C, and let B be an

orthonormal basis of H. Then H and `2 (B) are isomorphic as Hilbert spaces.

Proof. Let x ∈ H and define xb : B → C by

xb(u) := hx, ui for all u ∈ B.

68
 4.5. Unitary operators

Then Parseval’s identity says that u∈B |xb(u)|2 = kxk2 . In particular, we

P

have xb ∈ `2 (B) and kxbk = kxk. Thus we can define an isometric map
U : H → `2 (B) by
U (x) = xb for all x ∈ H.
It is elementary to check that U is linear. Moreover, U is surjective.
Indeed, let ξ ∈ `2 (B). As u∈B |ξ(u)|2 < ∞, the set
P

Bξ := {u ∈ B : ξ(u) 6= 0}

must be countable. Let {uj }j∈N be an enumeration of Bξ , where N =

{1, . . . , n} for some n ∈ N or N = N. If N = nN, we have ∞ |ξ(uj )|2 < ∞ ,
P
j=1
o
and this implies readily that the sequence j=1 ξ(uj ) uj k∈N is Cauchy,
Pk

hence convergent in H. Thus we may define x ∈ H by x := j∈N ξ(uj ) uj ,

P

and we then have


k) if u = uk for some k ∈ N ,
ξ(u
xb(u) = hx, ui = ξ(uj )h uj , ui =
X

j∈N
 0 if u ∈ B \ Bξ ,

i.e., xb(u) = ξ(u) for all u ∈ B. Hence, U (x) = ξ, showing that U is surjective.
We can now conclude that U is an isomorphism of Hilbert spaces from
H to `2 (B), as we wanted to show.


Remark 4.5.10. Theorem 4.5.9 is also true when H 6= {0} is a Hilbert

space over R, but one has then to replace `2 (B) with the real `2 -space
n o
`2R (B) := ξ : B → R : |ξ(u)|2 < ∞,
X

u∈B

considered as a Hilbert space over R.

Remark 4.5.11. If H 6= {0} is a Hilbert space over C, and B, B 0 are both

orthonormal bases of H, then we get from Theorem 4.5.9 that `2 (B) and
`2 (B 0 ) are isomorphic as Hilbert spaces. It can be shown that this implies
that (and in fact is equivalent to) B and B 0 having the same cardinality,
meaning that there exists a bijection between B and B 0 . (A similar statement
holds if H 6= {0} is a Hilbert space over R).

69
4. More on Hilbert spaces

4.6 Exercises
In the exercises of this chapter, H always denotes a Hilbert space over F,
unless otherwise stated.

Exercise 4.1. Consider X := `∞ (N) as a metric space w.r.t. d(f, g) =

kf − gku . Let A be the subset of X given by
n o
A := a(N ) : N ∈ N ,

where a(N ) (n) = 1 if 1 ≤ n ≤ N and a(N ) (n) = 0 if n > N .

a) Show that A is closed in X.
b) Let x ∈ X be given by x(n) = 1 + 1/n for all n ∈ N.
Show that d(x, A) = 1 and that 1 < d(x, a(N ) ) for all N ∈ N.

Exercise 4.2. Let c ∈ H, r > 0 and set B := Br (c) = {y ∈ H : ky−ck ≤ r}.

Check that B is closed and convex, and give a formula for xB when x ∈ H \B.

Exercise 4.3. Let S denote a nonempty subset of H and set

M := Span (S).
Verify that S ⊥ = M ⊥ . Then deduce that M = (S ⊥ )⊥ . Deduce also that
if N is a subspace of H, then N = (N ⊥ )⊥ .

Exercise 4.4. Let M be a closed subspace of H, and let x ∈ H.

Show that PM (x) = y0 for some y0 ∈ M if and only if x − y0 ∈ M ⊥ .
Show also that PM (x) is the unique vector y0 in M such that x − y0 ∈ M ⊥ .

Exercise 4.5. Assume P ∈ B(H) is a projection satisfying kP k = 1. Show

that P is the orthogonal projection of H on M := P (H).
Hint : Recall that H = P (H) ⊕ ker P (cf. Proposition 3.2.11).

Exercise 4.6. Consider H = L2 ([a, b], A, µ), where A denotes the σ-algebra
of all Lebesgue measurable subsets of [a, b], and µ is the usual Lebesgue
measure on A. Set
n Z o
M := [g] ∈ H : g ∈ L2 , g dµ = 0 .
[a,b]

Check that M is a closed subspace of H. Then, given [f ] ∈ H, find an

expression for the best approximation of [f ] in M .

70
 4.6. Exercises

Exercise 4.7. Let H1 , H2 be Hilbert spaces over F and consider H :=

H1 × H2 as a vector space over F. For (x1 , x2 ), (y1 , y2 ) ∈ H, set
D E
(x1 , x2 ), (y1 , y2 ) := hx1 , y1 i + hx2 , y2 i .

a) Check that this gives an inner product on H such that H is a Hilbert

space. Check also that the associated norm on H corresponds to the norm
k · k2 arising from the norms on H1 and H2 .
b) Set H
f := {(x , 0) : x ∈ H } and H
1 1 1 1
f := {(0, x ) : x ∈ H },
2 2 2 2

so H = H
f +̇ H
1
f (cf. Exercise 3.5).
2

Check that (H
f )⊥ = H
1
f and (H
2
f )⊥ = H
2
f . Deduce that the projection
1

of H on H
f along H
1
f is the orthogonal projection of H on H
2
f.
1

Exercise 4.8. Let (X, A, µ) be a measure space.

a) Show that Z
D E
[f ], [g] := f g dµ
X

gives a well-defined inner product on L2 = L2 (X, A, µ) (cf. Example 4.1.6).

b) Let E ∈ A. Set AE = {A ∩ E : A ∈ A} and µE = µ|AE . We recall
that (E, AE , µE ) is a measure space.
Show that there exists an isometric isomorphism from L2 (E, AE , µE )
onto the space ME defined in Example 4.1.6.

Exercise 4.9. Let H =

6 {0}. Show that the following conditions are
equivalent:

(a) H is separable;

(b) There is a sequence satisfying the assumptions in Example 4.2.5;

(c) H has a countable orthonormal basis.

Note that Example 4.2.5 shows that (b) ⇒ (c). So it suffices to show that
(a) ⇒ (b), and (c) ⇒ (a).

Exercise 4.10. Let H1 and H2 be Hilbert spaces over F, and let H be the
(external) direct product of H1 and H2 , as defined in Exercise 4.7. Assume
B1 and B2 are orthonormal bases for H1 and H2 , respectively.
Find an orthonormal basis B for H in terms of B1 and B2 .

71
4. More on Hilbert spaces

Exercise 4.11. Let H = L2 ([−1, 1], A, µ), where A denote the Lebesgue-
measurable subsets of [−1, 1] and µ is the restriction of the usual Lebesgue
measure to A.
For each nn ∈ {0} ∪ N, let pn+1 o: [−1, 1] → C be defined by pn+1 (t) = tn ,
and set S := [pn+1 ] : n ∈ {0} ∪ N ⊆ H.
a) Show that Span (S) is dense in H.
b) Apply the Gram-Schmidt
n orthonormalization
o process to S to obtain
an orthonormal basis B = [qn+1 ] : n ∈ {0} ∪ N for H, where each qn+1 is
the polynomial on [−1, 1] given by
q
n+ 1
2 dn  2 
qn+1 (t) = (t − 1)n
.
2n n! dtn
(These polynomials are called the normalized Legendre polynomials.)

Exercise 4.12. The concept of generalized sums can be used to provide an

alternative way of describing Fourier expansions in Hilbert spaces.
Let X be a normed space, J be a nonempty set, {xj }j∈J be a family of
elements of X, and x ∈ X. Then one says that the generalized sum j∈J xj
P

converges to x when the following holds: given ε > 0, there exists a finite
subset F0 ⊆ J such that for all finite subsets F of J containing F0 , we have
X
x− xj k < ε ,
j∈F

in which case we write

x=
X
xj .
j∈J

Consider a Hilbert space H and x ∈ H.

a) Show that we have

x=
X
hx, ui u .
u∈B

b) Show also that if M is a closed subspace of H and C is an orthonormal

basis for M , then we have

PM (x) =
X
hx, vi v .
v∈C

72
 4.6. Exercises

Exercise 4.13. In the context of Fourier analysis described in Example

4.2.12 (see also Example 4.2.2), the formula
2
kf k22 = fb(n)
X

n∈Z

is called Parseval’s identity. (The more general equality obtained in Theorem

4.2.8 c) is also often called Parseval’s identity.)
a) Set f (t) = t for all t ∈ [−π, π]. Compute the Fourier coefficients of f .
b) Use a) and Parseval’s identity to show that
∞
1 π2
=
X
.
n=1 n2 6

c) Set g(t) = et for all t ∈ [−π, π]. Use Parseval’s identity to obtain a
formula for the sum of the series
∞
X 1
.
n=1 n2 +1

Exercise 4.14. Let H be the L2 -space on [−π, π] w.r.t. to the normalized

Lebesgue measure µ, as in Example 4.2.2. Set

Heven := {[f ] ∈ H : f is even} and Hodd := {[f ] ∈ H : f is odd} .

We recall that f : [−π, π] → C is called even if f (−t) = f (t) for all t, while
it is called odd if f (−t) = −f (t) for all t.
a) Show that Heven is a closed subspace of H and that (Heven )⊥ = Hodd .
Describe the orthogonal projection P of H on Heven .
Hint: It might be helpful to consider the map [f ] → [fe ], where fe(t) := f (−t).
b) Find an orthonormal basis for Heven and one for Hodd .

Exercise 4.15. Let T ∈ B(H). Assume H0 is a dense subspace of H which

is invariant under T , and let T0 ∈ B(H0 ) denote the restriction of T to H0 .
Further, assume there exists some S0 ∈ B(H0 ) such that
D E D E
T0 (x), y = x, S0 (y) for all x, y ∈ H0 .

Show that T ∗ = S, where S ∈ B(H) is the unique extension of S0 provided

by Theorem 3.3.2.

73
4. More on Hilbert spaces

Exercise 4.16. Show that the formula for (TK )∗ in Example 4.3.7 is correct.
Hint : Consider H0 = [g] : g ∈ C([a, b])} and use Exercise 4.15.
Exercise 4.17. Prove Proposition 4.3.10.
Exercise 4.18. Let v, w ∈ H and consider the linear operator Tv,w : H → H
defined by
Tv,w (x) := hx, vi w for all x ∈ H .
Note that Tv,w has rank one if v, w ∈ H \ {0}.
a) Show that Tv,w is bounded with norm kTv,w k = kvk kwk. Then show
that (Tv,w )∗ = Tw,v .
b) Show that every T ∈ B(H) which has rank one is of the form T = Tv,w
for some v, w ∈ H \ {0}.
c) Assume T ∈ B(H) is a finite-rank operator, T 6= 0. Show that T may
be written as a finite sum of rank one operators in B(H).
Hint : Start by picking an orthonormal basis for T (H).
d) Show that if T ∈ B(H) is a finite-rank operator, then so is T ∗ .
Exercise 4.19. Let T ∈ B(H) and let M be a closed subspace of H.
Show that
M is invariant under T if and only if M ⊥ is invariant under T ∗ .

Exercise 4.20. Let T ∈ B(H).

a) Show that ker(T ) = ker(T ∗ T ) and T ∗ (H) = (T ∗ T )(H) .
b) Assume T is normal, i.e., satisfies T ∗ T = T T ∗ . Show that

ker(T ∗ ) = ker(T ) and T ∗ (H) = T (H) .

c) Assume T is normal and has an eigenvalue λ. Show that λ is an

∗
eigenvalue of T ∗ , and that EλT = EλT .
d) Assume H has a countably infinite orthonormal basis B = {uj }j∈N
and let S ∈ B(H) be the right shift operator (w.r.t. B). Set T = S ∗ .
Show that T is not normal by checking that T T ∗ = S ∗ S = IH , while
T ∗ T = SS ∗ = P , where P is the orthogonal projection of H on {u1 }⊥ .
Check also that 0 is an eigenvalue for T , while 0 is not an eigenvalue
of T ∗ = S. (This shows that c) does not necessarily hold when T is not
normal.)
Finally, if you are in the right mood, show that S has no eigenvalues,
while every λ satisfying |λ| < 1 is an eigenvalue of T .

74
 4.6. Exercises

Exercise 4.21. Let H and K be Hilbert spaces over F, and let T ∈ B(H, K).
a) Show that there exists a unique operator T ∗ ∈ B(K, H) (called the
adjoint of T ) satisfying that
D E D E
T (x), y = x, T ∗ (y) for all x ∈ H and all y ∈ K.

b) Let T 0 ∈ B(H, K) and α, β ∈ F. Let also L be a Hilbert space over F, and

let S ∈ B(K, L), so that ST ∈ B(H, L). Show that the following properties
hold:
• i) (α T + β T 0 )∗ = α T ∗ + β T 0 ∗ ; ii) (ST )∗ = T ∗ S ∗ ; iii) (T ∗ )∗ = T ;
• iv) kT ∗ k = kT k ; v) kT ∗ T k = kT k2 .
Exercise 4.22. Let (X, A, µ) be a measure space. One says that (X, A, µ)
is semifinite when the following condition holds: if E ∈ A and µ(E) = ∞,
then there exists F ⊆ E, F ∈ A such that 0 < µ(F ) < ∞.
a) Show that (X, A, µ) is semifinite whenever it is σ-finite.
Assume from now on that (X, A, µ) is semifinite. Set H := L2 (X, A, µ).
Let f ∈ L∞ and consider the multiplication operator Mf ∈ B(H) defined in
Example 4.4.4.
b) Show that kMf k = kf k∞ .
c) Show that if Mf is self-adjoint, then f is real-valued µ-a.e. (As
observed in Example 4.3.6, the converse is true without any restriction on
(X, A, µ).)
Exercise 4.23. Assume P ∈ B(H) is a self-adjoint projection, i.e., it
satisfies that P ∗ = P = P 2 . Show that P is the orthogonal projection of H
on M := P (H) (which is closed subspace of H, cf. Proposition 3.2.11).
Exercise 4.24. Let H 6= {0}.
a) Assume T ∈ B(H)
D is self-adjoint.
E Deduce from Theorem 4.4.9 that
T = 0 if and only if T (x), x = 0 for all x ∈ H.
b) Suppose F = R. Give an example with H = R2 showing that the
equivalence in a) may fail when T is not self-adjoint.

D c) Assume
E F = C and let T ∈ B(H). Show that T = 0 if and only if
T (x), x = 0 for all x ∈ H.

Exercise 4.25. Show that the set B(H)sa := {T ∈ B(H) : T ∗ = T } is

closed in B(H).

75
4. More on Hilbert spaces

Exercise 4.26. Let H 6= {0}. If T ∈ B(H) is self-adjoint, we have seen that

WT ⊆ R; of course, if F = R, this gives no information on T as this inclusion
is then true for any T in B(H). We assume therefore in this exercise that
H 6= {0} is a Hilbert space over C.
Let T ∈ B(H). Then show that the following assertions are equivalent:

(i) T is self-adjoint;

(ii) WT ⊆ R;
D E
(iii) T (x), x ∈ R for all x ∈ H.

Exercise 4.27. A self-adjoint operator T in B(H) is called positive when

D E
T (x), x ≥ 0 for all x ∈ H , (4.6.1)

in which case we write T ≥ 0.

(We note that if F = C and T ∈ B(H) satisfies (4.6.1), then T is
automatically self-adjoint, as follows from the previous exercise.)
a) Let S ∈ B(H), and let R ∈ B(H) be self-adjoint.
Check that S ∗ S ≥ 0 and R 2 ≥ 0. Then show that

kSk ≤ 1 ⇔ (IH − S ∗ S) ≥ 0 .

b) Let M be a closed subspace of H. Check that PM ≥ 0.

c) Assume that T, T 0 ∈ B(H) are positive and λ ∈ [0, ∞).
Check that T + T 0 and λ T are also positive.
d) Show that the set of positive operators in B(H) is closed in B(H).

Exercise 4.28. Let H = L2 ([0, 1]) (with usual Lebesgue measure) and let
T = Mf be the self-adjoint operator in B(H) given by multiplication with
the function f (t) = t on [0, 1], cf. Example 4.4.4. Show that T has no
(complex) eigenvalues.

Exercise 4.29. Let (X, A, µ) be a semifinite measure space (cf. Exercise

4.22), and let f ∈ L∞ . Suppose that the multiplication operator Mf on
H = L2 (X, A, µ) is unitary. Then show that |f | = 1 µ-a.e.
(As observed in Example 4.5.7, the converse statement is true without any
restriction on (X, A, µ).)

76
 4.6. Exercises

Exercise 4.30. Assume H 6= {0} is separable (cf. Exercise 4.9) and infinite-
dimensional. Let then B be an orthonormal basis for H indexed by Z, say
B = {vk }k∈Z . One may then define the bilateral shift operator V : H → H
(w.r.t. B) by

V (x) = for all x ∈ H, i.e., by

X
hx, vk i vk+1
k∈Z

n
V (x) = n→∞
lim for all x ∈ H.
X
hx, vk i vk+1
k=−n

a) Show that V is a unitary operator in B(H).

b) Describe V as a multiplication operator when H = L2 ([−π, π]) (with
normalized Lebesgue measure µ) and vk (t) = eikt for every k ∈ Z.
c) Assume F = C. Let U : H → `2 (Z) denote the isomorphism of
Hilbert spaces defined in the proof of Theorem 4.5.9. Show that U V U ∗ is
the bilateral forward shift operator on `2 (Z).

Exercise 4.31. Let T be a bounded operator on a Hilbert space H 6= {0}.

Check that the following properties of WT and NT hold:
n o
(a) WT ∗ = λ : λ ∈ WT ; hence, NT ∗ = NT .

(b) WT contains all the possible eigenvalues of T .

(c) WαT +βIH = α WT + β for all α, β ∈ F.

(d) WU T U ∗ = WT , hence NU T U ∗ = NT , for every unitary U ∈ B(H).

(e) WT is a compact subset of F when H is finite-dimensional.

It can also be shown that WT is a convex subset of F. This result is called

the Toeplitz-Hausdorff Theorem, but the proof is beyond the scope of these
notes.

77
 CHAPTER 5

On compact operators

5.1 Introduction to compact operators

between normed spaces
We had a very brief encounter with compact operators at the end of Section
3.1. For the ease of the reader, we recall their definition. In this section, X
and Y will denote normed spaces, both over F, unless otherwise specified.
Definition 5.1.1. An operator T ∈ L(X, Y ) is called compact if the se-
quence {T (xn )}n∈N has a convergent subsequence in Y whenever {xn }n∈N is
a bounded sequence in X.
We set K(X, Y ) := {T ∈ L(X, Y ) : T is compact }.
To appreciate this definition, the concept of relative compactness for
subsets of a metric space will be helpful.
A subset A of a metric space (Z, d) is called relatively compact in Z
if its closure A is a compact subset of Z. (Some authors say precompact
instead of relatively compact.) Equivalently, and this may be taken as the
definition for our purposes, a subset A of Z is relatively compact in Z if
and only if every sequence in A has a subsequence which converges in Z. In
comparison, we recall that A is compact if and only if every sequence in A
has a subsequence which converges in A.
We also remark that A ⊆ Z is bounded whenever A is relatively compact
in Z: indeed, if A is not bounded, then we can pick (any) z ∈ Z and find a
sequence {an }n∈N in A such that d(an , z) > n for all n ∈ N; it is then rather
easy to see that {an }n∈N has no convergent subsequence in Z, so A is not
relatively compact.
Proposition 5.1.2. Let T ∈ L(X, Y ). Then T is compact if and only if
T (B) is relatively compact in Y whenever B is a bounded subset of X.
5. On compact operators

Proof. Assume first that T is compact and let B ⊆ X be bounded. We want

to show that T (B) is relatively compact in Y . So let {yn }n∈N be a sequence
in T (B). For each n ∈ N we may then write yn = T (xn ) for some xn ∈ B.
As the sequence {xn }n∈N lies in B, it is bounded. Hence, by compactness of
T , {yn }n∈N = {T (xn }n∈N has a convergence subsequence in Y . Thus, T (B)
is relatively compact, as desired.
Conversely, assume that T maps bounded subsets of X into relatively
compact subsets of Y . We want to show that T is compact. So let {xn }n∈N
be a bounded sequence in X. Set B := {xn : n ∈ N}. Then B is a bounded
subset of X, so T (B) = {T (xn ) : n ∈ N} is relatively compact in Y . As
{T (xn )}n∈N is a sequence in T (B), we can conclude that it has a convergent
subsequence in Y . Thus, T is compact, as desired. 
Corollary 5.1.3. Assume T ∈ L(X, Y ) is compact. Then T is bounded.
Thus, K(X, Y ) ⊆ B(X, Y ).
Proof. Set B := X1 . Since B is a bounded subset of X, we get from
Proposition 5.1.2 that T (B) is relatively compact subset of Y . This implies
that T (B) is bounded. Hence we can find M > 0 such that kT (x)k ≤ M
for all x ∈ X1 , and it follows that T is bounded with kT k ≤ M . 
As we have seen previously in Section 3.1, cf. Proposition 3.1.9, an
important class of compact operators consists of the finite-rank operators in
B(X, Y ).
Example 5.1.4. Consider X = C([0, 1], R) with the uniform norm k · ku .
For g ∈ X, define T (g) : [0, 1] → R by
Z 1
[T (g)](s) = sin(s − t) g(t) dt for all s ∈ [0, 1] .
0

Since sin(s − t) = sin(s) cos(t) − cos(s) sin(t), we have that

Z 1  Z 1 
[T (g)](s) = cos(t) g(t) dt sin(s) − sin(t) g(t) dt cos(s)
0 0

for all s ∈ [0, 1]. It follows that T (g) ∈ X. Moreover, the map T : X → X
sending g to T (g) is clearly linear. As T (X) is 2-dimensional, T has finite-
rank. Further, since
Z 1 Z 1
[T (g)](s) ≤ | sin(s − t)g(t)| dt ≤ |g(t)| dt ≤ kgku
0 0

for all s ∈ [0, 1], we get that kT (g)ku ≤ kgku for all g ∈ X. Hence, T is
bounded. We can therefore conclude that T is compact.

80
 5.1. Introduction to compact operators between normed spaces

More generally, using the Arzelà-Ascoli Theorem (cf. Lindstrøm’s book),

it can be shown that if a function K : [a, b] × [c, d] → R is continuous, then
the associated integral operator T : C([c, d], R) → C([a, b], R), defined by
Z d
[T (g)](s) = K(s, t) g(t) dt for all s ∈ [a, b] ,
c
is compact.
Theorem 5.1.5. K(X, Y ) is a subspace of B(X, Y ). Moreover, if Y is
a Banach space, then K(X, Y ) is closed in B(X, Y ), and it follows that
K(X, Y ) is a Banach space.
Proof. We leave the proof of the first assertion as an exercise. So assume that
Y is Banach space, and let {Tn }n∈N be a sequence in K(X, Y ) converging
to some T ∈ B(X, Y ). To show that K(X, Y ) is closed in B(X, Y ), we have
to show that T is compact.
So let {xn }n∈N be a bounded sequence in X. Choose M > 0 such that
kxn k ≤ M for all n ∈ N.
• Since T1 is compact, there exists a subsequence {xnk }k∈N of {xn }n∈N
such that T1 (xnk ) → y1 as k → ∞ for some y1 ∈ Y .
We set x1,k := xnk for each k ∈ N. We then have T1 (x1,n ) → y1 as
n → ∞.
• Similarly, since {x1,n }n∈N is bounded and T2 is compact, we can find a
sequence {x2,n }n∈N , which is a subsequence of {x1,n }n∈N , and therefore
of {xn }n∈N , such that T2 (x2,n ) → y2 as n → ∞ for some y2 ∈ Y .
• Proceeding inductively, for each m ∈ N, m ≥ 2, we can find a sequence
{xm,n }n∈N , which is a subsequence of {xm−1,n }n∈N , and therefore of
{xn }n∈N , such that Tm (xm,n ) → ym as n → ∞ for some ym ∈ Y .
We now set x0k := xk,k ∈ X for each k ∈ N, and claim that
{T (x0k )}k∈N is a Cauchy sequence in Y . (5.1.1)
Since Y is complete, we will then be able to conclude that {T (x0k )}k∈N has
a convergent subsequence. As this subsequence will then be a subsequence
of the sequence {T (xn )}n∈N , we will thereby have shown that T is compact.
To establish (5.1.1), we first observe that for any k, l, m ∈ N, we have
kT (x0l ) − T (x0k )k ≤ k(T − Tm )(x0l ) + Tm (x0l ) − Tm (x0k ) + (Tm − T )(x0k )k
≤ k(T − Tm )(x0l )k + kTm (x0l ) − Tm (x0k )k + k(Tm − T )(x0k )k
≤ kT − Tm k kx0l k + kTm (x0l ) − Tm (x0k )k + kTm − T k kx0k k
≤ kTm (x0l ) − Tm (x0k )k + 2M kT − Tm k .

81
5. On compact operators

Let then ε > 0 and choose m ∈ N such that kT − Tm k < ε/3M . By

construction, for each k ≥ m, we have that Tm (x0k ) = Tm (xk,k ) is an element
of the sequence {Tm (xm,n )}n∈N , which is convergent to ym . It follows that
the sequence {Tm (x0k )}k∈N is convergent, hence that it is Cauchy. So we can
pick N ∈ N such that kTm (x0l ) − Tm (x0k )k < ε/3 for all k, l ≥ N . This gives
that

kT (x0l )−T (x0k )k ≤ kTm (x0l )−Tm (x0k )k+2M kT −Tm k < ε/3+2M (ε/3M ) = ε

for all k, l ≥ N . Hence we have shown that the claim (5.1.1) is true.
Finally, as Y is a Banach space, we know that B(X, Y ) is a Banach
space too, and this implies that K(X, Y ), being closed in B(X, Y ), is also a
Banach space. 

An immediate consequence is the following:

Corollary 5.1.6. Assume Y is a Banach space and set

F(X, Y ) := {T ∈ B(X, Y ) : T has finite-rank }.

Then we have
F(X, Y ) ⊆ K(X, Y ).

Example 5.1.7. Let 1 ≤ p < ∞ and set X := `p (N), which we know is

a Banach space w.r.t. k · kp . For each λ ∈ `∞ (N), we may consider the
multiplication operator Mλ ∈ B(X) given by

[Mλ (x)](n) = λ(n) x(n)

for all x ∈ X and all n ∈ N. One readily checks that kMλ k = kλk∞ .
Now, assume that λ ∈ c0 (N), i.e., limn→∞ λ(n) = 0. Then Mλ is compact.
Indeed, for each k ∈ N, let λ(k) ∈ `∞ (N) be defined by

λ(n) if 1 ≤ k ≤ n,
λ(k) (n) =
 0 otherwise,

for every n ∈ N. Then it is clear that each Mλ(k) has finite-rank; moreover,

kMλ − Mλ(k) k = kλ − λ(k) k∞ → 0 as k → ∞.

Thus Mλ ∈ F(X, X) ⊆ K(X, X).

82
 5.2. On compact operators on Hilbert spaces

We set K(X) = K(X, X), so that K(X) is a subspace of B(X). If

X is finite-dimensional, then every operator in B(X) has finite-rank, so
K(X) = B(X). On the other hand, if X is infinite-dimensional, then
K(X) 6= B(X), the reason being that the identity operator IX is not compact
in this case: indeed, if X is infinite-dimensional, then IX (X1 ) = X1 is closed,
but not compact, (cf. Exercise 3.2).
We also mention (cf. Exercise 5.1) that K(X) is a two-sided ideal in
B(X), meaning that we have

ST ∈ K(X) if S ∈ B(X) and T ∈ K(X), or if S ∈ K(X) and T ∈ B(X).

This property implies that no operator in K(X) can have a bounded inverse
when X is infinite-dimensional (for if T ∈ K(X) has an inverse T −1 ∈ B(X),
then we must have that IX = T −1 T ∈ K(X), so dim(X) < ∞).
We end this section with an interesting result concerning the possible
eigenvalues of a compact operator.

Theorem 5.1.8. Let T ∈ K(X). Then the following facts hold:

(a) Let δ > 0. Then {λ ∈ F : λ is an eigenvalue of T and |λ| > δ} is a
finite set.
(b) If λ ∈ F is a non-zero eigenvalue of T , then the associated eigenspace
is finite-dimensional.
(c) The set of eigenvalues of T (which may be empty) is countable and
bounded. If this set is countably infinite and {λk : k ∈ N} is an enumeration
of it, then limk→∞ λk = 0.

As we will be mostly interested in compact self-adjoint operators acting

on Hilbert spaces in this course, for which much more can be said (cf.
Theorem 5.3.4), we skip the proof of thi more general theorem.

5.2 On compact operators on Hilbert spaces

In view of Corollary 5.1.6, it is natural to wonder whether any compact
operator from a normed space to a Banach space may be approximated in
operator norm by bounded finite-rank operators. This problem was open
until 1973, when a counterexample was exhibited by P. Enflo. Happily, the
situation is as nice as possible when the target space is a Hilbert space.

Theorem 5.2.1. Let X be a normed space and H be a Hilbert space (both

over F). Then we have

F(X, H) = K(X, H) .

83
5. On compact operators

Proof. By Corollary 5.1.6, we only have to show that K(X, H) ⊆ F(X, H).
So let T ∈ K(X, H), and let ε > 0. We need to prove that there exists
S ∈ F(X, H) such that kT − Sk ≤ ε. Clearly, we can assume T 6= 0.
Set A := T (X1 ). Since X1 is bounded and T is compact, the set A
is compact in H. As H is a metric space, this implies that A is totally
bounded (cf. Proposition 3.5.12 in Lindstrøm’s book). Hence we can cover
A with some open balls B1 , . . . , Bn of radius ε/4, having respective centers
a1 , . . . , an ∈ A. For each j = 1, . . . , n, we can then find xj ∈ X1 such that
kaj − T (xj )k < ε/4.
Set now F := Span ({T (x1 ), . . . , T (xn )}), which is a finite dimensional
subspace of H, and let PF denote the orthogonal projection of H on F . Since
the range of PF T is contained in F , PF T has finite-rank, so PF T ∈ F(X, H).
We claim that
kT − PF T k ≤ ε .
Indeed, let x ∈ X1 . Then T (x) ∈ A, so T (x) ∈ Bj for some j ∈ {1, . . . , n}.
Hence,

kT (x) − T (xj )k ≤ kT (x) − aj k + kaj − T (xj )k < ε/4 + ε/4 = ε/2 .

Further, since T (xj ) ∈ F , we have that PF (T (xj )) = T (xj ). Thus, using

also that kPF k = 1, we obtain that

k(T − PF T )(x)k = kT (x) − T (xj ) + (PF T )(xj ) − (PF T )(x)k

 
≤ kT (x) − T (xj )k + kPF T (xj ) − T (x) k
≤ kT (x) − T (xj )k + kPF kkT (xj ) − T (x)k
= 2 kT (x) − T (xj )k
< 2 · ε/2 = ε .

As this holds for every x ∈ X1 , the claim follows. Hence, setting S := PF T ,

we are done. 
Remark 5.2.2. Let X be a normed space and H be a Hilbert space, and
let T ∈ K(X, H). Then it can be shown that T (X) is separable. We leave
this as an exercise.
Theorem 5.2.1 immediately gives:
Corollary 5.2.3. Let H be a Hilbert space. Set K(H) := K(H, H) and
F(H) := F(H, H). Then we have

F(H) = K(H) .

84
 5.2. On compact operators on Hilbert spaces

An application of this result is the following:

Corollary 5.2.4. Let H be a Hilbert space and let T ∈ K(H). Then
T ∗ ∈ K(H). In other words, K(H) is closed under the adjoint operation.
Proof. Using Corollary 5.2.3, we can find a sequence {Tn } in F(H) such
that kT − Tn k → 0 as n → ∞. Now, it is not difficult to see that F(H)
is closed under the adjoint operation (cf. Exercise 4.18). Hence, {Tn∗ } is a
sequence in F(H), and we have
kT ∗ − Tn∗ k = k(T − Tn )∗ k = kT − Tn k → 0 as n → ∞.
Thus, T ∗ ∈ F(H) = K(H). 
We recall from the previous section that if H is finite-dimensional, then
K(H) = B(H) = F(H), while K(H) 6= B(H) if H is infinite-dimensional.
An elementary argument showing that IH is not compact can be given
in this case: letting√then {uj }j∈N be any orthonormal sequence in H, we
have kuj − uk k = 2 for all j, k ∈ N, and it follows that the sequence
{IH (uj )}j∈N = {uj }j∈N does not have any convergent subsequence.
Let H be an infinite-dimensional Hilbert space H. An interesting class of
compact operators on H containing F(H) consists of the so-called Hilbert-
Schmidt operators. For simplicity, we only consider the case where H is
separable. We note that every orthonormal basis for H is then countable:
indeed, assume (for contradiction)
√ that H had an uncountable orthonormal
basis B. Then, as ku − u0 k = 2 for all distinct u, u0 ∈ B, we see that
any dense subset of H would have to be uncountable, contradicting the
separability of H.
Lemma 5.2.5. Assume H is a separable, infinite-dimensional Hilbert space
(over F). Let B = {uj }j∈N and C = {vj }j∈N be orthonormal bases for H,
and let T ∈ B(H). Then we have
∞ ∞
kT (uj )k2 = kT (vj )k2 .
X X

j=1 j=1

Proof. Using Parseval’s identity (two times), we get

∞ ∞ X
∞ D E2 ∞ X
∞ D E2
kT (uj )k2 = T (uj ), vk = uj , T ∗ (vk )
X X X

j=1 j=1 k=1 j=1 k=1

∞ X ∞ D E2 ∞ X ∞ D E2
= T ∗ (vk ), uj = T ∗ (vk ), uj
X X

j=1 k=1 k=1 j=1

∞
= kT (vk )k2 .
∗
X

k=1

85
5. On compact operators

Note that the change of order of summation at the second but last step
above is allowed since we are dealing with sums of non-negative numbers.
Applying what we have done to the case where B = C, i.e., uj = vj for every
j ∈ N, we get that
∞ ∞
kT (vj )k2 = kT ∗ (vk )k2 .
X X

j=1 k=1

Thus we obtain that

∞ ∞ ∞
kT (uj )k2 = kT ∗ (vk )k2 = kT (vj )k2 ,
X X X

j=1 k=1 j=1

as desired. 

Remark 5.2.6. An analogous result is true when H is finite-dimensional

and B, C are orthonormal bases for H.

Definition 5.2.7. Let H be a separable, infinite-dimensional Hilbert space

(over F). An operator T ∈ B(H) is called an Hilbert-Schmidt operator when
we have ∞
kT (uj )k2 < ∞
X

j=1

for some orthonormal basis B = {uj }j∈N of H, in which case we set

∞
X 1/2
kT k2 := kT (uj )k2 .
j=1

Lemma 5.2.5 shows that the definition of T being a Hilbert-Schmidt operator,

and the value of kT k2 , do not depend on the choice of orthonormal basis for
H. We also set

HS(H) := {T ∈ B(H) : T is a Hilbert-Schmidt operator} .

Proposition 5.2.8. Let H be a separable, infinite-dimensional Hilbert space

(over F).
Then HS(H) is a subspace of K(H), which contains F(H) and is closed
under the adjoint operation.
Moreover, the map T → kT k2 is a norm on HS(H), which satisfies

kT k ≤ kT k2

for every T ∈ HS(H).

86
 5.2. On compact operators on Hilbert spaces

Proof. We first note that it is evident from the proof of Lemma 5.2.5 that
T ∗ ∈ HS(H) whenever T ∈ HS(H).
Let B = {uj }j∈N be an orthonormal basis for H, and let T, T 0 ∈ HS(H).
Define ξ, ξ 0 ∈ `2 (N) by
ξ(j) := kT (uj )k and ξ 0 (j) := kT 0 (uj )k for each j ∈ N,
so that kξk2 = kT k2 and kξ 0 k2 = kT 0 k2 . Using the triangle inequality, first
in H, and then in `2 (N), we get
∞ ∞  2
k(T + T )(uj )k ≤
0 2
kT (uj )k + kT 0 (uj )k = kξ + ξ 0 k22
X X

j=1 j=1

≤ (kξk2 + kξ 0 k2 )2 = (kT k2 + kT 0 k2 )2 < ∞ .

This shows that T + T 0 ∈ HS(H) and
kT + T 0 k2 ≤ kT k2 + kT 0 k2 .
Moreover, one easily checks that λT ∈ HS(H) and kλT k2 = |λ| kT k2 for
every λ ∈ F. If kT k2 = 0, then we get that kT (uj )k = 0 for every j ∈ N,
and this clearly implies that T = 0.
Hence, we have shown so far that HS(H) is a subspace of B(H) which
is closed under the adjoint operation, and that k · k2 is a norm on HS(H).
To show that kT k ≤ kT k2 , let x ∈ H \ {0}. Set v1 = kxk 1
x and let
{vj }j≥2 be an orthonormal basis for {x} . Then {vj }j∈N is an orthonormal
⊥

basis for H, so we get

∞
kT (x)k2 = kxk2 kT (v1 )k2 ≤ kxk2 kT (vj )k2 = kT k22 kxk2 .
X

Thus, kT k ≤ kT k2 . j=1

Next, we show that T ∈ K(H). For each n ∈ N, let Pn denote the

orthogonal projection of H on Span ({u1 , . . . un }) and set Tn := T Pn . Then
we have ∞ n
kTn (uj )k2 = kT (uj )k2 < ∞ ,
X X

j=1 j=1

so Tn ∈ HS(H) for each n ∈ N. Hence,

∞
 X 1/2
kT − Tn k ≤ kT − Tn k2 = kT (uj )k2 → 0 as n → ∞ .
j=n+1

Since Tn ∈ F(H) for each n, Theorem 5.1.6 gives that T ∈ K(H). Hence,
HS(H) ⊆ K(H).
It only remains to show that F(H) ⊆ HS(H), but we leave this as an
exercise. 

87
5. On compact operators

Remark 5.2.9. For additional properties of HS(H), see Exercise 5.6.

Remark 5.2.10. If H 6= {0} is finite-dimensional and B = {uj }nj=1 is an

orthonormal basis for H, then we get a norm on B(H) by setting
n
X 1/2
kT k2 := kT (uj )k2
j=1

(which does not depend on the choice of orthonormal basis for H).
Letting A = [ai,j ] denotes the matrix of T w.r.t. B, one readily checks
that
n
 X 1/2
kT k2 = |ai,j |2 ,
i,j=1

i.e., kT k2 coincides with the so-called Fröbenius-norm of A.

Example 5.2.11. Set H = L2 ([a, b], A, µ), where A denotes the Lebesgue
measurable subsets of a closed interval [a, b] and µ is the Lebesgue measure
on A. Let K : [a, b] × [a, b] → C be a continuous function and let TK ∈ B(H)
denote the associated integral operator on H, which is the extension of the
integral operator TK : C([a, b]) → C([a, b]) given by
Z b
[TK (f )](s) = K(s, t) dt for f ∈ C([a, b]) and s ∈ [a, b].
a

cf. Example 3.3.6 and Exercise 3.18. Then TK is a Hilbert-Schmidt operator

on H (so TK is compact by Proposition 5.2.8).
To show this, we start by picking an orthonormal basis B = {[uj ]}j∈N for
H, where each uj is a continuous functions on [a, b]. (One may for example
construct B by applying the Gram-Schmidt orthonormalization process to
the monomials {tj−1 : j ∈ N}). We note that B := {[ uj ]}j∈N is then also an
orthonormal basis for H.
Let now s ∈ [a, b] and let ks ∈ C([a, b]) be given by ks (t) = K(s, t) for
all t ∈ [a, b]. Note that for each j ∈ N, we have
Z b Z D E
[TK (uj )](s) = K(s, t) uj (t) dt = ks (t) uj (t) dµ(t) = [ks ], [uj ] .
a [a,b]

Moreover, Parseval’s identity gives that

∞ D
X E 2 1/2
k [ks ] k2 = [ks ], [uj ] .
j=1

88
 5.2. On compact operators on Hilbert spaces

Thus, we obtain that

∞ ∞ D
2 E2
[TK (uj )](s) = [ks ], [uj ] = k [ks ] k22 .
X X

j=1 j=1

Now, using this and the Monotone Convergence Theorem, we get

∞ ∞ Z
2
k TK ([uj ]) k22 = [TK (uj )](s) dµ(s)
X X

j=1 j=1 [a,b]

Z ∞
X 2
= [TK (uj )](s) dµ(s)
[a,b] j=1
Z
= k [ks ] k22 dµ(s)
[a,b]
Z Z 
= |ks (t)|2 dµ(t) dµ(s)
[a,b] [a,b]
Z bZ b
= |K(s, t)|2 dt ds < ∞,
a a

which shows that TK ∈ HS(H) with kTK k2 ≤ |K(s, t)|2 ds dt.

RbRb
a a

In the previous example, one may allow the kernel K to be discontinuous

and still obtain an Hilbert-Schmidt operator TK , as long as K is square-
integrable w.r.t. to the product Lebesgue measure on [a, b] × [a, b]. However,
this requires a better knowledge of measure theory than the one we preassume
in these notes.

89
5. On compact operators

5.3 The spectral theorem for a compact

self-adjoint operator
Throughout this section we let H denote a Hilbert space (over F) different
from {0}. Our main goal is to generalize the spectral theorem for symmetric
real matrices known from linear algebra, and prove that every compact
self-adjoint compact operator T on H is diagonalizable in the sense that
there exists an orthonormal basis for H consisting of eigenvectors of T .
We begin with a series of lemmas.

Lemma 5.3.1. Assume T ∈ K(H) has a nonzero eigenvalue λ ∈ F. Then

the associated eigenspace Eλ := ker(T − λI) is finite-dimensional.

Proof. Assume for contraction that Eλ is infinite-dimensional. We may then

find a sequence {vn }n∈N of unit vectors in Eλ which are pairwise orthogonal.
By compactness of T , {T (vn )}n∈N has a convergent subsequence. So we may
as well assume that {T (vn )}n∈N is convergent, hence that it is a Cauchy
sequence. However, we have that

kT (vn ) − T (vm )k2 = kλ vn − λ vm k2 = 2 |λ|2 6= 0

for all m, n ∈ N. So {T (vn )}n∈N is not a Cauchy sequence, giving a contra-

diction. 

Lemma 5.3.2. Let T ∈ B(H) be self-adjoint, and assume T has an eigen-

value λ ∈ F. Then λ ∈ R.
Moreover, if λ0 is an eigenvalue of T distinct from λ, then Eλ ⊥ Eλ0 ,
i.e., hx, yi = 0 whenever x ∈ Eλ and y ∈ Eλ0 .

Proof. Let x ∈ Eλ . If kxk = 1, then we have

λ = λ hx, xi = hλ x, xi = hT (x), xi ∈ WT ⊆ R ,

so λ ∈ R. Moreover, assume that λ0 is an eigenvalue of T distinct from λ,

and let y ∈ Eλ0 . Then we have that λ0 ∈ R, so

λ hx, yi = hT (x), yi = hx, T (y)i = λ0 hx, yi .

Since λ0 6= λ, we get that hx, yi = 0. 

Lemma 5.3.3. Let T ∈ K(H) be self-adjoint. Then T has an eigenvalue

λ ∈ R such that |λ| = kT k.

90
 5.3. The spectral theorem for a compact self-adjoint operator

Proof. If T = 0, then the assertion is trivial. So assume that T 6= 0. Using

Theorem 4.4.9, we can find a sequence {xn }n∈N of unit vectors in H such
that |hT (xn ), xn i| → kT k as n → ∞. Since hT (xn ), xn i ∈ R for every n, we
can assume (by passing to a subsequence and relabelling) that

hT (xn ), xn i → λ as n → ∞, where λ = ±kT k. (5.3.1)

Moreover, since T is compact, we can also assume (by passing again to a

subsequence and relabelling) that T (xn ) → y as n → ∞ for some y ∈ H.
Note that the Cauchy-Schwarz inequality gives that

hT (xn ), xn i ≤ kT (xn )k for every n ∈ N,

so, letting n → ∞, we get that kyk ≥ |λ| > 0, so y 6= 0.

Now, using that T is self-adjoint, λ is real, kxn k = 1, and (5.3.1), we get
D E
kT (xn ) − λxn k2 = T (xn ) − λxn , T (xn ) − λxn
= kT (xn )k2 − 2λ hT (xn ), xn i + λ2 kxn k2
≤ kT k2 − 2λ hT (xn ), xn i + λ2
 
= 2λ λ − hT (xn ), xn i
→ 0 as n → ∞.

Thus, kT (xn ) − λxn k → 0 as n → ∞, and this gives that

ky − λxn k ≤ ky − T (xn )k + kT (xn ) − λxn k → 0 as n → ∞.

Hence,
T (y) = lim T (λxn ) = λ lim T (xn ) = λ y .
n→∞ n→∞

Since y 6= 0, λ is an eigenvalue of T , as we wanted to show.



We are now ready for the spectral theorem for a compact self-adjoint
operator T . Intuitively, we could hope to be able to construct an orthonormal
basis of eigenvectors for T by using Lemma 5.3.3 repeatedly as follows. Start
by picking a unit eigenvector v0 of T associated to the eigenvalue λ0 = ±kT k.
Next, consider the restriction T1 of T to {v0 }⊥ , and pick a unit eigenvector
v1 of T1 associated to the eigenvalue λ1 = ±kT1 k. Then continue this process
inductively. There are several technicalities involved in working out the
details of this approach. We will follow a more pedestrian route, which also
provides more information about T .

91
5. On compact operators

Theorem 5.3.4. Let T ∈ K(H) be self-adjoint. Then there exists an

orthonormal basis E for H which consists of eigenvectors of T .
More precisely, the following facts hold when T 6= 0:

(a) The set L consisting of all nonzero

h eigenvalues
i of T is a nonempty,
countable subset of the interval −kT k, kT k , containing kT k or −kT k.

(b) If L is countably infinite, and {λk : k ∈ N} is an enumeration of L,

then we have limk→∞ λk = 0.

(c) The eigenspace Eλ := ker(T − λI) is finite-dimensional for each λ ∈ L.

(d) For each λ ∈ L, let Eλ be an orthonormal basis for Eλ , and set

E 0 :=
[
Eλ .
λ∈L

Then E 0 is an orthonormal basis for T (H) = ker(T )⊥ , which is count-

able.

(e) If ker(T ) = {0}, set E0 := ∅ ; otherwise, let E0 be an orthonormal

basis for ker(T ). Then E := E0 ∪ E 0 is an orthonormal basis for H
which consists of eigenvectors of T .

(f ) Let Pλ denote the orthogonal projection of H on Eλ for each λ ∈ L.

Then Pλ Pλ0 = 0 whenever λ 6= λ0 belong to L. Moreover, T has a
spectral decomposition

T = (w.r.t. operator norm), (5.3.2)

X
λ Pλ
λ∈L
meaning that

– T = λ Pλ if L is finite ;
P
λ∈L

– limn→∞ kT − λk Pλk k = 0 if L is countably infinite

Pn
k=1

and {λk : k ∈ N} is an enumeration of L, as in (b).

Proof. We can clearly assume that T 6= 0.

(a): The set L is a subset of R by Lemma 5.3.2, which contains kT k or
−kT k by Lemma 5.3.3. If λ ∈ L, and v is an associated eigenvector in H1 ,
we have
|λ| = |hλv, vi| = |hT (v), vi| ≤ kT k .
Thus, L ⊆ [−kT k, kT k]
To show that L is countable, let ε > 0 and consider the subset of L given
by Lε := {λ ∈ L : |λ| ≥ ε}. Then Lε is finite.

92
 5.3. The spectral theorem for a compact self-adjoint operator

Indeed, assume Lε is nonempty. Then for each λ ∈ L, we can pick

vλ ∈ H1 such that T (vλ ) = λvλ ; for λ, λ0 ∈ Lε , λ =
6 λ0 , we then have
λvλ ⊥ λ0 vλ0 by Lemma 5.3.2, so we get

kT (vλ ) − T (vλ0 )k2 = kλvλ − λ0 vλ0 k2 = |λ|2 + |λ0 |2 ≥ 2ε2 .

Hence, if Lε was infinite, we could find a sequence in H1 which T maps into

a sequence with no convergent subsequence, contradicting the compactness
of T . Thus, Lε is finite.
Now, since L = n∈N L1/n , it follows that L is countable.
S

(b): Assume L is countably infinite and {λk : k ∈ N} is an enumeration

of L. Let ε > 0 be given. Then, as in (a), we get that the set K := {k ∈ N :
|λk | ≥ ε} is finite. So there exists N ∈ N such that K ⊆ {1, . . . , N }. For
every k ≥ N + 1, we then have |λk | < ε. This shows that limk→∞ λk = 0.
(c): This is a consequence of Lemma 5.3.1.
(d): We first remark that since T is self-adjoint, we have

T (H) = T ∗ (H) = (ker T )⊥ .

Next, it follows from Lemma 5.3.2 that Eλ ⊥ Eλ0 whenever λ 6= λ0 belong to

L. So it is clear that E 0 is an orthonormal set in H, which is countable since
each Eλ is finite and L is countable. Hence, E 0 is a countable orthonormal
basis for M := Span (E 0 ), and it remains only to show that M = ker(T )⊥ ,
i.e., that M ⊥ = ker(T ).

• ker(T ) ⊆ M ⊥ : Assume y ∈ ker(T ). Then for each λ ∈ L and v ∈ Eλ ,

we have
λ hv, yi = hT (v), yi = hv, T (y)i = hv, 0i = 0 .
Since λ 6= 0, this shows that y ∈ (E 0 )⊥ = M ⊥ .

• M ⊥ ⊆ ker(T ): It is easy to check that M is invariant under T . Hence,

M ⊥ is invariant under T ∗ = T (cf. Exercise 4.19). We may therefore
consider the restriction S of T to M ⊥ . Then S ∈ K(M ⊥ ): if not, then
there would exist a bounded sequence in M ⊥ , hence in H, which is
mapped by S, hence by T , to a sequence with no convergent subse-
quence, contradicting the compactness of T . Moreover, S is self-adjoint
(this is an easy exercise).
Now, assume that S = 6 0. Then Lemma 5.3.3 gives that S has an
nonzero eigenvalue µ. This implies that µ is a nonzero eigenvalue of T ,
i.e., µ ∈ L. But if v ∈ M ⊥ is an eigenvector for S associated with µ,

93
5. On compact operators

we then have that v ∈ Eµ ⊆ M , so v ∈ M ∩ M ⊥ = {0}, contradicting

that v 6= 0 (since v is an eigenvector).
This means that S has to be 0. Thus we get T (y) = S(y) = 0 for all
y ∈ M ⊥ , as desired.

(e): If ker(T ) = {0}, then we get from (d) that E = E 0 is an orthonormal

basis for ker(T )⊥ = {0}⊥ = H. If ker(T ) 6= {0}, then we have E0 ⊆ ker(T )
and E 0 ⊆ ker(T )⊥ , so it is clear that E is an orthonormal set. Moreover, we
have that
H = Span (E) .
Indeed, let x ∈ H. Then we may write

x = x M + xM ⊥ ,

where xM ∈ M = Span (E 0 ) and xM ⊥ ∈ M ⊥ = ker(T ) = Span (E0 ). So

we may choose {xn }n∈N ⊆ Span (E 0 ) and {yn }n∈N ⊆ Span (E0 ) such that
limn→∞ xn = xM and limn→∞ yn = xM ⊥ . This gives that

lim (xn + yn ) = xM + xM ⊥ = x .
n→∞

Hence, x ∈ Span (E). This shows that E is an orthonormal basis for H.

(f ): The first assertion follows readily from the fact that Eλ ⊥ Eλ0
whenever λ = 6 λ0 , cf. Lemma 5.3.2. Next, we consider the case where L is
countably infinite and {λk : k ∈ N} is an enumeration of L, leaving the
easier case where L is finite to the reader.
For each k ∈ N, set nk := dim(Eλk ) < ∞, and let {vk,1 , . . . , vk,nk } be
an enumeration of Eλ0 k . Then we have

E0 = Eλ0 k = {vk,l : k ∈ N, 1 ≤ l ≤ nk }.
[

k∈N

Consider x ∈ H. Since T (x) ∈ T (H) and E 0 is an orthonormal basis for

T (H), we get from Corollary 4.2.11 that
nk
m X nk
m X
T (x) = m→∞
lim hT (x), vk,l i vk,l = m→∞
lim hx, T (vk,l )i vk,l
X X

k=1 l=1 k=1 l=1

m nk
X  ∞
= lim hx, vk,l i vk,l = λk Pλk (x) .
X X
λk
m→∞
k=1 l=1 k=1

Let now ε > 0. We have to show that there exists N ∈ N such that
kT − nk=1 λk Pλk k ≤ ε for all n ≥ N .
P

94
 5.3. The spectral theorem for a compact self-adjoint operator

Using (b), we can choose N ∈ N such that |λk | < ε for all k > N .
Then for all n ≥ N and all x ∈ H, using continuity of the norm in H and
Pythagoras’ identity, we get
n ∞ ∞
  2 2
λk Pλk (x) = λk Pλk (x) = |λk |2 kPλk (x)k2
X X X
T−
k=1 k=n+1 k=n+1
∞
2
kPλk (x)k2 ≤ ε2 kxk2
X
≤ε
k=n+1

and the assertion follows.



Remark 5.3.5. Let us say that an operator T ∈ B(H) is diagonalizable if

there exists an orthonormal basis for H whose elements are eigenvectors
for T . The spectral theorem says that T is diagonalizable if T is compact
and self-adjoint. A more precise statement is as follows. We recall that
T ∈ B(H) is called normal if T ∗ commutes with T .
Assume that T ∈ K(H). If F = R, then T is diagonalizable if and only
if T is self-adjoint. On the other hand, if F = C, then T is diagonalizable if
and only if T is normal.
We leave the proof to the reader (cf. Exercises 5.13 and 5.14).

As a corollary of the spectral theorem, an analogue of the singular value

decomposition for matrices may be obtained for compact operators.
Indeed, let S ∈ K(H), S 6= 0. Then T := S ∗ S is self-adjoint and
compact, and T = 6 0 (as kT k = kS ∗ Sk = kSk2 6= 0). Hence, the spectral
theorem gives that we may find a countable orthonormal basis {vj }j∈N for
T (H) = ker(T )⊥ = ker(S ∗ S)⊥ = ker(S)⊥ consisting of eigenvectors for T .
For each j ∈ N , let µj denote the eigenvalue of T associated with vj . Note
that
D E D E D E
µj = µj vj , vj = T (vj ), vj = S(vj ), S(vj ) = kS(vj )k2 ≥ 0

for every j ∈ N . Since each µj is nonzero, we get that all µj ’s are positive.
For each j ∈ N , set

√ 1
σj := µj > 0 and uj := S(vj ) .
σj

95
5. On compact operators

The σj ’s are called the singular values of S. For all j, k ∈ N we have

D E 1 D E 1 D E
uj , uk = S(vj ), S(vk ) = T (vj ), vk
σj σk σj σk

µj D E 1 if j = k,
= vj , vk =
σj σk 0 otherwise,

so {uj : j ∈ N } is an orthonormal set in the range of S. Further, we have

the following decomposition of S:
S(x) = for all x ∈ H . (5.3.3)
X
σj hx, vj i uj
j∈N

Indeed, let x ∈ H and set M := T (H), so M ⊥ = ker(S).

With z := x − PM (x) ∈ M ⊥ , we get that
x = PM (x) + z = hx, vj i vj + z ,
X

j∈N

so
S(x) = hx, vj i S(vj ) + S(z) =
X X
σj hx, vj i uj ,
j∈N j∈N

as asserted in (5.3.3).
It readily follows that {uj : j ∈ N } is an orthonormal basis for S(H).
√
Finally we remark that the spectral theorem also gives that σj = µj → 0
as j → ∞ when N is countably infinite, and that
kSk = kT k1/2 = max{ µj : j ∈ N }1/2 = max{ σj : j ∈ N } .

5.4 Application: The Fredholm Alternative

A useful application of linear algebra, and one of its original motivation,
is the study of systems of linear equations, i.e., of equations of the type
A x = b, where A ∈ Mm×n (F), b ∈ Fm and the (unknown) vector x belongs
to Fn . More generally, one may consider equations of the form
T (v) = w (5.4.1)
where V, W are vector spaces (over F), T ∈ L(V, W ), w ∈ W and the
(unknown) vector x belongs to V . Whether such an equation is consistent,
i.e., has some solution(s), relies on whether w lies in the range of T , in which
case it follows readily that the solution set of (5.4.1) is given by
n o
v0 + ker(T ) := v0 + u | u ∈ ker(T ) (5.4.2)

96
 5.4. Application: The Fredholm Alternative

where v0 ∈ V is any vector satisfying (5.4.1), i.e., such that T (v0 ) = w.

In the rest of this section, we consider the case where V = W = H is a
Hilbert space (6= {0}), and T ∈ B(H). We can then exploit the relationship
between the fundamental subspaces of T and T ∗ , cf. Proposition 4.3.8.
For example, using that T (H) = ker(T ∗ )⊥ , we get that if T has closed
range (i.e., T (H) is closed), then (5.4.1) will be consistent if and only if w
is orthogonal to ker(T ∗ ).
In particular, if T has closed range and ker(T ∗ ) = {0} (i.e., T ∗ is one-to-
one), then T must be surjective, hence (5.4.1) is consistent for all w ∈ H.
Similarly, if T ∗ has closed range and ker(T ) = {0}, then it follows that T ∗
is surjective, so the equation T ∗ (v 0 ) = w0 is consistent for all w0 ∈ H.
On the other hand, if it happens that T is surjective, then we get that
ker(T ∗ ) = {0}, hence that the equation T ∗ (v 0 ) = w0 will have either no
solution or a unique solution. Similarly, if T ∗ is surjective, then ker(T ∗ ) =
{0}, and (5.4.1) will have either no solution or a unique solution.
A problem is that many bounded operators do not have a closed range.
Moreover, in general, it may be a difficult task to decide whether the range
of some given T ∈ B(H) is closed or not. However, we note that if T ∈ B(H)
has finite-rank, then it has closed range (as T (H) is finite-dimensional).
In the case where H is finite-dimensional, much more can be said. The
following terminology will be useful.
Definition 5.4.1. An operator F ∈ B(H) is said to satisfy the Fredholm
alternative if one of the following two (mutually exclusive) situations occurs:
(a) ker(F ) = ker(F ∗ ) = {0}, and the equations F (v) = w, F ∗ (v 0 ) = w0
have both a unique solution for all w, w0 ∈ H;
(b) 1 ≤ dim(ker(F )) = dim(ker(F ∗ )) < ∞, the equation F (v) = w is
consistent if and only if w ∈ ker(F ∗ )⊥ , and the equation F ∗ (v 0 ) = w0
is consistent if and only if w0 ∈ ker(F )⊥ .
Example 5.4.2. Assume that H is finite-dimensional and F ∈ B(H), i.e.,
F ∈ L(H). Then F satisfies the Fredholm alternative.
The crux is that we have dim(ker(F ∗ )) = dim(ker(F )). To show this,
we use the formula

dim(M ) + dim(M ⊥ ) = dim(H),

which is easily verified for any subspace M of H, and the dimension formula
for F . We get that

dim(ker(F ∗ )) = dim(F (H)⊥ ) = dim(H) − dim(F (H)) = dim(ker(F )).

97
5. On compact operators

Combining this fact with our previous observations in this section, it is

straightforward to deduce that either (a) or (b) in Definition 5.4.1 holds.

An important class of bounded operators satisfying the Fredholm alter-

native consists of operators of the form F = T − µI, where T is a compact
operator on H and µ ∈ F \ {0}. In the special case where T = TK is an inte-
gral operator, cf. Example 5.2.11, an equation of the form (TK − µI)(f ) = g,
i.e., TK (f ) − µf = g, is often called a Fredholm integral equation of the
second kind.1
Consider T ∈ K(H) and µ ∈ F \ {0}. Then it can be shown that the
following facts hold:

(i) T − µI has closed range;

(ii) dim(ker(T − µI)) = dim(ker((T − µI)∗ )) < ∞.

Since T ∗ is compact, we also get that T ∗ − µI = (T − µI)∗ has closed

range. Using these properties, and the general principles outlined before,
one readily arrives at the conclusion that F = T − µI satisfies the Fredholm
alternative, as asserted above. We don’t have time in this course to prove
that (i) and (ii) hold. Instead, we will illustrate how the spectral theorem
for compact self-adjoint operators can be applied to give a direct proof of
the following:

Theorem 5.4.3. Assume T ∈ K(H) is self-adjoint and µ ∈ F \ {0}. Then

F = T − µI satisfies the Fredholm alternative.

Proof. Assume first that µ is not an eigenvalue of T , i.e., ker(T − µI) = {0}.
Then the spectral theorem implies that the equation (T − µI)(x) = y
has a unique solution for all y ∈ H. (You are asked to check this in
Exercise 5.9.) Thus, F = T − µI is surjective, and this implies that
ker(F ∗ ) = ker(T − µI) = {0}, i.e., µ is not an eigenvalue of T . Arguing as
above, we get that the equation (T − µI)(x0 ) = y 0 , i.e., (T − µI)∗ (x0 ) = y 0
has a unique solution for all y 0 ∈ H. This shows that (a) in Definition 5.4.1
holds in this case.
Next, assume that µ is an eigenvalue of T , i.e., ker(T − µI) 6= {0}. Then
µ ∈ R, so F ∗ = F . Moreover, as µ 6= 0, we have that T 6= 0, and the spectral
theorem tells us that 1 ≤ dim(ker(F )) = dim(ker(T − µI)) < ∞. Hence,
to show that (b) in Definition 5.4.1 holds, it remains only to prove that the
1
Such equations, and Fredholm integral equations of the first kind (i.e., equations of
the form TK (f ) = g), were studied by I. Fredholm at the beginning of the 20th century.
They arise in some practical problems in signal theory and in physics.

98
 5.4. Application: The Fredholm Alternative

equation F (x) = y is consistent if and only if y ∈ ker(F )⊥ . This means that

we have to prove that the equation
T (x) − µx = y (5.4.3)
is consistent if and only if hy, zi = 0 for all z ∈ Eµ := ker(T − µI).
To prove this, let E 0 = {uj }j∈J be an enumeration of the orthonormal
basis for T (H) obtained in the spectral theorem for T , and let µj ∈ R \ {0}
denote the eigenvalue of T corresponding to each uj .
Since H is the direct sum of ker(T ) and ker(T )⊥ = T (H), we may write
y ∈ H as
y = y0 + hy, uj i uj ,
X

j∈J

where y0 denote the orthogonal projection of y onto ker(T ). Likewise, we

may assume that the (unknown) vector x in equation (5.4.3) is written as
x = x0 +
X
cj uj ,
j∈J

where x0 ∈ ker(T ) and {cj }j∈J ∈ `2 (J) are to be determined, if possible.

Plugging this into equation (5.4.3), we get the equivalent equation
−µx0 + (µj − µ) cj uj = y0 +
X X
hy, uj i uj .
j∈J j∈J

Clearly, we can set x0 := (−1/µ) y0 , and equation (5.4.3) is then consistent

if and only if the sequence {cj }j∈J ∈ `2 (J) can be chosen so that
(µj − µ) cj = hy, uj i for all j ∈ J. (5.4.4)
Now, as µ is a nonzero eigenvalue of T , we have that µ = µk for some k ∈ J.
Let uj1 , . . . , ujn denote the vectors in E 0 giving an orthonormal basis for
Eµ = Eµk . If j 6∈ {j1 , . . . , jn }, we have µj 6= µ, so
1
cj := hy, uj i
µj − µ
will satisfy (5.4.4) for every such j.
On the other hand, if j ∈ {j1 , . . . , jn }, we have µj − µ = 0. Hence,
(5.4.4) will be satisfied for j = j1 , . . . , jn if and only if we have hy, uj i = 0
for j = j1 , . . . , jn , i.e., if and only if hy, zi = 0 for all z ∈ Eµ . Moreover,
when this condition holds, we can choose cj1 , . . . , cjn freely and, regardless
of this choice, the constructed sequence {cj }j∈J is easily seen to belong to
`2 (J) (exercise: check this!), meaning that the associated vector x gives a
solution to (5.4.3). Thus, we have proved the desired equivalence. 

99
5. On compact operators

5.5 Exercises
Exercise 5.1. Let X, Y, Z denote normed spaces over F. Consider λ ∈ F,
T, T 0 ∈ B(X, Y ) and S ∈ B(Y, Z), so ST ∈ B(X, Z).
a) Show that λ T + T 0 ∈ K(X, Y ) if T, T 0 ∈ K(X, Y ).
b) Show that ST ∈ K(X, Z) if T ∈ K(X, Y ).
c) Show that ST ∈ K(X, Z) if S ∈ K(Y, Z).
d) Set K(X) = K(X, X). Deduce that
ST ∈ K(X) if S ∈ B(X) and T ∈ K(X), or if S ∈ K(X) and T ∈ B(X).
Exercise 5.2. Let X = `p (N), λ ∈ `∞ (N), and Mλ ∈ B(X) be the
associated multiplication operator, cf. Example 5.1.7.
Show that λ ∈ c0 (N) if Mλ is compact.
(It therefore follows that Mλ is compact if and only if λ ∈ c0 (N).)
Exercise 5.3. Let X be a normed space, H be a Hilbert space, and let
T ∈ K(X, H). Show that T (X) is separable.
Exercise 5.4. Let H beD an infinite-dimensional
E Hilbert space and let
T ∈ K(H). Show that T (un ), un → 0 as n → ∞ whenever {un }n∈N
is an orthonormal sequence in H.
Exercise 5.5. Let H be a Hilbert space and let P ∈ B(H) be a projection
(i.e. P 2 = P ). Show that P has finite-rank if (and only if) P is compact.
Exercise 5.6. Let H be a separable Hilbert space, H 6= {0}.
a) Show that F(H) ⊆ HS(H), and that F(H) is dense in HS(H)
w.r.t. k · k2 .
b) Assume that T ∈ HS(H) and S ∈ B(H). Show that both ST and
T S belong to HS(H), and that we have

kST k2 ≤ kSk kT k2 , kT Sk ≤ kT k2 kSk .

c) Let B = {uj }j∈J be an orthonormal basis for H, where J = {1, . . . , n}

if dim(H) = n < ∞, while J = N otherwise.
For T, T 0 ∈ HS(H), set
D E X D E
T, T 0 := T (uj ), T 0 (uj ) .
2
j∈J

Show that this gives a well-defined inner product on HS(H), and check that
the associated norm is the Hilbert-Schmidt norm k · k2 .

100
 5.5. Exercises

d) Show that HS(H) is complete w.r.t. k · k2 , so that HS(H) is a Hilbert

space w.r.t. the inner product from c).

Exercise 5.7. Let H = L2 (R, A, µ) where A denote all Lebesgue measurable

subsets of R and µ is the Lebesgue measure. For which f ∈ L∞ is the
multiplication operator Mf ∈ B(H) compact ?

Exercise 5.8. Let H be a Hilbert space, T ∈ K(H) and λ ∈ F, λ 6= 0.

Assume that there exists a sequence {xn }n∈N of unit vectors in H such that
kT (xn ) − λ xn k → 0 as n → ∞. Show that λ is an eigenvalue of T .

Exercise 5.9. Let H be a Hilbert space, and let T ∈ K(H) be self-adjoint.

Assume µ ∈ F, µ 6= 0 is not an eigenvalue of T , i.e. T − µIH is injective.
Let y ∈ H, let E 0 = {uj }j∈J be an enumeration of the orthonormal basis
for M = T (H) obtained in the spectral theorem for T , and let µj 6= 0 denote
the eigenvalue of T corresponding to uj .
a) Show that the series

X hy, uj i
uj
j∈J µj − µ

converges to some h ∈ H.
b) Set z := y − PM (y) and x := h − 1
µ
z. Show that (T − µIH )(x) = y.

c) Deduce that T − µIH is surjective (hence that it is bijective).

Exercise 5.10. Consider H = L2 ([−π, π]) (with respect to the normalized

Lebesgue measure). Let g ∈ C([−π, π]) be periodic, i.e. satisfies that
g(−π) = g(π), and extend g to a periodic function g̃ on R with period 2π.
Define G : [−π, π] × [−π, π] → C by G(s, t) = g̃(s − t).
a) Check that G is continuous, so that the associated integral operator
TG belongs to HS(H) (hence is compact).
c) Decide when TG is self-adjoint.
b) Let k ∈ Z and recall that ek (t) = eikt for all t ∈ [−π, π]. Check that
ek is an eigenvector for the operator TG . Deduce that TG is diagonalizable
(with respect to {ek }k∈Z ).
 1/2
c) Show that kTG k2 = kgk2 = −π |g(t)| dt .
1 Rπ 2
2π

101
5. On compact operators

Exercise 5.11. Consider H = L2 ([0, 1]) (with respect to Lebesgue measure)

and the integral operator TK ∈ B(H) associated with K(s, t) := min(s, t)
for all (s, t) in [0, 1] × [0, 1], cf. Example 5.2.11.
a) Explain why TK is self-adjoint and compact. Then check that the set
U := {[un ] : n ∈ N}, where
√  1 
un (t) := 2 sin (n − )π t for all t ∈ [0, 1], n ∈ N ,
2
is an orthonormal set of eigenvectors for TK .
b) It can be shown that U is an orthonormal basis for H. Is it possible
to deduce this from a) and the spectral theorem for TK ?

Exercise 5.12. Let S, T ∈ B(H).

a) Assume there exists an orthonormal basis for H whose elements are
eigenvectors for both S and T . Check that S commutes with T .
b) Assume S and T are compact and self-adjoint, and that S commutes
with T . Show that there exists an orthonormal basis for H whose elements
are eigenvectors for both S and T .
Hint: Start by considering an eigenvalue λ of T and study how S acts
on the corresponding eigenspace EλT .

Exercise 5.13. Assume H is a Hilbert space over R, and let T ∈ B(H).

a) Assume that T is diagonalizable (as defined in Remark 5.3.5). Check
that T is self-adjoint.
b) Let T be compact. Deduce that T is diagonalizable if and only if T is
self-adjoint.

Exercise 5.14. Assume H is a Hilbert space over C, and let T ∈ B(H).

a) Assume that T is diagonalizable (as defined in Remark 5.3.5). Check
that T is normal.
b) Show that T is normal if and only if Re(T ) and Im(T ) commutes
with each other.
c) Let T be compact. Show that T is diagonalizable if and only if T is
normal.
Hint: The implication (⇒) follows from a). For (⇐), use b) and Exercise
5.12 b).

102
 5.5. Exercises

Exercise 5.15. Let H be a separable Hilbert space with a countably infinite

orthonormal basis B = {vj }j∈N . Let {µj }j∈N be a bounded sequence in F
and let D ∈ B(H) denote the associated diagonal operator (w.r.t. B).
a) Show that D is compact if and only if limj→∞ µj = 0.
(Note: If you have looked at Example 5.1.7 and solved Exercise 5.2, this
should not be difficult).
b) Show that D is Hilbert-Schmidt if and only if {µj }j∈N ∈ `2 (N), in
P 1/2
which case we have kDk2 = ∞
j=1 |µj |2 .

Exercise 5.16. Let H be a separable Hilbert space of infinite dimension

and let T ∈ K(H) be selfadjoint, T 6= 0. Assume that you have found an
orthonormal basis B = {vj }j∈N for H consisting of eigenvectors for T , and
let µj ∈ R denote the eigenvalue of T corresponding to each vj .
a) Show that the sequence {µj }j∈N is bounded, hence that T is the
diagonal operator (w.r.t. B) associated with this sequence. Deduce from the
previous exercise that limj→∞ µj = 0.
b) As in the spectral theorem, set

L := {λ ∈ R | λ is a nonzero eigenvalue of T }.

Set also
e := {λ ∈ R | λ is an eigenvalue of T },
L
so L = L
e \ {0}. Show the following assertions:

(i) L
e = {µ | j ∈ N} and L = {µ | j ∈ N, µ 6= 0}.
j j j

(ii) If λ ∈ L and Nλ := {j ∈ N | µj = λ}, then Nλ is a finite subset of N

and {vj | j ∈ Nλ } is an o.n.b. for Eλ .

(iii) If µj 6= 0 for all j ∈ N, then ker(T ) = {0}.

(iv) If N0 := {j ∈ N | µj = 0} is nonempty, then {vj | j ∈ N0 } is an

o.n.b. for ker(T ).

Exercise 5.17. Let H = L2 ([0, 1]) (with usual Lebesgue measure) and let
T = Mf be the self-adjoint operator in B(H) given by multiplication with
the function f (t) = t on [0, 1], cf. Example 4.4.4.
Show that T (H) is not closed, i.e., that T does not have closed range.
Show also that T is not compact.

103
5. On compact operators

Exercise 5.18. Let H = `2 (N), let λ ∈ `∞ (N) be given by λ(n) = n1 for

all n ∈ N, and let T = Mλ ∈ B(H) denote the associated multiplication
operator. Note that T is compact, as follows from Example 5.1.7.
Show that T (H) = H and T (H) 6= H, so T does not have closed range.

Exercise 5.19. Let H be a Hilbert space and T ∈ B(H). Let us say that T
is bounded from below if there exists some α > 0 such that α kxk ≤ kT (x)k
for all x ∈ H. For example, T is bounded from below when T is an isometry.
Show that if T is bounded from below, then T has closed range.

Exercise 5.20. Finish the proof of Theorem 5.9 by checking that the
sequence {cj }j∈J constructed in the final paragraph (under the assumption
that y is orthogonal to Eµ ) belongs to `2 (J).

104