1

Functional Analysis and Applications

Lecture notes for MATH 797fn

Luc Rey-Bellet
University of Massachusetts Amherst

The functional analysis, usually understood as the linear theory, can be described as

Extension of linear algebra to infinite-dimensional

vector spaces using topological concepts

The theory arised gradually from many applications such as solving boundary value problems, solving
partial differential equations such as the wave equation or the Schrödinger equation of quantum mechanics,
etc... Such problems lead to a comprehensive analysis of function spaces and their structure and of linear
(an non-linear) maps acting on function spaces. These concepts were then reformulated in abstract form in
the modern theory of functional analysis. Functional analytic tools are used in a wide range of applications,
some of which we will discuss in this class.
2
Chapter 1

Metric Spaces

1.1 Definitions and examples

One can introduce a topology on some set M by specifying a metric on M .

Definition 1.1. A map d(·, ·) : M × M → R is called a metric on the set M if for all ξ, η, ζ ∈ M we have

1. (positive definite) d(ξ, η) ≥ 0 and d(ξ, η) = 0 if and if ξ = η.

2. (symmetric) d(ξ, η) = d(η, ξ).

3. (triangle inequality) d(ξ, η) ≤ d(ξ, ζ) + d(ζ, ξ)

Example 1.2. Some examples of metric spaces

Rb
1. M = C[a, b] with d(f, g) = a |f (t) − g(t)|p dt with 0 < p < ∞.

2. M = C[a, b] with d(f, g) = maxt∈[a,b] |f (t) − g(t)|

3. For any measure space (X, µ), M = Lp (X, µ) with d(f, g) = kf − gkp with 1 ≤ p ≤ ∞.

4. Let M be the set of all infinite sequences ξ = (x1 , x2 , · · · ) with xi ∈ C. Then for η = (y1 , y2 , · · · )
∞  i
X 1 |xi − yi |
d(ξ, η) = (1.1)
i=1
2 1 + |xi − yi |

defines a metric on M .

5. Let M be the set of all infinite sequences of 0 and 1: M = {ξ = (x1 , x2 , · · · ) ; xi ∈ {0, 1}}. Then
∞
X
d(ξ, η) = (xi + yi (mod)2)
i=1
= number of indices j at which ξ and η differ (1.2)

defines a metric on M and is called the Hamming distance and is used in coding theory.

3
4 CHAPTER 1. METRIC SPACES

6. A metric can be defined on arbitrary space M (without any linear structure), for example set

0 if ξ = η
d(ξ, η) = (1.3)
1 if ξ 6= η

and d is called the discrete metric and M a discrete space.

In a metric space (X, d) ones introduces naturally a concept of convergence as well as a topology.
Definition 1.3. 1. A sequence {ξi } is called a Cauchy sequence if for any  > 0 there exists N so that
d(ξi , ξj ) <  for all i, j ≥ N (or shorter if limi,j→∞ d(ξi , ξj ) = 0).
2. We say that ξ is the limit of the sequence {ξi } (or that ξi converges to ξ, or that limi→∞ ξi = ξ) if
limi→∞ d(ξi , ξ) = 0.
Definition 1.4. A metric space (M, d) is called complete if every Cauchy sequence {ξi } has a limit ξ ∈ M .
Theorem 1.5. The following metric spaces are complete.
1. Let M be a finite dimensional vector space with (arbitrary norm) k · k and metric d(ξ, η) = kξ − ηk.
2. M = Lp (X, µ) with d(f, g) = kf − gkp
3. M = C[a, b] with d(f, g) = maxt∈[a,b] |f (t) − g(t)|.
Proof. Consult your class notes for Math 624.

Definition 1.6. A brief reminder on some topological concepts.

• In a metric space (X, d)
Br (ξ) := {η : d(ξ, η) < r} (1.4)
is the open ball of radius r around ξ and

B r (ξ) := {η : d(ξ, η) ≤ r} (1.5)

is the closed ball of radius r around ξ

• A set N ⊂ M is called bounded if there exist a ball B such that N ⊂ B.
• A set N ⊂ M is called open if for every point ξ ∈ N there exists a open ball around ξ contained in
N.
• A set N ⊂ M is called closed if M \ N is open.
• For a set N ⊂ M the set N is the smallest closed set which contains N . The set N is called the
closure of N .
• A set N ⊂ M is called dense (in M ) if N = M .
• A set K ⊂ M is called compact (in M ) if every sequence in K contains a convergent subsequence
with limit in K.
1.2. BANACH FIXED POINT THEOREM 5

1.2 Banach fixed point theorem

Problem 1.7. As a motivation imagine we want to solve the fixed point problem

F (ξ) = ξ (1.6)

where F : M → M is some map (not necessarily linear).

The idea of the solution is simple. Pick a point ξ0 and define the sequence ξn inductively by

ξn+1 = F (ξn ) . (1.7)

If this sequence converges to ξ and F is continuous we have then

ξ = lim ξn+1 = lim F (ξn ) = F ( lim ξn ) = F (ξ) (1.8)

n→∞ n→∞ n→∞

and so ξ is a solution of the fixed point problem.

We have the following
Theorem 1.8. (Banach Fixed Point Theorem) Let (M, d) be a complete metric space and let F : M → M
be a contraction, i.e., there exists q ∈ [0, 1) such that for all ξ, η ∈ M

d(F (ξ), F (η)) ≤ qd(ξ, η) . (1.9)

Then F has exactly one fixed point ξ = limn F n (ξ0 ) for arbitrary ξ0 .
Proof. We have
d(ξn+1 , ξn ) ≤ qd(F (ξn ), F (ξn−1 )) ≤ · · · ≤ q n d(ξ1 , ξ0 ) . (1.10)
Using that q < 1 we have then
m
X
d(ξn+m , ξn ) ≤ d(ξn+k , ξn+k−1 )
k=1
Xm
≤ q n+k−1 d(ξ1 , ξ0 )
k=1
n
q
≤ d(ξ1 , ξ0 ) . (1.11)
1−q
Since (M, d) is complete, ξ = limn→∞ ξn exists.
To show that ξ is a fixed point we note that

d(f (ξ), ξ) ≤ d(f (ξ), f (ξn )) + d(ξn+1 , ξ) ≤ qd(ξ, ξn ) + d(ξ, ξn+1 ) (1.12)

and the right hand side can be made arbitrarily small for n large enough.
Finally to show uniqueness, if ξ and η are two fixed points then

d(ξ, η) = d(F (ξ), F (η)) ≤ qd(ξ, η) < d(ξ, η) . (1.13)

and this implies ξ = η.

Let us give a few applications of this theorem.

6 CHAPTER 1. METRIC SPACES

Example 1.9. Let us try to solve the set of linear equation

n
X
aik xk = zk (1.14)
i=1

where zk , k = 1, · · · , n are given and xk , k = 1, · · · , n are unknown. We rewrite it as a fixed point equation
n
X
xi = (aik + δik )xk − zi (1.15)
k=1
or
ξ = F (ξ) (1.16)
n n
where ξ = (x1 , · · · , xn ) ∈ R (or C ) and

F (ξ) = Cξ + ζ (1.17)

where ζ = (z1 , · · · , zn ) and C is the matrix with cik = aik + δik .

To apply the Banach fixed point theorem we pick a metric on Cn such that (Cn , d) is complete. For
example we can take d(ξ, η) = kξ − ηkp with p ≥ 1 and then we have

d(F (ξ), F (η)) = kC(ξ − η)kp (1.18)

For example if p = 1 we have

 
X X  XX X
kCξk1 = cik xk  ≤ |cik ||xk | ≤ max |cik | kξk1 . (1.19)


  k
i k i k I
| {z }
:=q1

If we can find a norm such that kC(ξ)k ≤ qkξk then the equation (1.9) has a unique solution.
The fixed point equation has the form ξ = Cξ + ζ or (1 − C)ξ = ζ which gives formally using a
Neumann series (which we will justify later)

ξ = (1 − C)−1 ζ = (1 + C + C 2 + · · · )ζ . (1.20)

Note also that the Banach fixed point algorithms gives the sequence

ξn+1 = Cξn + ζ (1.21)

and ξ is the limit of ξn independent of the starting point ξ0 . This iteration is easy to solve and gives
n−1
X
ξn = C n ξ0 + Ckζ (1.22)
k=1

and thus, formally, we find

∞
X
ξ = lim C n ξ0 + Ckζ (1.23)
n→∞
k=1

The second term is exactly the Neumann series and the first term should go to 0.
1.2. BANACH FIXED POINT THEOREM 7

Example 1.10. (Fredholm integral equation) A very similar argument applies to the equation
Z b
f (t) = λ k(t, s)f (s) ds + h(t) (1.24)
a
Rb
where we use the metric space C[a, b] with the maximum metric d(f, g) = a
|f (t) − g(t)|dt. This is a
fixed point equation f = F (f ) where
Z b
F (f )(t) = h(t) + λ k(t, s)f (s) ds . (1.25)
a

We have
Z 
 b 
d(F (f ), F (g)) = max |F (f )(t) − F (g)(t)| = max  λk(t, s)(f (s) − g(s)) ds
 
t∈[a,b] t∈[a,b]  a 
!
Z b
≤ |λ| max |k(t, s)| ds d(f, g) . (1.26)
t∈[a,b] a

From the Banach fixed point theorem we deduce that the Fredholm integral equation has a unique solution
provided
Z b !−1
|λ| ≤ max |k(t, s)| ds (1.27)
t∈[a,b] a

Example 1.11. (Solving differential equations) Consider an ordinary differential equation (initial value
problem)
dx(t)
= f (x(t)) , x(a) = x0 (1.28)
dt
where f : R → R. It is easy to see that x(t) is solution of (1.28) on the interval [a, b] if and only if we have
for t ∈ [a, b]
Z t
x(t) = x0 + f (x(s)) ds , a ≤ t ≤ b . (1.29)
a
We can interpret this equation as a fixed point equation in M = {x ∈ C[a, b] : x(0) = x0 } which is a
closed subspace of a Banach space and thus itself a Banach space. We define F : M → M ] by
Z t
F (x)(t) = x0 + f (x(s)) ds , a ≤ t ≤ b (1.30)
a

and thus x(t) is a solution of (1.28) on the interval [a, b] if and only if F (x)(t) = x(t) for t ∈ [a, b].
In order to apply the Banach fixed point theorem we assume that f is globally Lipschitz, i.e., there exists
a constant L > 0 such that for all x, y ∈ R
|f (x) − f (y)| ≤ L|x − y| (1.31)
Then if we equip C[a, b] with the maximum metric we have
Z t
d(F (x), F (y)) = sup |F (x(t)) − F (y(t))| ≤ sup |f (x(s)) − f (y(s))| ds
t∈[a,b] t∈[a,b] a
Z t
≤ L |x(s) − y(s)| ≤ L(b − a) d(x, y) . (1.32)
a
8 CHAPTER 1. METRIC SPACES

We can apply the Banach fixed point theorem provided q = L(b − a) < 1. This means there is a unique
solution of the differential equation on a suitable interval [a, b] is b is sufficiently close to a.
There are many generalizations of this result to systems of differential equations or even partial differ-
ential equations and the Lipschitz condition can be somewhat relaxed too.

1.3 Exercises
P∞ and let F : M 7→ M . Suppose there exists a sequence an is
Exercise 1. Let M be a complete metric space
a sequence of non-negative numbers with n=1 an < ∞ and

d(F n (x), F n (y)) ≤ an d(x, y) . (1.33)

Show that F has a unique fixed point.

Hint: Modify the proof of Banach fixed point theorem
Exercise 2. In this problem we consider the Volterra integral equation given by the fixed point equation
f = F (f ) where
Z t
F (t)(t) = λ k(t, s)f (s) ds + h(s) (1.34)
a

where h(t) ∈ C[a, b] and k(t, s) ∈ C[a, b] × C[a, b] are given. We will show that this equation has a unique
solution for any value of λ (compare with the Fredholm integral equation).
Let K the Volterra integral kernel be given by
Z t
Kf (t) = λ k(t, s)f (s) ds (1.35)
a
Pn
1. Show that F n (f ) = k=0 K k h + K nf .
n
2. Prove that |K n f (t)| ≤ C n (t−a)
n! supt |f (t)| for a suitable constant C.
3. Use the previous exercise to show the existence of a unique solution in C[a, b] to the Volterra integral
equation.

Exercise 3. Consider the differential equation (initial value problem)

dx(t)
= f (x(t)) , x(0) = x0 (1.36)
dt
where f : R → R is Lipschitz. Consider the metric space C[0, ∞) with the metric

d(x, y) = sup e−Dt |f (t)| . (1.37)

t≥0

Use this metric space for a suitable choice of D to show that the initial value problem (1.36) has a unique
solution for t ∈ [0, ∞).
Chapter 2

Normed Vector Spaces

For general metric spaces (M, d) the set M has no structure besides the topology induced buy the norm. We
concentrate now on the special case where

M = V = vector space over K with K = R or C

2.1 Some concepts from linear algebra

We recall some basic concepts from linear algebra slightly generalized to vector spaces which may have
infinite dimension. In this section we do not use any topological concepts yet.
Definition 2.1. 1. A set M ⊂ V is called linearly independent if every finite subset of M is linearly
independent.

2. The set E ⊂ V is called a Hamel basis (algebraic basis) of V if E is linearly independent and every
vector ξ ∈ V can be written uniquely as finite linear combination of elements in E.
Using Zorn’s Lemma one can prove that
Theorem 2.2. Let M ⊂ V be linearly independent. Then V has a Hamel basis which contains M .
We use the following notations: for N, M ⊂ V , ξ ∈ V and a ∈ K

ξ+M = {ξ + η ; η ∈ M } ,
N +M = {ξ + η ; ξ ∈ N, η ∈ M } ,
aM = {aξ ; ξ ∈ M } .

Definition 2.3. 1. If M and N are subspaces of V and M ∩ N = {0} then M + N is a subspace and
one write M + N as M ⊕ N (direct sum).

2. If V = M ⊕ N we say that M and N are complementary subspaces.

Proposition 2.4. To each subspace M ⊂ V there exists a complementary subspace N (not uniquely de-
fined).

9
10 CHAPTER 2. NORMED VECTOR SPACES

Proof. Let EM a Hamel basis of M and E a Hamel basis of V which contains EM (see the proof of Theorem
2.2). Then E \ EM generates a subspace N which is complementary to M .

Definition 2.5. If V = M ⊕ N and dimN = n < ∞ then we say that M has codimension n, i.e.,

codim M = dim M . (2.1)

One check easily that

Proposition 2.6. If V = M ⊕ N1 = M ⊕ N2 . Then dim N1 = dimN2 .
Definition 2.7. Suppose V = M ⊕ N . Then any ξ ∈ V has a unique decomposition

ξ = α+β (2.2)

with α ∈ M , β ∈ N . The projection of ξ on M along N is given by

Pξ = α. (2.3)

One verifies easily that

Lemma 2.8. We have P 2 = P .
Conversely we have
Lemma 2.9. Suppose P : V → V is linear map such that P 2 = P . Then V = M ⊕ N where M = P V
and N = (1 − P )V .
Proof. For any ξ ∈ V we have
ξ = P ξ + (1 − P )ξ (2.4)
|{z} | {z }
∈M ∈N

and thus V = M + N . The sum is direct, since if

ξ ∈ M ∩ N = P V ∩ (1 − P )V (2.5)

we have, on one hand, ξ = P α and so

P ξ = P 2α = P α = ξ . (2.6)

On the other hand ξ = (1 − P )β and so we obtain

ξ = P ξ = P (1 − P )β = (P − P 2 )β = 0 . (2.7)

Thus M ∩ N = {0}.

Using projections we can prove

Theorem 2.10. Suppose T : V → W is a linear map. Let P a projection on the nullspace of T and Q a
projection along the range of T . Then there exists a linear map S : W → V such that

ST = 1V − P , T S = 1W − Q . (2.8)

The map T is bijective if and only P = Q = 0.

2.2. NORM 11

Proof. Let us denote N = N (T ) ⊂ V the nullspace (or kernel) of T and and M = R(T ) ⊂ W the range
of T . We have
V = N ⊕ V1 , W = W1 ⊕ M . (2.9)
and P is the projection on N along V1 and Q the projection on W1 along M .
We define T0 : V1 → M by T0 ξ = T ξ. Then T0 is bijective and so T0−1 : M → V1 exists. If

S = T0−1 (1 − Q) : W → V1 (2.10)

we have 
0 if ξ ∈ V
ST ξ = (2.11)
ξ if ξ ∈ V1
and so
ST = 1V − P . (2.12)
Arguing similarly we have
T S = 1W − Q . (2.13)

2.2 Norm
Suppose we have a metric d on a vector space V . It is natural to ask whether the metric respects the linear
structure of V . By that we mean that
1. d is translation invariant, i.e., d(ξ + α, η + α) = d(ξ, η) for all α ∈ V .
2. Under scalar multiplication we have d(aξ, aη) = |a|d(ξ, η).
Property 1. implies that d(ξ, η) = d(ξ −η, 0). If we set then kξk := d(ξ, 0) we have then from the properties
of the distance that
(N1) kξk ≥ 0 and ξ = 0 if and only if ξ = 0.
(N2’) kξk = k − ξk (symmetry)
(N3) kξ + ηk ≤ kξk + kηk
while from Property 2., instead of (N2’) we obtain the stronger
(N2) kaξk = |a|kξk
Definition 2.11. (Normed vector spaces)
1. A map k · k : V → R is called a norm on V if it satisfies the condition (N1), (N2), and (N3).
2. The pair (V, k · k) is called a normed vector space.
3. A complete normed vector space is called a Banach space.
Note that in a normed vector space we have convergence in the sense of the metric defined by d(ξ, η) =
kξ − ηk or equivalently ξn converges to ξ if and only if limn→∞ kξ − ξn k = 0.
12 CHAPTER 2. NORMED VECTOR SPACES

Example 2.12. We give examples of normed vector spaces (see Math 623-624 for details and proofs).
1. For p ≥ 1,
X
lp = |xi |p )1/p < ∞

ξ = (x1 , x2 , x3 , · · · ) xi ∈ C with kξkp := ( (2.14)
i

is a Banach space. The completeness is non-trivial as is the triangle inequality (a.k.a Minkowsky
inequality).
2. The space
l∞ =

ξ = (x1 , x2 , x3 , · · · ) xi ∈ C with kξk∞ := sup |xi | < ∞ (2.15)
i

is a Banach space as well as

ξ ∈ l∞ , ξ = (x1 , x2 , x3 , · · · ) lim xi = 0

c0 = (2.16)
i

and
ξ ∈ l∞ , ξ = (x1 , x2 , x3 , · · · ) lim xi exists

c = (2.17)
i

3. The space

C n [a, b] =

f : [a, b] → R : f n times continuously differentiable (2.18)

is a Banach space with the norm

n
X
kf k = max |f (k) (t)| (2.19)
t∈[a,b]
k=0

where f (k) is the k-th derivetive of f .

4. The space 
BV [a, b] = f : [a, b] → R : f of bounded variation (2.20)
is a Banach space with the norm
kf k = |f (a)| + V (f ) (2.21)
where V (f ) is the variation of f on [a, b] given by
N
X
V (f ) = sup |f (xk ) − f (xk−1 | (2.22)
P
k=1

where P is set of partition of [a, b]: a = x0 < x1 < · · · < xn = b.

The following facts are very easy but also very important

Proposition 2.13. Let (V, k · k) be a normed vector space.

1. The linear operations are continuous
2. The map ξ → kξk is continuous.
2.3. CONTINUOUS LINEAR MAPS 13

Proof. If ξn → ξ and ηn → η then ξn + tηn → ξ + tη since

k(ξ + t)η − (ξn + tηn k ≤ kξ − ξn k + |t|kη − ηn k . (2.23)

If ξn → ξ then kξn k → kξk since

|kξk − kξn k| ≤ kξ − ξn k (2.24)
by the (reverse) triangle inequality.

Definition 2.14. (Comparison of norms) Suppose two norms k · k1 and k · k2 are given on a vector space
V.

1. The norm k · k1 is stronger than k · k2 if there exists C > 0 such that for all ξ ∈ V .

kξk2 ≤ Ckξk1 (2.25)

2. The The norms k · k1 and k · k2 are equivalent if there exists constants C and D such that

Dkξ1 k ≤ kξk2 ≤ Ckξk1 . (2.26)

Clearly equivalent norm induce the same topology since convergence of one sequence in one norm
implies the convergence of the sequence in the other norm.
Theorem 2.15. If dim(V ) < ∞ then all norms on V are equivalent.

Proof. : This is left as an exercise. Use Bolzano-Weierstrass.

2.3 Continuous linear maps

Let (V, k · kV ) and (W, k · kW ) be two normed vector space. In the sequel, whenever there is no risk of
confusion we shall drop the index V or W from the norm and simply denote it by k · k. We also consider a
linear map
T : V →W. (2.27)
Unless explicitly specified we will deal only with linear maps in the sequel. We also use the notation T ξ for
T (ξ)
Definition 2.16. Let T : V → W be a linear map.

1. The map T is bounded if there exists a constant C ≥ 0 such that

kT ξk ≤ Ckξk (2.28)

for all ξ ∈ V .

2. The norm of T , denoted by kT k is the smallest C such that (2.28) holds, i.e.

kT ξk
kT k := sup = sup kT ξk . (2.29)
ξ∈V,ξ6=0 kξk ξ∈V,kξk=1
14 CHAPTER 2. NORMED VECTOR SPACES

3. The set of bounded linear maps is denoted by


L(V, W ) = T : V → W, T linear and bounded (2.30)

and we write
L(V ) := L(V, V ) . (2.31)

Theorem 2.17. Let T be a linear map. Then T is bounded if and only if T is continuous.

Proof. Suppose T is bounded, then we have

kT ξ − T ηk = kT (ξ − η)k ≤ kT kkξ − ηk (2.32)

and this implies that T is (Lipschitz) continuous.

Conversely let us assume that T is continuous but not bounded. Then there exists a sequence ξn such
that
kT ξn k ≥ nkξn k . (2.33)
Let us set
ξn
ηn = √ .
nkξn k
Then we have
1 √
kηn k = √ , kT ηn k > nkηn k = n. (2.34)
n
This means that ηn → 0 but T ηn is divergent which contradicts the continuity of T .

Example 2.18. Let us consider some examples of bounded (and not bounded operators). Many more exam-
ples to come.

1. The identity operator 1 defined by 1ξ = ξ is bounded with k1k = 1.

2. The differentiation operator is not bounded. Take for example the space V which consists of polyno-
mials p(t) on [0, 1] with the sup-norm and set T p(t) = p0 (t). Then if pn (t) = xn we have kpn k = 1
for all n but kT pn k = kntn−1 k = n and this shows that T is not bounded.
Note that differentiation is a very natural operation so to include it in our consideration we will con-
sider unbounded linear operator later on.

3. The integral operator

Z b
T f (t) = k(t, s)f (s) ds (2.35)
a

is bounded on C[a, b] equipped with the sup-norm if k(t, s) ∈ C([a, b] × [a, b]). Indeed we T f (t) is
continuous in t by e.g. the dominated convergence theorem and
Z b Z b
kT f k = sup | k(t, s)f (s) ds| ≤ sup |f (s)| sup |k(t, s)| ds (2.36)
t a s t a
| {z }
≡C
2.3. CONTINUOUS LINEAR MAPS 15

Rb
and C is finite since k is continuous. Actually we can show that kT k = supt a |k(t, s)| ds. If
Rb
k(t, s) is nonnegative then we simply take f = 1 and we have kT 1k = supt a |k(t, s)| ds. For a
Rb Rb
general k we pick t0 such that supt a |k(t, s)| ds = a |k(t0 , s)| ds. Then we would like to pick
Rb Rb
the function f (s) = sign k(t0 , s) so that T f (t0 ) = a k(t0 , s)f (s)ds = a |k(t0 , s)| ds. But this
f is not continuous and we need to use an approximation argument. Consider the function φn (t)
which is piecewise linear, increasing, continuous, with φ(t) = sign t if |t| > 1/n. We set fn (s) =
φn (k(t0 , s)). We have
Z b Z b
1
T fn (t0 ) = k(t0 , s)fn (s) ds = |k(t0 , s)| ds − (b − a) . (2.37)
a a n
Rb
From this it follows that kT k = a
|k(t0 , s)| ds.

4. The Fourier transform is defined by

Z
fˆ(k) = f (x)e−i2πkx , (2.38)
R

and we write T (f ) = fˆ. In Math 623-624 ones proves (after some effort) that T : L1 (R, dx) →
C0 (R) is a bounded operator with norm 1. Here C0 (R) is the Banach space of continuous functions
such that lim|x|→∞ |f (x)| → 0 equipped with the sup-norm. One also proves (after some more
effort) that T : L2 (R, dx) → L2 (R, dx) defines an unitary transformation (i.e. T is invertible and
kT f k = kf k for al f ).

Theorem 2.19. L(V, W ) is a normed vector space.

Proof. Let T , S ∈ V and a a scalar.

• (N1) We have kT k ≥ 0. If kT k = 0 then we have kT ξk ≤ kT kkξk = 0 and so T = 0. Conversely if

T = 0 then obviously kT k = 0.

• (N2) kaT k = sup kaT ξk = |a| sup kT ξk = |a|kT k

kξk=1 kξk=1

• (N3) k(T + S)ξk ≤ kT ξk + kSξk and thus kT + Sk ≤ kT k + kSk.

Theorem 2.20. If W is a Banach space then L(V, W ) is a Banach space.

Proof. Let Tn be Cauchy sequence in L(V, W ). For any ξ ∈ V we have

kTn ξ − Tm ξk ≤ kTn − Tm kkξk (2.39)

and thus Tn ξ is a Cauchy sequence in W . Since W is complete this sequence has a limit in W and we set

η = lim Tn ξ (2.40)
n→∞
16 CHAPTER 2. NORMED VECTOR SPACES

We define then T : V → W by T ξ = η. The linearity of Tn immediately implies that T is linear. What

remains to prove is that T is bounded and Tn → T . Given  > 0 pick N such that for n, m > N we have
kTn ξ − Tm ξk ≤ kTn − Tm kkξk ≤ kξk (2.41)
Using the continuity of the norm and taking m → ∞ in (2.41) we have
kTn ξ − T ξk ≤ kξk (2.42)
for any n ≥ N . This implies that kT − TN k ≤  and so kT k ≤ kT − TN k + kTN k < ∞ and so T is
bounded. This also implies that kT − Tn k ≤  for n ≥ N and so Tn → T .

Theorem 2.21. For S ∈ L(U, V ) and T ∈ L(V, W ) we have T S := T ◦ S ∈ L(U, W ) and

kT Sk ≤ kT kkSk (2.43)
Proof. kT Sk = sup kT Sξk ≤ kT k sup kSξk = kT kkSk.
kξk=1 kξk=1

Corollary 2.22. For T ∈ L(V ) we have

1. kT n k ≤ kT kn
2. The limit r(T ) = lim kT n k1/n exists and is called the spectral radius of T .
n→∞
n 1/n
3. r(T ) = inf kT k ≤ kT k
n

Proof. 1. is immediate. For 2. and 3. set cn = kT n k/kT kn . Then we have 0 ≤ cn ≤ 1 and cn+m ≤ cn cm
1/n
and this implies that cn is a bounded decreasing sequence and so c = limn→∞cn exists. Therefore cn is
also a decreasing with limit
kT n k1/n kT n k1/n
lim = inf (2.44)
n→∞ kT k n kT k

The notation ”spectral radius” will become clear in the sequel.

As an application we consider the Neumann series which is is a generalization of the geometric series.
P∞ k
Suppose V is a Banach space,P∞ T ∈ L(V ) with kT k = δ < 1. Then the series k=0 T converges since
k
since the partial sum Sn = k=0 T satisfy for n > m
n n
X X 1
kSn − Sm k ≤ kT k k ≤ δ k ≤ δ m+1 (2.45)
1−δ
k=m+1 k=m+1
k
P
and thus form a Cauchy sequence and we have limn Sn = k=0∞ T .
Furthermore we have
(1 − T )Sn = Sn (1 − T ) = 1 − T n+1 → 1 (2.46)
and thus we conclude that 1 − T is invertible and
∞
X
(1 − T )−1 = Tk . (2.47)
k=0

One can prove the stronger result

2.3. CONTINUOUS LINEAR MAPS 17

Theorem 2.23. Let V be a Banach space and T ∈ L(V ) with r(T ) < 1. Then (1 − T )−1 exists and is
given by the Neumann series
∞
X
(1 − T )−1 = Tk (2.48)
k=0

T K converges whenever k kT k k does. If r(T ) = δ < 1 then for any δ 0 with

P P
Proof. Note first that
δ <Pδ < 1 there exists N such that for n > N we have kT n k1/n ≤ δ 0 . This means that kT n k ≤ (δ 0 )n and
0

so k kT k k converges. The rest is as before.

Linear operators occur naturally as derivative and this is true in infinite-dimensional spaces as well.
Definition 2.24. Let V and W be normed vector spaces and U ⊂ V open. The map F : U → W is said to
be differentiable at η ∈ U if there exists a continuous linear map F 0 (a) such that

F (ξ) = F (η) + F 0 (η)(ξ − η) + R(ξ)kξ − ηk , (2.49)

where R : U → W satisfy limξ→η R(ξ) = 0.

Let is consider a few examples
Example 2.25. 1. Let V = W = C[a, b] and let k ∈ C[0, 1] × [0, 1] and g be twice continuously
differentiable. Let us consider the map
Z b
F (x)(t) = k(t, s)g(x(s)) ds (2.50)
a

To compute the derivative we pick h ∈ C[0, 1] and using the mean value theorem we obtain
Z b
F (x + h) − F (x) = k(t, s) [g(x(s) + h(s)) − g(x(s))] ds
a
Z b  
0 1 00 2
= k(t, s) g (x(s))h(s) + g (a(s))h(s) ds (2.51)
a 2

where a(s) is a value between x(s) and x(s) + h(s). This computation shows that the linear F 0 (x) :
C[0, 1] → C[0, 1] given by
Z b
F 0 (x)(h)(t) = k(t, s)g 0 (x(s))h(s) ds (2.52)
a

is a good candidate for the derivative. Clearly this map is bounded since k(t, s)g 0 (x(s)) ∈ C[0, 1] ×
[0, 1]. Moreover we have the bound
Z 
 b 1 
00
sup  k k(t, s)g (a(s))h(s) ds ≤ Ckhk2
2
(2.53)
 
t  a 2 

since g 00 is bounded by assumption. This shows that F is differentiable.

18 CHAPTER 2. NORMED VECTOR SPACES

2. Let E be a Banach space and consider the Banach space L(E). We say that T ∈ L(E) is an isomor-
phism if T is linear, bijective, continuous and if A−1 is continuous. We will see in fact in Section 2.8
that by the open map theorem the continuity of A−1 follows from the other assumptions.
In any case the set
GL(E) = {T ∈ L(E) : , T an isomorphism} (2.54)
is an open subset of L(E). Indedd if T is an isomorphism then T + H is an isomorphism provided
kHk ≤ kT −1 k−1 . This follows from T + H = T (1 + T −1 H), from kT −1 Hk ≤ kT −1 kkHk < 1
and from the geometric series.
Let us consider the map F : GL(E) → L(E) given by

F (T ) = T −1 . (2.55)

We claim that the F is differentiable and that we have

F 0 (T )H = −T −1 HT −1 . (2.56)

The continuity of F 0 (T ) is clear since we have kF 0 (T )Hk ≤ kA−1 k2 kHk. Furthermore we have

F (T + H) − F (T ) = (T + H)−1 − T −1 = (1 + T −1 H) − 1 T −1
 

X∞
= (−T −1 H)j T −1 (2.57)
j=1

if kHk ≤ kT −1 k−1 (geometric series). The first term in the series is −T −1 HT −1 and the remainder
can be bounded by kT −1 k2 kHk(1 − kT −1 kkHk)−1 . The reminder divided by kHk tends to 0,
showing the differentiability.

Pointwise (strong) convergence: In L(V, W ), in addition to norm convergence ( ≡ uniform convergence)

there is the weaker notion of point wise convergence.
Definition 2.26. Let {Tn } a sequence in L(V, W ) and T ∈ L(V, W ). We say that Tn converges strongly to
T and write
T = s − lim Tn (2.58)
n→∞

if for any ξ ∈ V we have

lim Tn ξ = T ξ . (2.59)
n→∞

Whenever we need to differentiate between different types of convergence we will write n − limn→∞ Tn
for the convergence in norm.
The following two lemmas are very easy and the proof is left to the reader.
Lemma 2.27. If n − limn→∞ Tn = T then s − limn→∞ Tn = T .
Lemma 2.28. The linear operations are continuous with respect to strong convergence.
However note that if n − limn→∞ Tn = T then Tn k → kT k but this does not necessarily hold in the
case of strong convergence. Similarly if n − limn→∞ Tn = T and n − limn→∞ Sn = S then we have
n − limn→∞ Sn Tn = ST but this does not necessarily hold for strong convergence.
2.4. LINEAR FUNCTIONALS AND DUAL SPACES 19

Example 2.29. Let V = lp and for ξ = (x1 , x2 , · · · ) let us set

Tn ξ = (x1 , · · · , xn , 0, 0, · · · ) . (2.60)

Then we have Tn ξ − ξ → 0 and so s − limn→∞ Tn = 1. But since

(1 − Tn )ξ = (0, · · · , 0, xn+1 , xn+2 , · · · ) (2.61)

for any n we can find ηn with kηn = 1k = 1 and (1 − Tn )ηn = ηn . This implies that k1 − Tn k = 1 for all
n and so n − limn→∞ Tn is certainly not equal to 1.

2.4 Linear functionals and dual spaces

Definition 2.30. If V is a Banach space over K = R or C then L(V, K) is a Banach space. We write

V 0 := L(V, K) (2.62)

and V 0 is called the dual space. The elements of V 0 are called linear functionals on V and for λ ∈ V 0 the
norm is given by
kλk = sup |λ(ξ)| . (2.63)
ξ∈V
kξk=1

Let us work out an example in detail

Example 2.31. If V = lp for p > 1 then V 0 = lq where p−1 + q −1 = 1.
Proof. Note first that if η = (y1 , y2 , · · · ) ∈ lq then λη (ξ) defined by
∞
X
λη (ξ) = xk yk (2.64)
k=1

defines a bounded linear functional on lp since by Hölder’s inequality we have

|λη (ξ)| ≤ kξkp kηkq , (2.65)

and so kλη k ≤ kηkq .

Let us denote
k = {δik }∞ p
i=1 ∈ l . (2.66)
Any ξ ∈ lp can be written as the convergent series in lp
∞
X
ξ = x k k . (2.67)
k=1

Let λ ∈ V , since it is continuous, we have

X
λ(ξ) = xk yk with yk = λ(k ) . (2.68)
k

and we set η = {yk } and λ has the from λη .

20 CHAPTER 2. NORMED VECTOR SPACES

(n)
Let us pick ξ (n) = {xk } with
(
|yk |q
(n) yk if k ≤ n and yk 6= 0
xk = . (2.69)
0 otherwise

We have then
n
X
λ(ξ (n) ) = |yk |q
k=1

≤ kλkkξ (n) kp
n
!1/p
(n)
X
= kλk |xk |p
k=1
n
!1/p
X
(q−1)p
= kλk |yk |
k=1
n
!1−1/q
X
= kλk |yk |q . (2.70)
k=1

We obtain then !1/q

n
X
q
|yk | ≤ kλ . (2.71)
k=1

Since n is arbitrary we conclude that η ∈ lq and kηkq ≤ λ. Combining with Hölder’s inequality shows that
kΛη k = kηkq and that the map λ 7→ η from (lp )0 to lq is a norm preserving isomorphism.

Example 2.32. In a similar way one shows that Lp (X, µ)0 = Lq (X, µ) for an arbitrary measure space
(X, µ) and 1 < p < ∞.
Example 2.33. With some minor modifications the same proof works for l1 and we have (l1 ) = l∞ .
However it is not true that (l∞ )0 = l1 but rather we have
l1 = (c0 )0 (2.72)
Proof. Let η ∈ l1 and ξ ∈ c0 . Then the series λη (ξ) = k xk yk converges and by Hölder’s equality we
P
have
|λη (ξ)| ≤ kξk∞ kηk1 . (2.73)
Thus we kλη k ≤ kηk1 .
(N )
For any  > 0, let N be such that k≥N |yk | < . Pick then ξ (N ) = {xk } with
P

(
|yk |
(N ) yk if k ≤ n and yk 6= 0
xk = . (2.74)
0 otherwise

We have then X
|λη (ξ)| = |yk | ≥ (kηk1 − )kξk . (2.75)
k≥N
2.5. HAHN-BANACH THEOREM 21

Therefore kλη k ≥ kηk1 and so kλη k = kηk1 .

remains to show that every linear functional on c0 has the above form. If ξ ∈ c0 then we can write
ItP
∞
ξP= k=1 xk k (not true in l∞ !) and so for any λ ∈ c00 there exists η = {yk } such that λ(ξ) = λη (ξ) =
∞ (n)
k=1 xk yk . Picking ξ as in (2.74) we have kξ (n) k∞ = 1 and
n
!
X
kλη (ξ (n) ) = |yk | kξ (n) k∞ (2.76)
k=1

from which we conclude that kλη k ≥ kηk1 , i.e. η ∈ l1 .

Schauder basis and separability:

Definition 2.34. A metric space is called separable if it contains a countable dense set.
One shows that lp is separable but l∞ is not separable.

In the construction of dual spaces we used the fact that any ξ ∈ lp , 1 ≤ p < ∞ can be written uniquely
as !1/p
X X
p
ξ = x k k , with kξkp = |xk | <∞ (2.77)
k=1 k=1

In general we have
Definition 2.35. A countable subset B = {k } of a normed
P∞ vector space V is called a Schauder Basis of V
if each vector ξ ∈ V can be written uniquely as ξ = k=1 xk k .
Without difficulty one shows that
Lemma 2.36. A normed vector space which has a Schauder basis is separable.

It is a remarkable and deep fact that the converse does not hold: there exists Banach spaces which are
separable but do not have a Schauder basis. (See P. Enflo. A counterexample to the approximation problem
in Banach spaces, Acta Math 130, 309, (1973).)

2.5 Hahn-Banach theorem

As we will see a very important question in functional analysis is how to extend a functional defined on a
subspace W of a vector space V to all V while respecting some properties.
Theorem 2.37. Hahn-Banach (real vector spaces) Let V a vector space over R and p : V → R a convex
function on V , i.e. we have
p(aξ + (1 − a)η) ≤ ap(ξ) + (1 − a)p(η) (2.78)
for any a ∈ [0, 1] and all ξ, η ∈ V .
22 CHAPTER 2. NORMED VECTOR SPACES

Let φ be a linear functional defined on the subspace W ⊂ V such that we have

φ(ξ) ≤ p(ξ) for all ξ ∈ W . (2.79)

Then there exists a linear functional Φ defined on V such that

Φ(ξ) = φ(ξ) for all ξ ∈ W .

Φ(ξ) ≤ p(ξ) for all ξ ∈ V . (2.80)

A very important example of a convex function p is a norm (or a semi-norm) on V , i.e., p(ξ) = kξk.

Proof. The idea of the proof is to show that one can extend φ from W to W + Rη for any ξ ∈
/ W . The rest
follows from an applications of Zorn’s lemma.
Let us choose η ∈/ W and let denote W1 = W + Rη. By linearity it is enough to specify c := φ(η) to
define an extension φ1 on W1 since we have then

φ1 (ξ + aη) = φ1 (ξ) + aφ1 (η) = φ(ξ) + ac . (2.81)

To obtain the desired bound on φ1 we need that for all ξ ∈ W and a ∈ R we must have

φ1 (ξ + aη) = φ(ξ) + ac ≤ p(ξ + aη) . (2.82)

We restrict ourselves to a > 0. Then we looking for a c such that

φ(ξ) + ac ≤ p(ξ + aη)

φ(ξ) − ac ≤ p(ξ − aη)

or
1
c ≤ (p(ξ + aη) − φ(ξ))
a
1
c ≥ (−p(ξ − aη) + φ(ξ))
a
(2.83)

Such a c exists provided we can prove that

1 1
sup (−p(ξ − aη) + φ(ξ)) ≤ inf (p(ξ + aη) − φ(ξ)) (2.84)
ξ∈W , a>0 a ξ∈W , a>0 a

that is
1 1
(−p(ξ1 − a1 η) + φ(ξ1 )) ≤ (p(ξ2 + a2 η) − φ(ξ2 )) (2.85)
a1 a2
for all a1 , a2 ≥ 0 and ξ1 , ξ2 ∈ W . Since a1 and a2 are non-negative this is equivalent to

a2 φ(ξ1 ) + a1 φ(x2 ) ≤ a1 p(ξ2 − a2 η) + a2 p(ξ1 − a1 η) . (2.86)

2.5. HAHN-BANACH THEOREM 23

On the other hand we have use the assumption

a2 φ(ξ1 ) + a1 φ(x2 ) = φ(a2 ξ1 + a1 ξ2 )

 
a2 a1
= (a1 + a2 )φ ξ1 + ξ2
a1 + a2 a1 + a2
 
a2 a1
≤ (a1 + a2 )p ξ1 + ξ2
a1 + a2 a1 + a2
 
a2 a1
= (a1 + a2 )p (ξ1 − a1 η) + (ξ2 + a2 η)
a1 + a2 a1 + a2
≤ a2 p(ξ1 − a1 η) + a1 p(ξ2 + a2 η) . (2.87)

This is exactly (2.86) and thus we have proved the existence of c := φ1 (η).
We now use Zorn’s lemma. Let A the set of all extension ψ won φ auf some subspace Wψ with the
property that φ = ψ on W and ψ(ξ) ≤ p(ξ) for all ξ ∈ Wψ . In A we can define a partial ordering through
ψ1 ≺ ψ1 if Wψ1 ⊂ Wψ2 and ψ1 = ψ2 on Wψ1 . Suppose B ⊂ A is totally ordered and define then Ψ
on ∪ψ∈B Wψ through Ψ(ξ) = ψ(ξ) on Wψ . By construction ψ ≺ Ψ for all ψ ∈ B and so B has an upper
bound. By Zorn’s lemma A has a maximal element Φ which satisfies Φ = φ on W and Φ(ξ) ≤ p(ξ). Finally
WΦ must be V since otherwise one could extend Φ as before and this contradicts maximality.

This theorem as an extension to complex vector spaces.

Theorem 2.38. Hahn-Banach (real vector spaces) Let V a vector space over C and p : V → R such that
we have p(aξ + bη) ≤ |a|p(ξ) + |b|p(η) for any a, b ∈ C with |a| + |b| = 1 and all ξ, η ∈ V .
Let φ be a complex linear functional defined on the subspace W ⊂ V such that we have

|φ(ξ)| ≤ p(ξ) for all ξ ∈ W . (2.88)

Then there exists a linear functional Φ defined on V such that

Φ(ξ) = φ(ξ) for all ξ ∈ W .

|Φ(ξ)| ≤ p(ξ) for all ξ ∈ V . (2.89)

Proof. The functional φr (ξ) = Reφ(ξ) is a real functional on V viewed as a vector field over R. In addition

φr (iξ) = Reφ(iξ) = Reiφ(ξ) = −Imφ(ξ) . (2.90)

and so
φ(ξ) = φr (ξ) − iφr (iξ) . (2.91)
Conversely given any real linear functional Φr on V let us define Φ(ξ) = Φr (ξ) − iΦr (iξ). It is certainly
linear over R and we have

Φ(iξ) = Φr (iξ) − iΦr (−ξ) = Φr (iξ) + iΦr (ξ) = iΦ(ξ) (2.92)

and so Φ is linear over C.

By the real Hahn-Banach theorem we can extend φr to Φr on V such that Φr (ξ) ≤ p(ξ) and Φ(ξ) =
Φr (ξ) − iΦr (iξ) is complex linear. With θ = argΦ(ξ) we find

|Φ(ξ)| = e−iθ Φ(ξ) = Φ(e−iθ ξ) = Φr (e−iθ ) ≤ p(e−iθ ξ) ≤ |e−iθ |p(ξ) = p(ξ). (2.93)
24 CHAPTER 2. NORMED VECTOR SPACES

It is hard to overemphasize how important this theorem is for the foundations of functional analysis. Let
us first derive some corollaries and then discuss a number of applications.

Corollary 2.39. Let (V, k · k) be a normed vector space, W ⊂ V a subspace, and φ ∈ W 0 . Then there exists
λ ∈ V 0 such that λ = φ on W and kλk = kφk.

Proof. Take p(ξ) = kφkkξk and apply Hahn-Banach.

Corollary 2.40. Let (V, k · k) be a normed vector space, 0 6= ξ ∈ V . Then there exists λ ∈ V 0 such that
λ(ξ) = kξk and kλk = 1.

Proof. Let W = Kξ be the subpsace spanned by ξ and set φ(aξ) = akξk. This is a linear functional with
kφk = 1. Now use Corollary 2.39.

Corollary 2.41. Let (V, k · k) be a normed vector space, W ⊂ V a subspace. Let ξ ∈ V be such that

inf kξ − ηk = δ > 0 . (2.94)

η∈W

Then there exists λ ∈ V 0 such that kλk = 1, λ(ξ) = δ and λ(η) = 0 for η ∈ W .

Proof. Consider the subspace W1 = W ⊕ Kξ and let define λ1 on W1 by

λ1 (η + aξ) = aδ (2.95)

Clearly we have λ1 (η) = 0 for η ∈ W and λ1 (ξ) = δ. The functional λ1 is linear and bounded with norm
1: using the assumption (2.94) we have

1 1
kη + aξk = k − a(− η − ξ)k = |a|k − η − ξk ≥ |a|δ = |λ1 (η + aξ)k (2.96)
a a

and so kλ1 k ≤ 1. On the other hand for  > 0 there exists η ∈ W such that

δ ≤ kξ − ηk ≤ δ(1 + ) . (2.97)

Then we have
1
λ1 (ξ − η) = δ ≥ kξ − ηk , (2.98)
1+
that is kλ1 k ≥ (1 + )−1 . So kλ1 k = 1. Now use Corollary 2.39 for any kξk there exists
2.6. APPLICATIONS OF HAHN-BANACH THEOREM 25

2.6 Applications of Hahn-Banach theorem

Dual and bidual spaces The Hahn-Banach space is useful to derive the properties of the dual space V 0 as
well as the bidual V 00 = (V 0 )0 of a Banach space. First we derive a dual representation of the norm
Theorem 2.42. Let V be a normed vector space. Then we have
|λ(ξ)|
kξk = sup (2.99)
λ∈V 0 ,λ6=0 kλk

In particular if ξ0 is such that λ(ξ0 ) = 0 for all λ ∈ V 0 then ξ0 = 0.

Proof. On one hand from |λ(ξ)| ≤ kλkkξk we have
|λ(ξ)|
kξk ≥ sup . (2.100)
λ∈V 0 ,λ6=0 kλk

Using corollary 2.40 for a given fixed ξ there exists λξ ∈ V 0 such that kλk = 1 and

|λξ (ξ)|
kξk = λξ (ξ) = . (2.101)
kλξ k

The next results describe the relation between a Banach space V and his bidual V 00 .
Theorem 2.43. Let V be a normed vector space. For ξ ∈ V define an element ξˆ ∈ V 00 by
ˆ
ξ(λ) = λ(ξ) , λ∈V0. (2.102)

Then the map

J : V → V 00
ξ 7→ ξˆ (2.103)

is an isometric isomorphism from V to a subspace of V 00 .

ˆ
Proof. Since |ξ(λ)| = |λ(ξ)| ≤ kλkV 0 kξkV is ξˆ a bounded linear functional on V 0 with
ˆ V 00 ≤ kξkV .
kξk (2.104)
ˆ V 00 = kξkV . By corollary 2.40 given ξ there
This shows that J(V ) ⊂ V 00 and it remains to show that kξk
exists λξ ∈ V 0 such that λξ (ξ) = kξk. Then we have
ˆ V 00 =
kξk sup ˆ
|ξ(λ)| ˆ ξ )| = λξ (ξ) = kξkV .
≥ |ξ(λ (2.105)
λ∈V 0 ,kλk=1

Therefore J is an isometry of V onto its range.

The following theorem suggest

26 CHAPTER 2. NORMED VECTOR SPACES

Definition 2.44. A Banach space V is called reflexive (or self-dual) if J : V → V 00 is bijective.

With respect to separability we have

Theorem 2.45. A normed vector space V is separable if V 0 is separable.

Proof. Consider a dense set {λk }k≥1 in V 0 and for each k pick ξk ∈ V with kξk k = 1 and such that

1
λk (ξk ) ≥ kλk k . (2.106)
2
Now let W be the countable set of finite linear combinations of the ξk with rational coefficients.
By contradiction let us assume that W is not dense and so there exists ξ ∈ V such that

inf kξ − ηk = δ > 0 . (2.107)

η∈W

By corollary 2.41 there exists λ ∈ V 0 such that kλk = 1, λ(ξ) = δ and λ|W = 0. But since {λk } is dense
in V 0 there exists a subsequence λki such that limi→∞ kλ − λki k = 0. On the other hand we have

1
kλ − λki k ≥ |(λ − λki )(ξki )| = |λki (ξki | ≥ kλki k (2.108)
2
But this implies that limi kλki k and hence λ = 0. This is a contradiction since 0 ∈ W .

Corollary 2.46. A separable Banach space V with a non separable dual space V 0 cannot be reflexive.

Proof. If V were reflexive then V 00 = V would be separable and hence V 0 would be separable by Theorem
2.45.

Example 2.47. Every Hilbert space is reflexive. From Riesz representation theorem (see Math 623-624)
a Hilbert space H is isometrically isomorphic to its dual H 0 which is itself a Hilbert space. Hence H is
isometricaly isomorphic to its bi-dual H 00 .

Example 2.48. lp is reflexive and separable (see hwk). On the other hand l∞ is not separable (see HWK)
from which it follows by corollary 2.46 that l1 is not reflexive and hence (l1 )00 = (l∞ )0 6= l1 .

Example 2.49. C[a, b] is not reflexive since its dual space (C[a, b])0 is not separable. To see this note that
for any s ∈ [a, b] the functionals
λs (f ) = f (s) (2.109)
satisfy kλs k = 1 and λs − λs0 k = 2. Since there are uncountably such functionals separability is excluded.

We shall not prove the following important and classical result (see your measure theory class for details).
Let us consider theR Banach space M ([a, b]) of all finite complex Borel measures on [a, b] with total variation
norm kµkvar = [a,b] d|µ| as well as the Banach space N BV [a, b] of function of bounded variation with
F (a) = 0 with norm kF kN BV = V (F ) where V (F ) is the variation of F .
2.6. APPLICATIONS OF HAHN-BANACH THEOREM 27

Theorem 2.50. Any bounded linear functional λ on C[a, b] can be written as

Z
λ(f ) = f dµ µ Borel measure
[a,b]
Z
= f dF Lebesgue − Stieljes integral (2.110)
[a,b]

with
kλk = kµkvar = V ar(F ) . (2.111)
Hence
0
(C[a, b]) = N BV [a.b] = M [a, b] . (2.112)
Finally we construct the dual or adjoint operator to a bounded operator using Hahn-Banach theorem.
Definition 2.51. Let V and W be normed vector spaces and T : V → W a bounded linear operator. Then
the adjoint operator T 0 : W 0 → V 0 is defined by
(T 0 λ)(ξ) = λ(T ξ) , ξ ∈ V, λ ∈ W 0 . (2.113)
Example 2.52. Let V = W = l1 and T the shift operator defined for ξ = (x1 , x2 , · · · ) by
T ξ = (0, x1 , x2 , x3 , · · · ) (2.114)
P∞
We have (l1 )0 = l∞ with η(ξ) = k=1 xk yk . So
∞
X ∞
X
(T 0 η)(ξ) = η(T ξ) = yk xk−1 = yk+1 xk , (2.115)
k=1 k=1

hence we have
T 0 η = (y2 , y3 , · · · ) (2.116)
0
It is easy to check that kT k = kT k = 1.
Theorem 2.53. The adjoint operator T 0 is linear bounded and we have
kT 0 k = kT k . (2.117)
0
Proof. The fact T is linear is easy and left to the reader. We have
kT 0 λ(ξ)k = kλ(T ξ)k ≤ kλkkT ξk ≤ kλkkT kkξk (2.118)
and this implies that kT 0 λk ≤ kλkkT k and so kT 0 k ≤ kT k.
To prove that kT 0 k ≥ kT k we use that, by Corollary 2.40, for any ξ ∈ V there exists λξ ∈ W 0 such that
kλξ k = 1 λξ (T ξ) = kT ξk (2.119)
We have then
kT ξk = kλξ (T ξ)k
= k(T 0 λξ )(ξ)k
≤ kT 0 λξ kkξk
≤ kT 0 kkλξ kkξk
= kT 0 kkξk (2.120)
28 CHAPTER 2. NORMED VECTOR SPACES

and therefore kT k ≤ kT 0 k.

Remark 2.54. It is straightforward to see that the map T 7→ T 0 satisfies

1. (T + S)0 = T 0 + S 0
2. (αT )0 = αT 0
3. (ST )0 = T 0 S 0

In the case where V = H is a Hilbert space, then there is, by Riesz representation theorem, a natural
antilinear isometry
J : H0 →H
λ 7→ λ̂ (2.121)
where
λ(ξ) = (ξ, λ̂) , ξ∈H (2.122)
(Recall (·, ·)) is the scalar product in H linear in the first argument, anti-linear in the second argument.)
Definition 2.55. The (Hilbert space) adjoint operator T ∗ : H → H for T ∈ L(H) is defined by
T ∗ = JT 0 J −1 . (2.123)
Then we have for any ξ, η ∈ H
(T ξ, η) = (R−1 η)(T ξ) = (T 0 R−1 η)(ξ) = (ξ, RT 0 R−1 η) = (ξ, T ∗ η) . (2.124)
Note that the map T 7→ T ∗ is antilinear, i.e. cT ∗ = c̄T ∗ for c ∈ C. By slight abuse of notation T ∗ is
called the adjoint operator to T .
Theorem 2.56. Let H be a Hilbert space and T, S ∈ L(H). Then we have
1. T 7→ T ∗ is an anti linear isometry of L(H) onto itself.
2. (T ∗ )∗ = T
3. (T S)∗ = S ∗ T ∗
4. If T −1 ∈ L(H) then T ∗ −1 ∈ L(H) and (T ∗ )−1 = (T −1 )∗ .
5. kT ∗ T k = kT k2
Proof. 1. follows from Riesz representation theorem. 2. follows from the reflexivity of H and H = H 0 . 3.
and 4. are left to the reader. For 5. note that kT ∗ T k ≤ kT ∗ kkT k = kT k2 . Conversely using theorem 2.42
we have
|(T ∗ T ξ, η)| |(T ∗ T ξ, ξ)| |(T ξ, T ξ)| kT ξk2
kT ∗ T kkξk ≥ kT ∗ T ξk = sup ≥ = = (2.125)
η:η6=0 kηk kξk kξk kξk

and hence kT k2 ≤ kT ∗ T k.
2.7. BAIRE THEOREM AND UNIFORM BOUNDEDNESS THEOREM 29

2.7 Baire theorem and uniform boundedness theorem

The uniform boundedness theorem is quite useful in applications and as the open mapping theorem and
closed graph theorem of next section it derives from a common source: the so-called Baire category theorem.
Here, as opposed to the Hahn-Banach theorem the completeness of the space plays a crucial role.

Definition 2.57. A subset M of a metric space X is said to be

1. nowhere dense in X if its closure M has no interior points.

2. of the first category in X if M is the union of countable many sets each of which is nowhere dense.

3. of the second category in X if M is not of the first category.

Theorem 2.58. (Baire category theorem) A complete non-empty metric space X is of the second category
in itself. In particular if X 6= ∅ is complete and
∞
[
X = Ak , Ak closed (2.126)
k=1

then at least one Ak contains a nonempty open subset.

Proof. Let us assume that X is of the first category and so

∞
[
X = Mk , (2.127)
k=1

with each Mk nowhere dense. We will construct a Cauchy sequence ξk whose limit belongs to no Mk ,
therefore contradicting the representation 2.127.
By assumption M1 is nowhere dense so that M1 does not contain a nonempty open set. But X does (e.g.
c c
X itself). This implies that M1 6= X and thus M1 = X \ M1 is open and non-empty. So we pick ξ1 ∈ M1
and an open ball around it
c 1
B1 = B1 (ξ1 ) ⊂ M1 1 < (2.128)
2
By assumption M2 is nowhere dense in X so that M2 does not contain a nonempty open set. Hence it does
c
not contain the open ball B1 /2 (ξ1 ). Therefore M2 ∩ B1 /2 (ξ1 ) is nonempty and open so that we may
choose an open ball in this set , say,

c 1 1
B2 = B2 (ξ2 ) ⊂ M2 ∩ B1 /2 (ξ1 ) 2 < < (2.129)
2 4
By induction we find a sequence of balls

1
Bk = Bk (ξk ) k < (2.130)
2k
such that Bk ∩ Mk = ∅ and
Bk+1 ⊂ Bk /2 (ξk ) ⊂ Bk . (2.131)
30 CHAPTER 2. NORMED VECTOR SPACES

Since k < 2k the sequence ξk is Cauchy and limk→∞ ξk = ξ ∈ X. Furthermore for every n and m ≥ n
we have Bm ⊂ Bn /2 (ξn ) so that
n n
d(ξn , ξ) ≤ d(ξn , ξm ) + d(ξm , ξ) < + d(ξm , ξ) −→ . (2.132)
2 2
c
as m → ∞. Hence for every n, ξ ∈ BN ⊂ M n and so ξ ∈
/ ∪k Mk = X. This is a contradiction.

An important immediate consequence of Baire’s theorem is the uniform boundedness theorem (Banach-
Steinhaus theorem). It is quite remarkable since it shows that if a sequence Tn is point wise bounded, it is is
bounded in norm!.

Theorem 2.59. (Uniform boundedness theorem) Let {Tn } be a sequence of bounded linear operators
Tn ∈ L(V, W ) from a Banach space V into a normed vector space W . Assume that for ever ξ ∈ V there
exists a constant cξ such that
sup kTn ξk ≤ cξ . (2.133)
n

Then there exists a c such that

sup kTn k ≤ c (2.134)
n

Proof. Let us define

Ak = {ξ : sup kTn ξk ≤ k} , (2.135)
n
S
and it is easy to see that Ak is a closed set and we have V = k Ak . Since V is complete, by Baire theorem
some Ak contains an open ball, say,
B0 = Br (ξ0 ) ⊂ Ak0 . (2.136)
For an arbitrary ξ 6= 0 let η ∈ B0 be given by

r ξ
η = ξ0 + . (2.137)
2 kξk

and since η ∈ Ak0 we have supn kTn ηk ≤ k0 . We have

2kξk
ξ = (η − ξ0 ) (2.138)
r
and for any n

2kξk 2kξk 4k0

kTn ξk = kTn (η − ξ0 )k ≤ , kTn ηk + kTn ξ0 )k ≤ kξk (2.139)
r r r
and hence
4k0
sup kTn k ≤ . (2.140)
n r
This concludes the proof.
2.7. BAIRE THEOREM AND UNIFORM BOUNDEDNESS THEOREM 31

Example 2.60. Bilinear functionals Let V and W be Banach spaces. Consider a bilinear functional

B :V ×W →C (2.141)

which is continuous in each variable. That is for fixed ξ ∈ V

B(ξ, ·) : W → C (2.142)

is linear and continuous and for fixed ηinW

B(·, η) : V → C (2.143)

is linear and continuous. We claim that continuity in each variable implies continuity, i.e., if (ξn , ηn ) →
(0, 0) then B(ξn , ηn ) → 0.
Proof. Define Tn : W → C by
Tn η = B(ξn , η) . (2.144)
For any n, by the continuity in the second variable we have that Tn is a bounded operator. For any ηinW by
continuity in the first variable we have limn kTn ηk = 0 and hence supn kTn ηk ≤ cη < ∞. By the uniform
boundedness theorem there exists a c > 0 such that

sup kTn k ≤ c (2.145)

n

or
|B(ξn , ηn )| = kTn ηn k ≤ ckηn k −→ 0 (2.146)
as n → ∞. Hence we have continuity.

Note further that continuity of B is equivalent to

|B(ξ, η)| ≤ ckξkkηk (2.147)

for all ξ, η. And this proved exactly as for linear maps.

We apply this to symmetric operators in Hilbert spaces and we show that unbounded symmetric operators
are necessarily defined only on a subspace of H but not on all of H.
Theorem 2.61. (Hellinger-Toeplitz) Let H be a Hilbert space and let T : H → H be a linear operator
defined on all H and we have (T ξ, η) = (ξ, T η) for all ξ, η ∈ H. Then T is bounded.
Corollary 2.62. Let H be a Hilbert space and let T : DT → H be a linear operator defined on all DT ⊂ H
and we have (T ξ, η) = (ξ, T η) for all ξ, η ∈ DT . Then DT 6= H
Proof. The bilinear map
B(ξ, η) = (T ξ, η) (2.148)
is continuous in η and since
B(ξ, η) = (ξ, T η) (2.149)
it is also continuous in ξ. By the previous example B is continuous in both variables jointly and thus there
exists c > 0 such that
|(T ξ, η)| ≤ ckξkkηk (2.150)
32 CHAPTER 2. NORMED VECTOR SPACES

for all ξ, η ∈ H. In particular if η = T ξ we have

kT ξk2 = (T ξ, T ξ) = (T ξ, η) ≤ ckξkηk = ckξkkT ξk . (2.151)

Hence we have kT k ≤ c.

Example 2.63. (Convergence of Fourier series) Consider a function f periodic of period 2π. Its Fourier
coefficients are
Z 2π
1
cn = f (t)e−int dt , n ∈ Z (2.152)
2π 0

and the Fourier partial sums are

X
SN (f )(t) = cn einx (2.153)
|k|≤N

As one learn in an analysis class we have pointwise (or even uniform) convergence of SN (f )(x)tof (x) if
the function f is sufficiently smooth (say f is C 1 ). At discontinuity points f may or may not converge but
interestingly enough even at points where f is continuous SN (f ) need not converge. One can construct
explicit examples but we prove this here using the uniform boundedness theorem. First we recall that, using
trigonometric formulas one can write

2π
sin((n + 21 )t)
Z
1
SN (f )(t) = DN (t − s)f (s) ds with DN (t) = (2.154)
2π 0 sin( 21 t)

We apply the uniform boundedness theorem by consider the Banach space X of continuous periodic of
period 2π with kf k = supt |f (t)| and let us define the linear functional

λn (f ) = Sn (f )(0) (2.155)

One checks that

Z 2π
1
|λn (f )| ≤ kf k |DN (s)| ds (2.156)
2π 0

By using the same argument as for computing the norm of the Fredholm integral operator we obtain

Z 2π
1
kλn k = |Dn (s)| ds (2.157)
2π 0
2.8. OPEN MAPPING AND CLOSED GRAPH THEOREMS 33

Now we use the inequality | sin(t)| < t on [0, π] and obtain

 sin((n + 21 )t) 
Z 2π  
1
kλn k =   dt
2π 0  sin( 21 t) 
 
1 2π sin((n + 12 )t)
Z
> dt
π 0 t
1 (2n+1)π |sin v|
Z
= dt
π 0 v
2n Z
1 X (k+1)π |sin v|
= dt
π v
k=0 kπ
2n Z (k+1)π
1X 1
≥ |sin v| dt
π (k + 1)π kπ
k=0
2n
2 X 1
= −→ 0 (2.158)
π2 (k + 1)
k=0

as n → ∞. Hence the sequence kλn k is unbounded. By the uniform boundedness theorem this implies that
supf |λn (f )| cannot be bounded and hence there exists at least one f such that λn (f ) = Sn (f )(0) diverges.
That is the Fourier series diverges at t = 0.

2.8 Open mapping and closed graph theorems

After the Hahn-Banach theorem and the uniform boundedness theorem we now attack the third ”big” the-
orem of functional analysis, the open mapping theorem. It is well-known that the continuity of a map
f : X → Y between metric spaces is equivalent to the property that for any open set O ⊂ Y , the set f −1 (0)
is also open. By contrast let us define

Definition 2.64. Let M and N be metric spaces. The F : DF → N with domain DF is called an open
mapping if for every open set O ∈ DF the image F (O) is an open set in Y .

Remark 2.65. In general continuous mapping are not open, e.g. f (t) = sin(t) maps the open set (0, 2π)
onto the closed set [−1, 1].

We have the remarkable result

Theorem 2.66. (Open mapping theorem) Let V and W be Banach spaces. A surjective bounded linear
operator T : V → W is an open mapping.

An immediate consequence of this theorem is

Theorem 2.67. Inverse mapping theorem LetLet V and W be Banach spaces and let T : V → W be a
bounded linear bijective map. Then T −1 is continuous and thus bounded.

The proof of theorem 2.66 relies on

34 CHAPTER 2. NORMED VECTOR SPACES

Proposition 2.68. A bounded linear operator T : V → W from a Banach space V onto a Banch space
W has the property that the image of T (B0 ) of the open unit ball B0 = B1 (0) ⊂ V contains an open ball
around 0 ∈ W
Proof. The proof consists of three steps:
1. The closure of the open ball B1 = B1/2 (0) contains an open ball B ∗ .

2. If Bn = B2−n (0) then the closure T (Bn ) contains an open ball Vn around 0 ∈ Y .
3. T (B0 ) contains an open ball around 0.

Step 1.: Let B1 = B1/2 (0). Then we have

∞
[
V = kB1 . (2.159)
k=1

Since T is surjective and linear,

∞
[ ∞
[ ∞
[


W = T (V ) = T kB1 = kT (B1 ) = kT (B1 ) , (2.160)
k=1 k=1 k=1

where the last equality follows from the fact that the union is Y , hence we did not add any points by taking
the closure. By Baire category theorem there exist some k such that kT (B1 ) contains an open ball and hence
T (B1 ) contains an open ball, say B ∗ = B (η0 ) ⊂ T (B1 ). Then we also have

B ∗ − η0 = B (0) ⊂ T (B1 ) − η0 (2.161)

Step 2.: We prove that B ∗ − η0 ⊂ T (B0 ) where B0 = B1 (0) by proving that (see (2.161)) that

T (B1 ) − η0 ⊂ T (B0 ) . (2.162)

Let η ∈ T (B1 ) − η0 . Then η + η0 ∈ T (B1 ) and we remember that also η0 ∈ T (B1 ). Then there exists
αn = T βn ∈ T (B1 ) and δn = T γn ∈ T (B1 ) such that

lim αn = η + η0 , , lim δn = η0 . (2.163)

n n

Since βn , γn ∈ B1 of radius 1/2 we have kβn − γn k < 1 and βn − γn ∈ B0 . From

lim T (βn − γn ) = η (2.164)

n→∞

we conclude that η ∈ T (B0 ). This concludes the proof of (2.162) and thus

B ∗ − η0 = B (0) ⊂ T (B0 ) (2.165)

By the linearity of T if Bn = B2−n (0), we have T (Bn ) = 2−n T (B0 ) and thus

Vn = B/2n (0) ⊂ T (Bn ) (2.166)

2.8. OPEN MAPPING AND CLOSED GRAPH THEOREMS 35

Step 3: Finally we prove that

V1 = B/2 (0) ⊂ T (B0 ) . (2.167)
Let η ∈ V1 . By (2.166) with n = 1, for any  > 0 there exist ξ1 ∈ B1 such that

kη − T ξ1 k ≤ /4 . (2.168)

Then η − T ξ1 ∈ V2 and by (2.166) with n = 2 we see that η − T ξ1 ∈ V2 ⊂ T (B2 ). Repeating the same
argument we find ξ2 ∈ B2 such that

kη − T ξ1 − T ξ2 k ≤ /8 (2.169)

and hence η − T ξ1 − T ξ2 ∈ V3 ⊂ T (B3 ), an so on. In the nth step we select ξn inBn such that
n
X 
kη − T ξk k ≤ (2.170)
2n+1
k=1

Let us set ζn = ξ1 + · · · ξn then since kξk k ≤ 2−k ζn is Cauchy sequence and ζn → ξ ∈ V . Also ξ ∈ B0
since B0 has radius 1. Since T is continuous T ζn → T x and by (2.170) we have T ξ = η and so η ∈ T (B0 ).

Proof of theorem 2.66. This follows from the previous proposition by using linearity to translate and dilate
balls.

The next theorem is also an immediate consequence of the open mapping theorem. Suppose T : D(T ) →
W where D(T ) is a subspace of V and is called the domain of T , and V, W are Banach spaces. We do not
assume that T is bounded and in general D(T ) 6= V .
Definition 2.69. The graph of T is the set

Γ(T ) = {[ξ, η] ∈ V × W ; ξ ∈ DT , η = T ξ} (2.171)

Note that Γ(T ) is a subspace of V × W which we can make it into a normed vector space with the norm

k[ξ, η]k := kξk + kηk (2.172)

Definition 2.70. A map T : D(T ) → W (V ,W Banach spaces) is closed if the graph Γ(T ) is closed.
Equivalently T is closed if for any sequence {ξn } such that

ξn −→ ξ and T ξn → eta (2.173)

then
Tx = η (2.174)
Clearly bounded operators are closed, but the converse is not true.
Theorem 2.71. (Closed graph theorem) Let V and W be Banach spaces and T : D(T ) → W a linear
operator. If D(T ) is closed and if T is closed then T is bounded.
36 CHAPTER 2. NORMED VECTOR SPACES

Corollary 2.72. Let V and W be Banach spaces and T : V → W a linear operator. Then T is closed if
and only if T is bounded.

Proof. If D(T ) and Γ(T ) are closed then they are Banach spaces and we define the map P Γ(T ) → D(T )
by
P ([ξ, T ξ]) = ξ (2.175)
i.e., P is the projection on the first component. The map P is a bijection and by the inverse mapping theorem
P −1 is bounded. This means

kP −1 ξk ≤ ckξk or kξk + kT ξk ≤ Ckξk (2.176)

. Thus we have kT ξk ≤ (C − 1)kxik and so T is bounded.

As an application we prove
Theorem 2.73. Suppose than V is a Banach space with respect to the two norms kξk1 and kxik2 which
are compatible in the sense that if a sequence {ξn } converges in both norms then the two limits are equal.
Then the two norms are equivalent in the sense that there exists constants cand C such that ckξk1 ≤ kξk2 ≤
Ckξk1 .

Proof. Consider the identity map 1 : (V, k · k1 ) → (V, k · k2 ) given by 1(ξ) = ξ. Compatibility means that
the map 1 is closed. By the closed graph theorem it is bounded in both directions.

We conclude with an example of a closed operator which is not bounded

Example 2.74. Let V = C[0, 1] with the sup-norm and let T f (t) = f 0 (t) be the differentiation operator
with domain D(T ) = C 1 [0, 1] of continuously differentiable functions. The operator is unbounded (take
fn = tn ) and closed since if fn converges uniformly to f and fn0 converge uniformly to h then we have
using uniform convergence to interchange integral and limit
Z t Z t Z t
h(s) ds = lim fn0 (s) ds = lim fn0 (s) ds = f (t) − f (0) , (2.177)
0 0 n 0 n

that is f 0 (t) = h.

2.9 Exercises
Exercise 4. Prove that lp is a Banach space.
Exercise 5. Consider the normed vector space BV [a, b] with kf kBV = f (a) + V (f ) where V (f ) is the
variation of f on [a, b] and let kf k∞ = supt |f (t)|.

1. Show that the norm kf k∞ is weaker than kf kBV .

2. Show that BV [a, b] with k · kBV is a Banach space. (You may use part 1.)

Exercise 6. Prove Theorem 2.15.

2.9. EXERCISES 37

Exercise 7. Compute the spectral radius of the Volterra integral K operator given in Eq. (1.35) of Exercise
2
Rb
Exercise 8. Let m denote Lebesgue measure. Show that the integral operator T f = a k(t, s)f (s) ds
defined a bounded operator on L2 ([a, b], m) provided k ∈ L2 ([a, b] × [a, b], m × m).
Exercise 9. Let V be a Banach space.

1. Show that if T ∈ L(V ) then eX defined by

∞
X 1 k
eX = X , (2.178)
k!
k=1

defines a linear operator in L(V ).

2. Show that eX+Y = eX eY whenever X and Y commute, i.e. XY = Y X.

3. Show that eX is invertible and (eX )−1 = e−X .

4. Show that the (non-linear) map X 7→ eX is differentiable and compute the derivative (eX )0 . Show
that
(eX )0 6= eX . (2.179)

Exercise 10. Show that lp 1 ≤ p < ∞ is separable but that l∞ is not separable.
Exercise 11. To show that (l∞ )0 6= l1 consider the subspace c and define a function λ on c by

λ(ξ) = lim ξn . (2.180)

n→∞

Show that λ extends to functional on l∞ and deduce from this that (l∞ )0 6= l1 .
Exercise 12. Let V and W be normed vector spaces and T ∈ L(V, W ). If T −1 exists and is bounded show
that (T −1 )0 = (T 0 )−1 .
Exercise 13. To illustrate the Hahn-Banach theorem and its consequences:

1. For Corollary 2.39, consider the functional λ on the euclidean plane R2 given by λ(ξ) = α1 x1 +α2 x2 ,
its linear extensions λ̃ to R3 and the corresponding norms λ̃.

2. For Corollary 2.40, let V = R2 , find the functional λ.

Exercise 14. Let V , W be normed vector spaces and {Tn } a sequence of bounded operators.

1. Suppose V is a Banach space. Show that if {Tn } converges strongly to T then T is a bounded operator.

2. If V is not complete then T need not to be bounded. To see this let E ⊂ l∞ be the subspace of
sequence which contains only a finite number of nonzero terms and define A by

T (x1 , x2 , · · · ) = (x1 , 2x2 , 3x3 , · · · ) (2.181)

Show that A is not bounded but can be written as the strong limit of a sequence bounded operators.
38 CHAPTER 2. NORMED VECTOR SPACES

Exercise 15. Let V , W be Banach spaces and {Tn } a sequence of bounded operators. Show that the
sequence {Tn } converges strongly to a bounded operator T if

1. The sequence {kTn k} is bounded.

2. The sequence {kTn ξk} converges for ξ in a dense set in V .

Hint: Use the first part of previous exercise.

Exercise 16. 1. Show by an example that in Baire’s category theorem the completeness condition cannot
be omitted.

2. Show by an example that in Baire’s category theorem condition of the countability of the decomposi-
tion (see Eq. (2.126)) cannot be omitted.
Hint: Do not look for complicated metric spaces.
Exercise 17. (Weak convergence) Let V be a normed vector space. We say that a sequence {ξn } converges
weakly to ξ if
lim λ(ξn ) = λ(ξ) (2.182)
n
0
for all λ ∈ V . Show the following properties of weak convergence

1. The weak limit of {ξn }, if it exists, is unique.

2. If ξn converges weakly to ξ then kξn k is bounded. Hint: Use the uniform boundedness theorem.

3. If ξn converges to ξ then ξn converges to ξ weakly but the converse is not necessarily true. Hint: Try
a separable Hilbert space or lp ....

4. Show that in finite dimensional spaces weak convergence and strong convergence are equivalent.

5. Show that if {ξn } is a sequence such that (i) supn kξn k ≤ c < ∞ and (ii) limn λ(ξn ) = λ(ξ) for a
dense set of λ in V 0 then ξn converges weakly to ξ.

6. Show that if V = lp , p > 1, then ξn converges weakly to ξ if and only if the sequence {kξn k} is
(n) n) (n)
bounded and limn xk = xk for all k. (Here we have denoted ξn = (x1 , x2 , · · · ...))

7. Show that in l1 weak convergence and strong convergence are equivalent.

Exercise 18. In this problem we call a map an open mapping if for any open set O ⊂ V the image T (O) is
an open set of T (V ).

1. Suppose N is a closed subspace of a Banach space V and consider the quotient space quotient V /N .
Show that V /N can be made into a Banach space with the norm

kξk
b = inf kηk . (2.183)
η∈ξb

2. Show that if V is a Banach space then any λ ∈ V 0 is an open mapping. Show also that if T : V → W
has finite-dimensional range then T is an open mapping. Hint: Use part 1.

3. Find an example of a bounded linear map which is not an open mapping.

2.9. EXERCISES 39

Exercise 19. Let V and W be Banach spaces and T : V → W a bounded linear maps such that the range
of T , R(T ) is finite-codimensional subspace of W . Show that R(T ) is closed.
Hint: Use the closed graph theorem. Extend T to V ⊕ Z such that the range of T is all of W .
Exercise 20. Suppose V is a Banach space, Y and Z closed complementary subspaces of V such that
V = Y ⊕ Z. Let PY be the projection on Y along Z and PZ be the projection on Z along Y . Show that PY
and PZ are continuous.
Hint: Use the closed graph theorem
40 CHAPTER 2. NORMED VECTOR SPACES
Chapter 3

Spectral theory

3.1 Spectrum and resolvent

Let us first recall the spectral theory in finite-dimensional space (linear algebra). For T ∈ L(Cn ), λ ∈ C
is called an eigenvalue of T if T − µ1 is singular, i.e., if det(T − µ1) = 0. The set of eigenvalues of T is
called the spectrum of T . Since det(T − µ1) is a polynomial of order n then the spectrum of T contains
at least one point and at most n points. If µ is an eigenvalue, then the eigenvalue equation T ξ = µξ has
at least one non-trivial solution. Such solutions are called eigenvectors for the eigenvalue λ. If µ is not an
eigenvalue then T − µ1 is regular and so (T − µ1)−1 exists.
In infinite dimensional vector spaces is the spectral analysis hugely more complicated, but also much
more interesting than in finite-dimensional spaces. From a practical point of view understanding the spec-
trum of an operator is essential part of understanding the operator itself!

Convention/notation: V is a complex vector space and T ∈ L(V ). For µ ∈ C we set

Tµ = T − µ1 (3.1)

Definition 3.1. Let T ∈ L(V )

1. µ is a regular value of T if Tµ is bijective. (Hence Tµ−1 ∈ L(V ) by the inverse mapping theorem).

2. The resolvent set of T, denoted by ρ(T ) is the set of regular values of T ,

ρ(T ) = {µ ∈ C : µ regular value of T } , (3.2)

and the resolvent of T is

Rµ (T ) = (T − µ1)−1 (3.3)

3. The spectrum of T is
σ(T ) = C \ ρ(T ) , (3.4)
and λ ∈ σ(T ) is called a spectral value.

41
42 CHAPTER 3. SPECTRAL THEORY

4. µ ∈ σ(T ) is called an eigenvalue of T if the equation

T ξ = µξ (3.5)

has a non-trivial solution ξ ∈ V . Such a solution ξ is called an eigenvector for the eigenvalue µ and
the subspace
Eµ (T ) = {ξ ; T ξ = µξ} (3.6)
is called the eigenspace of T for the eigenvalue µ. The point spectrum of T is

σp (T ) = {µ ; µ eigenvalue of T } (3.7)

5. The continuous spectrum of T is

σc (T ) = {µ ∈ σ(T ) \ σp (T ) ; Dµ = Tµ (V ) is dense in V and Tµ−1 exists, but is unbounded}

(3.8)

6. The residual spectrum of T is

σc (T ) = {µ ∈ σ(T ) \ σp (T ) ; Dµ = Tµ (V ) is not dense in V } (3.9)

We clearly have
Lemma 3.2. The sets σp (T ), σc (T ), and σr (T ) are mutually disjoint and

C = ρ(T ) ∪ σ(T ) = ρ(T ) ∪ σp (T ) ∪ σc (T ) ∪ σr (T ) (3.10)

It is not immediately obvious that σr (T ) is not empty.

Example 3.3. Let T be the right shift operator on l2 , i.e for ξ = (x1 , x2 , · · · ) we have

T ξ = (0, x1 , x2 , · · · ) . (3.11)

Then 0 is a spectral value since T (l2 ) = {ξ ; x1 = 0} is not dense in l2 . On T (l2 ) the inverse if the left
shit Sξ = (ξ2 , ξ3 , · · · ). On the other hand 0 is not en egenvalue since the equation T ξ = 0 only the trivial
solution ξ = 0. So 0 ∈ σr (T ).

Example 3.4. Let T be the multiplication operator on L2 [0, 1] given by

T f (t) = tf (t) . (3.12)

We have kT f k2 = t2 |f (t)|2 dt ≤ kf k2 showing that kT k ≤ 1. it is left to the reader to prove that

R

kT k = 1. note that

• If µ ∈
/ [0, 1] then µ ∈ ρ(T ). Indeed we have
1
Rµ f (t) = f (t) (3.13)
t−µ

and kRµ k ≤ dist(µ, [0, 1])−1

3.1. SPECTRUM AND RESOLVENT 43

• If µ ∈ [0, 1], µ is not eigenvalue since tf (t) = µf (t) has only the solution f (t) = 0 a.e.

• If µ ∈ [0, 1], then Tµ (L2 [0, 1]) is dense since it contains all L2 functions which vanish in some
neighborhood of t0 = µ. Then Tµ−1 f (t) = (t − µ)−1 f (t) is defined on a dense set but unbounded.

Thus we have
σ(T ) = σc (T ) = [0, 1], σp (T ) = σr (T ) = ∅ (3.14)

The map

R· (T ) : ρ(T ) −→ L(V )
µ 7−→ Rµ (T )

is an operator-valued function on the resolvent set ρ(T ). We first investigate in this function can be under-
stood as an ”analytic” function in some way. Since L(V ) is a Banach space we need to develop a bit the
theory of analytic Banach-spaced valued functions.

Definition 3.5. Let Ω ⊂ C be an open set, V a Banach space and

ξ(·) : Ω −→ V (3.15)

a map with valued in the Banach space V .

1. The map ξ(z) is called strongly differentiable at z0 ∈ Ω if

1
ξ 0 (z0 ) = lim (ξ(z0 + h) − ξ(z0 )) (3.16)
h→0 h

exists. The map ξ(z) is called strongly analytic in Ω if it is strongly differentiable at any z ∈ Ω.

2. The map ξ(z) is called weakly differentiable at z0 ∈ Ω if for any linear functional λ ∈ V 0 the complex
valued function
z 7→ λ(ξ(z)) (3.17)
is differentiable at z0 . The map ξ(z) is called weakly analytic in Ω if it is weakly differentiable at any
z ∈ Ω.

We have

Theorem 3.6. A Banach-space valued map ξ(z) is strongly analytic if and only if it is weakly analytic.

As a warm-up we have

Lemma 3.7. Let {ξn } be a sequence in the Banach space V . Then the sequence {ξn } is Cauchy if and only
if the sequence {λ(ξn )} is Cauchy, uniformly for all λ ∈ V 0 , kλk ≤ 1.
44 CHAPTER 3. SPECTRAL THEORY

Proof. On one hand we have kλ(ξn ) − λ(ξm )k ≤ kλk|ξn − ξm k = kξn − ξm k for kλk ≤ 1. On the other
hand we have kξn − ξm k = supkλk=1 |λ(ξn ) − λ(ξm )| by Theorem 2.42.

Proof of Theorem 3.6. Clearly strongly analytic implies weakly analytic. So let us assume that ξ is weakly
analytic in Ω ⊂ C, let z0 ∈ Ω and Γ a circle around z0 contained in Ω. For any λ ∈ V 0 we have by Cauchy
Theorem
  Z    
ξ(z0 + h) − ξ(z0 ) d 1 1 1 1 1
λ − λ(ξ(z0 )) = − − λ(ξ(z) dz .
h dz 2πi Γ h z − (z0 + h) z − z0 (z − z0 )2
(3.18)
Since λ(ξ(z)) is continuous on Γ there exists a constant cλ such that

sup |λ(ξ(z))| ≤ cλ (3.19)

z inΓ

So the family of

ξ(z) ; V 0 −→ C
λ 7−→ λ(ξ(z)) (3.20)

for z ∈ Γ is a point wise bounded family of linear maps (use the isomorphism V → V 00 ). From the uniform
boundedness theorem there exists a c < ∞ such that

sup kξ(z)k = c (3.21)

z∈Γ

So we can bound (3.18) by

Z    
1  1 1 1 1 
|(3.18)| ≤ ckλk  − − 2
 d|z ≤ Const|h|kλk (3.22)
2π Γ h z − (z0 + h) z − z0 (z − z0 ) 

Therefore h1 λ(ξ(z0 + h) − ξ(z0 )) converges uniformly in λ for all λ with kλk ≤ 1. By the previous lemma
this implies that h1 (ξ(z0 + h) − ξ(z0 )) is Cauchy for |h| → 0. This concludes the proof.

The theorem just proved is very useful since it allows us to speak simply of analytic Banach-space valued
functions. All the theorem from analytic function theory can be used since they apply to the ordinary analytic
function λ(ξ(z)) and so we can ”lift” these results to the strongly analytic function ξ(z).

Theorem 3.8. Let V be a Banach space and T ∈ L(V ).

1. ρ(T ) is an open set in C.

2. The resolvent µ 7→ Rµ (T ) is analytic in ρ(T ).

Proof. We use the Neumann series (see Theorem 2.23). if µ0 ∈ ρ(T ) consider the open ball

Ω0 = µ : |µ − µ0 | < kRµ0 k−1 .


(3.23)
3.1. SPECTRUM AND RESOLVENT 45

We have

Tµ = Tµ0 − (µ − µ0 )1
= Tµ0 [1 − (µ − µ0 )Rµ0 ] . (3.24)

Since k(µ − µ0 )Rµ0 k < 1 we can invert Tµ and using Neumann series we obtain
∞
−1
X
Rµ = [1 − (µ − µ0 )Rµ0 ] Rµ0 = (µ − µ0 )k Rµk+1
0
. (3.25)
k=0

This is a convergent series in L(V ) and so Rµ exist for µ ∈ Ω0 and so Ω0 ⊂ ρ(T ). The power series also
shows that Rλ is is analytic.

Theorem 3.9. Let V be a Banach space and T ∈ L(V ). Then the spectrum σ(T ) is closed and non-empty.

Proof. As the complement of ρ(T ), the spectrum σ(T ) is closed. By contradiction let us assume that σ(T )
is empty. Then ρ(T ) = C and Rµ is an entire function. For |µ| ≥ kT k we have the convergent series
∞
X
Rµ (T ) = (T − µ1)−1 = −µ−1 (1 − µ−1 T )−1 = −µ−1 (µ−1 T )k (3.26)
k=0

and
∞
−1
X µ−1
kRµ (T )k ≤ |µ| kµ−1 T kk = . (3.27)
1 − kµ−1 T k
k=0

and thus kRµ (T )k −→ 0 as |µ| → ∞. Hence Rµ (T ) is is a bounded entire function and by Liouville
Theorem it as to be constant and identically null. But this is impossible and so σ(T ) is not empty.

Eq. (3.26) also shows that

Corollary 3.10. We have
σ(T ) ⊂ {µ : |µ| ≤ kT k} (3.28)
This can be strengthened into the following theorem which justifies the terminology spectral radius for
r(T ).
Theorem 3.11. Let V be a Banach space and T ∈ L(V ). Then we have

r(T ) = sup |µ| = max |µ| (3.29)

µ∈σ(T ) µ∈σ(T )

Proof. The series (3.26) is nothing but the Laurent series P for Rµ (expansion in power of 1/µ around ∞).
From complex analysis we know for a power series f (z) = n≥0 an z n with convergence radius r converges
absolutely for |z| < r but is not analytic in {|z| < (r + )}. For the power series (3.26) this means that the
series converges exactly outside the circle of radius supν∈σ(T |ν|.
We also know that the convergence radius of a power series is given by the formula

r−1 = lim sup |an |1/n , (3.30)

n→∞
46 CHAPTER 3. SPECTRAL THEORY

that is in our case

r−1 = lim sup kT n k1/n = lim kT n k1/n = r(T ) (3.31)
n→∞ n→∞

Finally note that since the spectra is closed we can replace the sup by a max.

Note that L(V ) in addition of being a Banach space is also an algebra with the property that kT Sk ≤
kT kkSk. So we can certainly define polynomial of an element of L(V )
N
X
p(T ) = ak T k (3.32)
k=1

an z n we can define
P
or more generally for entire function f (z) = n

∞
X
f (T ) = an T n (3.33)
k=1

Note also that the formula (3.33) makes sense if f has a convergence radius larger than the spectral radius
r(T ).
In order to be more general we do not use a power series but instead use Cauchy integral formula.
Definition 3.12. Let T ∈ L(V ), f (z) a function analytic in a domain Ω containing σ(T ). Let C be a contour
in Ω ∩ ρ(T ) such that C winds once around any point in σ(T ) and winds zero time around any point in ΩC .
Set Z Z
1 −1 1
f (T ) = (z − T ) f (z)dz = − Rz (T )f (z)dz . (3.34)
2πi C 2πi C
Note that the definition does not depend on the choice of C.
As a preparation for the next theorem we prove
Lemma 3.13. (Resolvent formula)

Rµ (T ) − Rν (T )
Rµ (T )Rν (T ) = (3.35)
µ−ν

Proof. We have
(T − µ1) − (T − ν1) = (ν − µ)1 (3.36)
and so multiplying by Rµ Rν we have

Rν − Rµ = (ν − µ)Rµ Rν . (3.37)

The next theorem provides the basis for the functional calculus.
Theorem 3.14. (Functional calculus and spectral mapping theorem) Let T ∈ L(V ).

1. If f is a polynomial or analytic in a disk of radius greater than σ(T ) then the definition (3.34) coin-
cides with the formulas (3.32) and (3.33).
3.1. SPECTRUM AND RESOLVENT 47

2. The map (3.34) from the algebra of analytic functions on an open set containing σ(T ) into L(V ) is a
homomorphism.

3.
σ(f (T )) = f (σ(T )) (3.38)

4. If f is analytic in an open set containing σ(T ) and g is analytic in an open set containing f (σ(T )). If
h = g ◦ f then
h(T ) = f (g(T )) (3.39)
P∞
Proof. 1. Using the representation Rµ (T ) = − k=0 µ−k−1 T k we obtain by Cauchy integral formula
Z
1
T n
= − (T − z1)−1 z n dz (3.40)
2πi C

and this shows that the formulas coincide.

For 2. we note that the mapping f 7→ f (T ) is obviously linear. To show it is multiplicative we use the
resolvent formula of Lemma 3.13. If f and g are analytic in a domain containing σ(T ) we pick two contours
C and D as in definition 3.12 with D lying inside C. Then we have
 2 Z Z
1
f (T )g(T ) = Rz (T )Rw (T )f (z)g(w)dzdw
2πi C D
 2 Z Z
1 Rz (T ) − Rw (T )
= f (z)g(w)dzdw
2πi
C D z−w
 2 Z Z 
1
= (z − w)−1 g(w)dw Rz (T )f (z)dz
2πi C D
 2 Z Z 
1 −1
− (z − w) f (z)dz Rw (T )g(w)dw (3.41)
2πi D C

Since D lies inside C we have D (z − w)−1 g(w)dw = 0 while C (z − w)−1 f (z) = 2πif (w) and thus
R R

Z
f (T )g(T ) = − Rw (T )f (w)g(w)dw . (3.42)
D

and thus f (T )g(T ) = h(T ).

For 3. we have to show that µ belongs to the spectrum of f (T ) if and only if µ is of the form

µ = f (ν) , ν ∈ σ(T ) . (3.43)

If µ is not of the form (3.43), then f (z) − µ does not vanish on σ(T ). Therefore g(z) ≡ (f (z) − µ)−1 is
analytic in an open set containing σ(T ). According to part 2. (f (T ) − µ1)g(T ) = h(T ) = 1. So g(T ) is
the inverse of (f (T ) − µ) and so µ ∈
/ σ(f (T )).
Conversely suppose that µ is of the form (3.43). Define k(z) by

f (z) − f (ν)
k(z) = . (3.44)
z−ν
48 CHAPTER 3. SPECTRAL THEORY

The function k(z) is analytic in an open set containing σ(T ) so k(T ) can be defined using by (3.34). Since
(z − ν)k(z) = f (z) − f (ν) by part 2. we have

(T − ν1)k(T ) = f (T ) − f (ν)1 . (3.45)

Since ν ∈ σ(T ) the first factor is not invertible and so f (T ) − f (ν) is not invertible either.
For 4. by assumption g(w) is analytic on f (σ(T )). Since by 3. the spectrum of f (T ) is f (σ(T )) we can
apply (3.34) to g in place of f and f (T ) in place of T :
Z
1
g(f (T )) = (w1 − f (T ))−1 g(w)dw (3.46)
2πi D
If w ∈ D then w − f (z) is an analytic function on σ(T ), then applying (3.34) again
Z
−1 1
(w − f (T )) = (z1 − T )−1 (w − f (z))−1 dz (3.47)
2πi C
provided C does not wind around any point of D. Combining the tow formulas we find
 2 Z Z
1
g(f (T )) = (z − T )−1 (w − f (z))−1 g(w)dzdz . (3.48)
2πi D C

We interchange the order to the integral and since D winds around every point z ∈ C we have by Cauchy
integral formula Z
1
(w − f (z))−1 g(w)dw = g(f (z)) = h(z) . (3.49)
2πi D
Setting this back in (3.48) we find that g(f (T )) = h(T ).

Suppose σ(T ) can be decomposed into N pairwise disjoint closed components:

σ(T ) = σ1 ∪ · · · ∪ σN , σj ∩ σk = ∅ (3.50)

Since the σi are closed they are at positive distance from each other and we can pick contours Cj that winds
once around each point of σj but not σk , k 6= j. We set
Z
1
Pj = (z1 − T )−1 dz (3.51)
2πi Cj

Theorem 3.15. The Pj are disjoint projections, i.e.

Pj2 = Pj and Pj Pk = 0 f or j 6= k (3.52)

and we have X
Pj = 1 . (3.53)
j
SN
Proof. Pick open set Ωi such that Ωi contains σi and the Ωi are pairwise disjoint. Set Ω = j=1 Ωj . Then
consider the function fi (z) which is equal to 1 in Ωi and 0 in Ω \ Ωi . These functions are analytic in Ω and
satisfy fi (z)2 = fi (z) as well as fi fj = 0 for i 6= j. By functional calculus we obtain the theorem.
3.1. SPECTRUM AND RESOLVENT 49

Theorem 3.16. Let V be a Banch space and T ∈ LV and let T 0 ∈ L(V 0 ) it adjoint. Then

σ(T ) = σ(T 0 ) (3.54)

Corollary 3.17. If T is a bounded operator on a Hilbert space and T ∗ its (Hilbert-)adjoint. Then

σ(T ∗ ) = σ(T ) (3.55)

We need the following lemma. For a bounded linear operator T we denote N (T ) the nullspace of T and
R(T ) the range of T . For a subspace W ⊂ V we define the annihilator of W , denoted by W ⊥ as the set of
linear functional which vanish on W . For a subset W 0 of V 0 we define the annihilator of W 0 (denoted by
W 0 ⊥) as the set of all vectors in V annihilated by all functional in W 0 .
Lemma 3.18. We have
N (T 0 ) = R(T )⊥ , N (T ) = R(T 0 )⊥ (3.56)
Proof. Use the duality relation λ(T ξ) = T 0 λ(ξ). The details are left as an exercise.
The theorem 3.16 is an immediate consequence of
Theorem 3.19. T ∈ L(V ) is invertible if and only if T 0 ∈ L(V 0 ) is invertible.
Proof. IF T is invertible with inverse S then

T S = ST = 1V . (3.57)

Taking the adjoint gives

S 0 T 0 = T 0 S 0 = 1V 0 (3.58)
which shows that S is the inverse of T . If V is a reflexive space then the relation between T and T 0 is
0 0

symmetric so the proof is complete. In case V is not reflexive an additional argument is needed. If T 0 is
invertible then by taking adjoint
T 00 S 00 = S 00 T 00 = 1V 00 (3.59)
Since T 00 and IV 00 restricted to V are equal to T and IV respectively it follows that the nullspace of T is
trivial and so T is one- to-one. So S 00 restricted to the range of T is inverse to T . Suppose now that the range
of T is not all of V . Then we can find by Hahn- Banach a non-zero functional λ ∈ V 0 with λ = 0 on the
range of T . According to the previous lemma such a functional belongs to the nullpsace of T 0 . Since T 0 is
invertible this is impossible.

Example 3.20. We return to the shift T : l2 → l2 given by T ξ = (0, x1 , x2 , · · · ) with adjoint Sξ =

(x2 , x3 , · · · ). We claim that
σ(T ) = σ(S) = {z ∈ C : |z| ≤ 1} (3.60)
It is easy to check that kLk = 1 and similarly that kLn k = 1 for all n and thus r(T ) = 1. So no number z
with |z| ≥ 1 belongs to σ(S). Next let us try to find eigenvalues for L: (x2 , x3 , cdots) = µ(x1 , x2 , cdots)
which implies
xn = µn−1 x1 (3.61)
2
P
We have then n |xn | < ∞ if and only if |µ| < 1. So any µ in the open disk {z : |z| < 1} is an
eigenvalue. Since the spectrum is closed then σ(S)is the close unit disk. Finally the spectrum of T is the
same as the spectrum of S. The reader should check whether T has any eigenvalues.
50 CHAPTER 3. SPECTRAL THEORY

Example 3.21. Nilpotent and quasi-nilpotent operators

An operator is called nilpotent if N k = 0 for some k > 1. Then σ(N k ) = {0k and by the spectral
mapping theorem σ(N ) = {0}.
The Voltera operator acting on C[a, b]
Z t
Kf (t) = K(s, t)f (t) dt (3.62)
0

with continuous kernel k(t, s) was shown to have spectral radius r(T ) = 0 in the exercises. Therefore

σ(K) = {0}. (3.63)

An operator K with σ(K) = {0} but N k 6= 0 for all k is called quasi-nilpotent.

Example 3.22. Fourier transform The Fourier transform T is defined by

Z
fb(k) = T f (k) = f (x)e−i2πxk (3.64)
R

As we have show in Math 623/624 T is an invertible norm preserving map from L2 (R) onto L2 (R). The
inverse is given by Z
f (x) = fb(k)ei2πxk dk . (3.65)
R

Consider the mapping R given by

Rf (x) = f (−x) , (3.66)
and obviously R2 = 1. From (3.65) we have
Z
f (−x) = fb(k)e−i2πxk dk . (3.67)
R

So we have T 2 = R and thus

T4 = 1 (3.68)
By the spectral mapping theorem we immediately obtain that

σ(T ) ⊂ {+1, −1, +i, −i} (3.69)

More details on the spectrum of T will be given in the HWK.

3.2 Compact operators: Basic properties

One of the workhorse of analysis in finite-dimensional spaces is the Bolzano-Weierstrass theorem which
asserts that that a set is compact if and only if it is closed and bounded. In infinite dimensional spaces this
theorem fails.

Theorem 3.23. The closed unit ball of a Banach space V is compact if and only if V is finite dimensional.
3.2. COMPACT OPERATORS: BASIC PROPERTIES 51

Before we prove the theorem note that if H is a separable Hilbert space with an orthonormal basis {ξn }
then we have
kξn − ξm k2 = (ξn − ξm , ξn − ξm ) = kξn k2 + kξm k2 = 2 . (3.70)
√
and so the sequence ξn satisfies kξn k = 1 for all n and kξn − ξm k = 2 for n 6= m. Hence {ξn } has no
convergent subsequence and so the unit ball is not compact.
In order to adapt this argument to a general Banach space we need the following lemma.
Lemma 3.24. Let V be a Banach space and W a proper closed subspace of V . Then there exists

ζ ∈ V with kζk = 1 and kζ − ηk > 1/2 for all η ∈ W . (3.71)

Proof. Since W is a proper there exists ξ ∈ V with ξ ∈

/ W and since W is closed

inf kξ − ηk = d > 0 . (3.72)

η∈W

We pick η0 ∈ W such that kξ − η0 k ≤ 2d and set ζ 0 = ξ − η0 . Then we have

kζ 0 k ≤ 2d and kζ 0 − ηk = kξ − (η + η0 )k ≥ d for all η ∈ W. (3.73)

Finally we set ζ = ζ 0 /kζ 0 k so that kζk = 1 and

kζ − ηk > 1/2 for all η ∈ W. (3.74)

Equipped with this lemma we obtain easily

Proof of Theorem 3.23. Let {ξn } be the sequence constructed inductively as follows. Pick an arbitrary ξ1
with kξ1 k = 1. Then if Vn is the subspace spanned ξ1 , · · · , ξn , it is a closed proper subspace since it is
finite-dimensional and thus by Lemma 3.24 there exists ξn+1 with kξn+1 k = 1 and kξn+1 − ξl k > 1/2 for
l = 1, · · · , n. The sequence {ξn } does not have a convergent subsequence.

A bounded operator has the property to maps bounded sets into bounded sets. A operator will be called
compact if it transforms bounded sets into sets whose closure compact. To make this precise we recall
Definition 3.25. Let X be a complete metric space. The set S is called precompact if its closure S is
compact.

• S is precompact if and only if any sequence {ξn } ⊂ S contains a Cauchy subsequence.

• S is precompact if and only if for any  > 0 S can be covered by finitely many balls of radius less
than .

Definition 3.26. Let V and W be Banach spaces.

1. A linear map T : V → W is called compact if the image of the unit ball T (B1 (0)) is precompact.
(This is implies that T maps any bounded set into a precompact set.)

2. We denote by C(V, W ) the set of all compact operators and set C(V ) ≡ C(V, V )
52 CHAPTER 3. SPECTRAL THEORY

Let us derive some of the elementary properties of compact operator.

Theorem 3.27. Let U, V, and W be Banach spaces.

1. If T, S ∈ C(V, W ) then T + S ∈ C(V, W )

2. If T ∈ C(V, W ) and α ∈ K then αT ∈ C(V, W ).

3. If T ∈ C(U, V ) and S ∈ L(V, W ) then ST ∈ C(U, W )

4. If T ∈ L(U, V ) and S ∈ C(V, W ) then ST ∈ C(U, W )

5. If {Tn } is a sequence in C(V, W ) with kTn − T k → 0 then T ∈ C(V, W ).

This means that the set of compact operators is a closed subspace of the vector space of bounded opera-
tors and that C(V ) is a closed two-sided ideal in the algebra L(V ).

Proof. For 1. let {ξn } be a sequence in B1 (0) then since T is compact there exists a subsequence {ξnk }
such that {T ξnk } converges. Since S is compact there exists a subsubsequence {ξnkj } such that {Sξnkj }
and {(T + S)ξnkj } converges too.
2. is a special case of 3. and 3. itself follows from the fact that a bounded map maps precompact set into
precompact sets. In turn 4. is obvious.
For 5. given  > 0 we pick n such that kTn − T k ≤ /2 and since Tn is compact then Tn (B1 (0)) can be
covered by finitely many balls of radius /2. Then T( B1 (0)) can be covered by finitely many balls of radius
.

Definition 3.28. Let V and W be Banach spaces. A linear operator T ∈ L(V ) is a finite rank operator if its
range R(T ) is finite dimensional.

Clearly finite-rank operators are compact. A finite rank operator T ∈ L(V, W ) can always be written as
follows. Pick η1 , · · · , ηN ∈ W and λ1 , · · · , λN ∈ V 0 and set
N
X
Tξ = λj (ξ)ηj (3.75)
j=1

In general it is not true that the set of finite rank operators is dense in the set of compact operators even
on separable Banach spaces. It is true in Hilbert spaces as well as in many standard Banach spaces. We
will prove this property for separable Hilbert spaces in the next section. As we will see in examples one can
sometimes prove that an operator is compact by proving that it can be approximated by finite-rank operators.
We conclude this section with a first series of examples of compact operators. In order to prove com-
pactness we will use two classical results of analysis. The first one characterizes compact sets for spaces of
continuous functions.

Theorem 3.29. (Arzela-Ascoli) Let K be a compact metric space and let C(K) the Banach space of
complex-values continuous functions with norm kf k = supx∈K |f (x)|. Let {fα }α∈I be a collection of
function such that

1. {fα } is uniformly bounded: there exists c < ∞ such that supα kfα k ≤ c.
3.2. COMPACT OPERATORS: BASIC PROPERTIES 53

2. {fα } is equicontinuous: given  > 0 there exists δ > 0 such that

d(x, y) ≤ δ =⇒ kfα (x) − fα (y)k ≤  for all α ∈ I (3.76)

Then the family {fα } is a precompact subset of (C(K).

A related theorem gives a compactness criterion for L2 (Q) where Q is a bounded domain.
Theorem 3.30. (Rellich) Let Q be open and bounded with smooth boundary. Let {fα }α∈I be a family of
functions in L2 (Q) such that
1. The family {fα } is uniformly bounded: there exists c < ∞ such that supα kfα k ≤ c.
2. The derivative of {fα } are uniformly bounded: there exists c < ∞ such that supα k∂xi fα k ≤ c for
i = 1, 2, · · · , d.
Then the family {fα } is a precompact subsets of L2 (Q).
Example 3.31. (Integral operator I) Consider the integral operator
Z b
T f (t) = k(t, s)f (s) ds . (3.77)
a

If k(t, s) ∈ C[a, b] × [a, b] then we show that T : C[a, b] → C[a, b] is compact. Let {fn } be a bounded
sequence, e.g. kfn k ≤ 1 for all n, then we show that the sequence {T fn } satisfies the conditions of Arzela-
Ascoli theorem. We have
Z b
|T f (t) − T f (t0 )| ≤ |k(t, s) − k(t0 , s)||f (s)| ds
a
Z b
≤ kfn k |k(t, s) − k(t0 , s)| ds (3.78)
| {z } a
≤1

By the continuity of k(t, s) the right hand side of (3.78) goes to 0 as t → t0 uniformly in n and thus {T fn }
is an equicontinuous family.

Example 3.32. (Integral operator II) Consider an integral operator on some separable Hilbert space
L2 (X, µ) Z
T f (t) = k(t, s)f (s) dµ . (3.79)

with k(t, s) ∈ L2 (X × X, µ × µ). It was proved in the exercise that T is bounded with
Z Z 1/2
2
kT k ≤ |k(t, s)| dµ(t)dµ(s) . (3.80)

We show that T is compact by showing that is it the limit of a sequence of finite-rank operators. Note
that a finite rank integral operators has the form
N
X Z
T f (t) = fj (s) gj (s)f (s) ds (3.81)
j=1
54 CHAPTER 3. SPECTRAL THEORY

for some fj , gj ∈ L2 . We now expand the k(t, s) (for fixed s) using an o.n.b. {fj } of L2 we have
∞
X Z
k(t, s) = fj (t)kj (s) , kj (s) = k(t, s)fj (t) dµ(t) (3.82)
j=1

By Parseval equality we have

Z ∞
X
|k(t, s)|2 dµ(t) = |kj (s)|2 (3.83)
j=1

and Z Z ∞ Z
X
|k(t, s)|2 dµ(s)dµ(t) = |kj (s)|2 dµ(s) . (3.84)
j=1

Let us define
N
X
kN (t, s) = fj (t)kj (s) (3.85)
j=1
R
and define TN f (t) = kN (t, s)f (s)dµ(s). This is a finite rank operator and we have by the same calcula-
tion has for kT k
Z Z ∞ Z
X
2 2
kT − TN k ≤ |kN (t, s)| dµ(t)dµ(s) = |kj (s)|2 dµ(s) (3.86)
j=N

Because of (3.83), the right side of (3.86) goes to 0 as N → ∞. Hence T as the limit of a sequence of
finite-rank operators is compact.

Example 3.33. (Laplace equation) It is well known that the boundary value problem

∆u = f in Q u = 0 on ∂Q (3.87)

has a unique solution u for every f ∈ C ∞ (Q). Let us denote by S the linear map

u = Sf (3.88)

which gives the solution of (3.87). We claim that S defines a compact map from L2 (Q) into L2 (Q).
We have
Lemma 3.34. Let f be compactly supported in Q. Then we have
Z Z X
|f |2 dx ≤ C |fj |2 dx with fj ≡ ∂xj f . (3.89)
j

Proof. Since f vanishes on the boundary of Q we have, at any point x ∈ Q

Z x
f (x) = f1 dx1 (3.90)
xb
3.2. COMPACT OPERATORS: BASIC PROPERTIES 55

where xb is a point on the boundary of Q with the same x2 , · · · , xd coordinates as x. By Cauchy-Schwartz

we have Z
f (x)2 ≤ d |f1 |2 dx , (3.91)

and integrating over Q gives the result.

2 1/2
RP
Let us denote kf k0 = kf kL2 and kf k1 = ( j |fj | dx) . Then the lemma asserts that kf k0 ≤
ckf k1 for smooth. Now multiply (3.87) by u, integrate over Q and integrate by parts:
Z X Z
− |uj |2 dx = f u dx (3.92)
j Q

Using Cauchy-Schwartz and the lemma this yields

kuk21 ≤ kf k0 kuk0 ≤ , ckf k0 kuk0 . (3.93)

from which we obtain

kuk1 ≤ ckf k0 , and kuk0 ≤ , c2 kf k0 . (3.94)
This shows that the image of the unit ball {kf k0 ≤ 1} is mapped by S to solutions u which satisfies the
conditions of Theorem 3.30. Hence S is a compact map.

Example 3.35. (Heat equation) Let us consider the heat equation (initial value problem)

ut = ∆u , u(x, 0) = u(x) (3.95)

for function u(x, t), with x ∈ Q a bounded open domain with smooth boundary and t > 0. It is well-known
that the initial value problem where u(x, 0) = u(x) is given has a unique solution for all t > 0. Let us
denote by ST the operator mapping the initial condition to the solution at time T

ST (u) = u(x, T ) (3.96)

We have
Theorem 3.36. The map ST : L2 (Q) → L2 (Q) is compact for any T > 0.

Proof. Multiply (3.95) by u and integrate with respect to x ∈ Q and t ∈ [0, T ]. After integrating by parts
one gets
Z T Z Z Z
1 2 1 2
|u| (x, T ) dx −
uut dxdt = |u| (x, 0) dx
0 Q Q 2 Q 2
Z TZ Z TZ X
= u∆u = − |uj |2 dxdt ≤ 0 (3.97)
0 Q 0 Q j

and thus Z Z
|u|2 (x, T ) dx ≤ |u|2 (x, 0) dx (3.98)
Q Q
56 CHAPTER 3. SPECTRAL THEORY

i.e. the L2 norm of u is decreasing in t or

kST k ≤ 1 . (3.99)
Next let us multiply (3.95) by t∆u and integrate with respect to x and t. We get
Z TZ Z TZ
tut ∆u dxdt = t(∆u)2 dxdt ≥ 0 (3.100)
0 Q 0 Q

Integrating by parts the left side of (3.100) with respect to x and then t we find
Z TZ Z TZ X
tut ∆u dxdt = − t utj uj dxdt
0 Q 0 Q j
Z T Z
1 d X
= − t |uj |2 dxdt
2 0 Q dt j
Z T Z Z
1 d X T
= |uj |2 dxdt − |uj (x, T )|2 dx (3.101)
2 0 Q dt j 2 Q

Combining with (3.97) we find

Z Z T Z
T 1 d X
|uj (x, T )|2 dx ≤ |uj |2 dxdt
2 Q 2 0 Q dt j
Z Z
1 2 1
= |u(x, 0)| dx − |u(x, T )|2 dx
2 Q 2 Q
Z
1
≤ |u(x, 0)|2 dx (3.102)
2 Q

Combining (3.98) and (3.102) show that

1
kST uk0 ≤ kuk0 kST uk1 ≤ kuk0 (3.103)
2T
which implies that ST is compact by Rellich Theorem.

3.3 Spectral theory of compact operators

The general result is the following
Theorem 3.37. Let V be a Banach space and T ∈ L(V ) a compact operator.
1. Spectrum:
(a) If λ 6= 0 is in σ(T ) then λ ∈ σp (T ).
(b) σp (T ) is countable and the only possible accumulation point is 0 which may ar may not belong
to σp (T ).
(c) For λ 6= 0, the eigenspace Eλ = {ξ : T ξ = λξ} is finite dimensional.
3.3. SPECTRAL THEORY OF COMPACT OPERATORS 57

(d) The adjoint T 0 has the same eigenvalues and the eigenspaces have the same dimensions.
2. Fredholm Alternative: For Tµ = T − µ1 we have either (a) or (b)
(a) Tµ ξ = η and Tµ0 α = β have unique solutions ξ and α for every η ∈ V , β ∈ V 0 . In particular if
η = 0, β = 0 then ξ = 0 and α = 0.
(b) Tµ ξ = 0 and Tµ0 α = 0 have n linearly independent solutions ξ1 , ξn (resp. α1 , · · · , αn ) and
Tξ = η and T 0 α = β have solution if and only if

αk (η) = 0 , k = 1, · · · , n resp. β(ξk ) = 0k = 1, · · · , n (3.104)

We shall not prove this theorem here (see e.g. Lax for a proof) in full generality but we are going to
concentrate on the case of separable Hilbert spaces. We show first that finite rank operators are dense in
separable Hilbert spaces. This remains true for non-separable Hilbert spaces but maybe surprisingly this in
general not true even for separable Hilbert spaces, although this holds for many of the usual Banach spaces.
Theorem 3.38. Let H be a separable Hilbert space and T ∈ C(H). Then there exists a sequence of finite
rank operator {Tn } such that kTn − T k → 0.
Proof. Pick an o.n.b {ξi } of H and let us set
n
X
Pn ξ = (ξ , ξk )ξk , Tn ≡ T Pn . (3.105)
k=1

Then we have

kT − Tn k = sup k(T − Tn )ξk

kξk=1

= sup kT (1 − Pn )ξk
kξk=1

= sup kT ξk ≡ λn (3.106)
kξk=1,Pn ξ=0

Since Pn+1 ξ = 0 implies Pn ξ = 0 we have

λn+1 ≤ λn , , lim λn = λ exists . (3.107)

n→∞

We need to show λ = 0. For any n let us pick ηn with kηn k = 1, Pn ηn = 0 and kT ηn k ≥ λn /2. We have

lim(ηn , ξ) = 0 for all ξ ∈ H , (3.108)

n

and thus
lim(T ηn , ξ) = (ηn , T ξ) = 0 for all ξ ∈ H , (3.109)
n

Suppose that limn kT ηn k 6= 0, then for any  > 0 there exists a subsequence ηnk such that

kT ξnk k >  , k = 1, 2, · · · (3.110)

Since T is compact there exists a subsubsequence {ηnkj }such that {T ηnkj } is convergent, i.e.

lim T ηnkj = η 6= 0 (3.111)

j
58 CHAPTER 3. SPECTRAL THEORY

But then
lim(T ηnkj , η) = kηk2 6= 0 (3.112)
j

which contradicts (3.109). Therefore we have limn kT ηn k 6= 0 and so λ = 0.

As preparation we have
Theorem 3.39. (Analytic Fredholm theory) Let Ω ⊂ C be a connected domain and T (·) : Ω → C(V ) an
analytic operator-valued map taking value in the compact operators for all z ∈ Ω. Then we have one of the
two following alternatives

1. 1 is in the spectrum of T for all z ∈ Ω.

2. 1 ∈ ρ(T (z)) for all z ∈ Ω \ Σ where Σ is a discrete set without accumulation point. The resolvent
(T (z) − 1)−1 is meromorph in Ω, analytic in Ω \ Σ. The residue by the poles are finite rank operators
and for z ∈ Σ, T (z)ξ = ξ has a nontrivial solution space of finite dimension.

The proof of this theorem as well as the next few ones uses the following idea and a basic construction
which we now explain. The idea is to reduce the solvability of T (z)ξ = ξ to the solvability of f (z) = 0 for
some analytic function. Now we have either f ≡ 0 or f vanish on a discrete set. In order to do this we will
write f (z) has the determinant of some matrix obtained from a finite rank operator S(z) and the solvability
of T (z)ξ = ξ is equivalent to the solvability of T (z)ξ = ξ.
The basic construction is as follows: Given z0 ∈ Ω and z ∈ Ω0 ≡ {z : kT (z) − T (z0 )k < 1/2} we pick
Tb an operator of finite rank such that
kT (z0 ) − Tbk < 1/2 (3.113)
Then
kT (z) − Tbk < 1 , (3.114)
and
A(z) = (1 − (T (z) − Tb)) (3.115)
is invertible for z ∈ Ω0 . We now consider the finite rank operator

S(z) = TbA(z)−1 (3.116)

and note that

T (z) − 1 = (S(z) − 1)A(z) (3.117)

Proof of Theorem 3.39. With S(z) as in (3.117) we note that

T (z) − 1 invertible ⇐⇒ S(z) − 1 invertible. (3.118)

and
T (z)ξ = ξ has nontrivial solutions ⇐⇒ S(z)ξ = ξ has nontrivial solutions (3.119)
Note that the range of S(z) = TbA(z)−1 is always contained in H0 = R(Tb) which is independent of z and
dim (H0 ) = n < ∞. We decompose

H = H0 ⊕ H1 with H1 = H0⊥ (3.120)

3.3. SPECTRAL THEORY OF COMPACT OPERATORS 59

We have a unique decomposition ξ = ξ0 + ξ1 for any ξ ∈ H and if we write

 
ξ0
ξ= (3.121)
ξ1

and we have the matrix-like representation

 
A00 A01
A= , Aij ∈ L(Hj , Hi ) . (3.122)
A10 A11

Since R(S(z)) = H0 we have (with Sij ≡ Sij (z))

 
S00 S01
S(z) = (3.123)
0 0

and so     
S00 S01 ξ0 (S00 − 1)ξ0 + S01 ξ1
(S(z) − 1)ξ = = (3.124)
0 0 ξ1 −ξ1
Therefore (S(z) − 1)ξ = 0 implies ξ1 = 0 and (S00 − 1)ξ0 = 0 and thus

dim {ξ : S(z)ξ = ξ} = dim {ξ0 : S00 (z)ξ0 = ξ0 } = n < ∞ (3.125)

In particular S(z)ξ = ξ has a nontrivial solution if and only if

f (z) ≡ det(S00 (z) − 1) = 0 . (3.126)

Since S00 (z) can be represented as a n × n matrix and is analytic then f (z) is Ω0 and

Σ0 = {z ∈ Ω0 ; f (z) = 0} (3.127)

is either Ω0 or is a discrete set.

In conclusion if f (z) ≡ 0 in Ω0 then S(z)ξ = ξ has nontrivial solution for all z ∈ Ω0 and so does
T (z)ξ = ξ and 1 belong to the spectrum of T (z) for all z ∈ Ω0 .
If f (z) 6≡ 0 then Σ0 = {z ∈ Ω0 ; f (z) = 0} is a discrete set without accumulation points and if z ∈
/ Σ0
we have

(T (z) − 1)−1 = A(z)−1 (S(z) − 1)−1

(S00 − 1)−1 (S00 − 1)−1 S01
 
= A(z)−1 .
0 −1
(3.128)

Since (S00 − 1)−1 = f (z)−1 C for some analytic matrix C, (S00 − 1)−1 is meromorph in Ω0 with poles on
the zeros of f (z) and residuum of the form
 
C CS01
Res(f (z)−1 ) (3.129)
0 0

which is of finite rank. The same holds for T (z).

60 CHAPTER 3. SPECTRAL THEORY

Lemma 3.40. Let T ∈ L(H) be compact. Then for µ 6= 0 the range R(T − µ1) is closed.
Proof. Replacing T by z −1 T we need to show that R(T − 1) is closed. By (3.117) we have

(T − 1) = (S − 1)A (3.130)

and so R(T − 1) = R(S − 1). We also have

(S00 − 1)−1 (S00 − 1)−1 S01

 
−1
(S(z) − 1) = (3.131)
0 −1

Let {ηn } be a Cauchy sequence in R(S − 1), i.e.

      
ηk0 S00 − 1 S01 ξk0 (S00 − 1)ξk0 + +S01 ξk1
= = (3.132)
ηk1 0 −1 ξk1 −ξk1

Therefore ξk1 is a Cauchy sequence and limk ξk1 = ξ1 and limk S01 ξk1 = S01 ξ1 . Therefore (S00 − 1)ξk0
is a Cauchy sequence. Since dim (H0 ) < ∞ then the range of (S00 − 1) is closed.

Lemma 3.41. Under the same assumptions as in Theorem 3.39 for z ∈ Ω we have

dim N (T (z) − 1) = dim N (T ∗ (z) − 1) . (3.133)

Proof. Since T (z) − 1 = (S(z) − 1)A(z) we have N (T (z) − 1) = A(z)−1 N (S(z) − 1). Also we have

(T (z) − 1)∗ = A(z)∗ (S(z) − 1)∗ (3.134)

we have
dim N (T (z)∗ − 1) = dim N (S ∗ (z) − 1) . (3.135)
So it is enough to prove the theorem for S instead of T . Since
 
S00 S01
S= (3.136)
0 0

we have N (S − 1) = N (S00 − 1). Since

 
S00 − 1 0
S∗ − 1 = ∗ (3.137)
S01 −1

we have
∗
   
∗ ξ0 (S00 − 1)ξ0
0 = (S − 1) = ∗ (3.138)
ξ1 S01 ξ0
if and only if
∗ ∗
ξ0 ∈ N (S00 − 1) and ξ1 = S01 ξ0 . (3.139)
Thus we have
∗
dim N (S(z) − 1) = dim N (S00 − 1) = dim N (S00 − 1) = dim N (S(z)∗ − 1) . (3.140)
3.3. SPECTRAL THEORY OF COMPACT OPERATORS 61

Lemma 3.42. For T ∈ L(H) the following are equivalent.

1. T is compact.

2. T ∗ T is compact.

3. T T ∗ is compact.

4. T ∗ is compact.

Proof. 1. If T is compact, then T ∗ ∈ L(H) and so T ∗ T and T T ∗ are compact.

2. If T ∗ T is compact then for any bounded sequence ξn there exists a subsequence ξk0 = ξnk such that
T T ξk0 converges. Then we have
∗

kT (ξk0 − ξl0 )k2 = (T (ξk0 − ξl0 ) , T (ξk0 − ξl0 )) = ((ξk0 − ξl0 ) , T ∗ T (ξk0 − ξl0 ))
= kξk0 − ξl0 kkT ∗ T (ξk0 − ξl0 )k ≤ 2M kT ∗ T (ξk0 − ξl0 )k . (3.141)

and so T ξk0 converges and so T is compact.

3. Finally we have T ∗ T compact ⇒ T compact ⇒ T T ∗ compact ⇒ T ∗ compact.

We are now in the position to prove

Theorem 3.43. (Riesz-Schauder) Let T ∈ L(H) compact. Then we have

1. Every spectral point µ ∈ σ(T ), µ 6= 0 is an eigenvalue of T .

2. The point spectrum σp (T ) is countable, µ = 0 is the only possible accumulation point of σp (T ).

Proof. Let us set T (z) = zT . Then is zT analytic in C with values C(H). The conditions of Theorem 3.39
are satisfied with Ω = C. The first case cannot occur since 1 would be a spectral value of zT for all z, i.e.
1/z would be a spectral value of T for all z ∈ C and thus σ(T ) = C which is not possible for a bounded
operator.
Therefore the set
Σ = {z ; 1is not a singular value of T (z)} (3.142)
is a discrete set without accumulation point. Furthermore we have

1
z ∈ Σ ⇔ zT − 1 not bijective ⇔ T − 1 not bijective (3.143)
z
So the set  
1
σ
b(T ) = µ ; µ = with z ∈ Σ (3.144)
z
is contained in σp (T ) (and is is equal to it if 0 ∈ σp (T )).

Theorem 3.44. (Fredholm alternative) Let T ∈ L(H) compact. Then we have for Tµ = T − µ1, µ 6= 0
of one the two options
62 CHAPTER 3. SPECTRAL THEORY

1. 
Tµ ξ = η
has a 1d dimensional solution space (3.145)
Tµ∗ ξ = η

Tµ ξ = 0
has only the trivial solution (3.146)
Tµ∗ ξ = 0

2.

Tµ ξ = 0 has a solution space Mµ (dim (Mµ ) < ∞)

Tµ∗ ξ = 0 has a solution space Mµ∗ (dim (Mµ∗ ) = dim (Mµ ))

Tµ ξ = η has a solution if and only if η ⊥ Mµ∗

Tµ∗ ξ = η has a solution if and only if η ⊥ Mµ

Proof. Let µ 6= 0. If µ 6∈ σp (T ) then 1. holds and ξ = Rµ (T )η is the unique solution of Tµ ξ = η. The rest
is obvious.
If µ ∈ σp (T ) then Mµ = N (Tµ ) is the eigenspace of T for the eigenvalue µ and so dim(Mµ ) < ∞
and dim(Mµ ) = dim(Mµ∗ ) by Lemma 3.41. It remains to show that Tµ ξ = η has a solution if and only if
η ∈ N (Tµ∗ )⊥ .
If Tµ ξ = η then for ζ ∈ N (Tµ∗ ) we have

(η, ζ) = (Tµ ξ, ζ) = (ξ, Tµ∗ ζ) = 0 (3.147)

and so η ⊥ N (Tµ∗ ).
Conversely we show that if T µξ = η has no solution then η ∈ / N (Tµ∗ )⊥ . By Lemma 3.40, R(Tµ ) ≡ H0
is closed. If Tµ ξ = η has no solution then there exists δ > 0 such that

inf kβ − ηk = inf kTµ ξ − ηk = δ (3.148)

β∈H0 ξ∈H

By Corollary chb3 to Hahn-Banach theorem the exists λ ∈ H 0 such that λ(η) = δ and λ(Tµ ξ) = 0 for all
ξ ∈ H. By Riesz representation theorem there exists ζ ∈ H such that

(η, ζ) = δ 0 = (Tµ ξ, ζ) = (ξ, Tµ∗ ζ) for all ξ ∈ H . (3.149)

So Tµ∗ ζ = 0 and so ζ ∈ N (Tµ∗ ). Since (η, ζ) = δ > 0 we have on one hand ζ 6= 0 and, on the other hand,
that η ∈ N (Tµ∗ )⊥ .

Before we pursue we we recall a few facts and definitions form Hilbert space theory.
• If T = T ∗ then the eigenvalues are of T are real and the eigenvectors for distinct eigenvalues are
orthogonal.
• T is called a positive operator if the quadratic form (T ξ, ξ) ≥ 0 (it is important to work in complex
vector spaces here!) and for positive operator every eigenvalue is nonnegative.
• If T = T ∗ from 2.56 we know that kT ∗ T k = kT k2 and so kT 2 k = kT k2 . By induction kT 2n k =
kT k2n and thus r(T ) = inf n kT n k1/n = kT k.
3.3. SPECTRAL THEORY OF COMPACT OPERATORS 63

• T is called normal if T ∗ T = T ∗ .

Theorem 3.45. (Hilbert-Schmidt) Let T ∈ L(H) be compact and self-adjoint. Then there exists an or-
thonormal basis {ηi } of H such that T ηi = λi ηi with limi λi = 0. If µ1 , µ2 , · · · are the pairwise distinct
eigenvalues of T and E1 , E2 , · · · the corresponding eigenspace of T . Then we have
X
H = E1 ⊕ E2 ⊕ · · · T = λPi (3.150)
i

where Pi is the orthogonal projection on Hi .

Proof. Eigenspaces for different eigenvalues are orthogonal, so we can choose in each eigenspace an or-
thonormal basis and this gives an orthonormal sequence {ηi }. Let H1 be the closed linear span of {etai }.
Clearly T maps H1 into itself. Let us write H = H1 ⊕ H1⊥ . and let us assume H1⊥ is not trivial. Since
T = T ∗ we have T maps H1⊥ into itself. Let T = T |H1 and T2 = T |H1 ⊥ . The operator T2 is self-adjoint
and compact.
From Riesz-Schauder theorem if µ ∈ σ(T2 ) with µ 6= 0 then µ ∈ σp (T2 ). But then µ is also an
eigenvalue of T which is impossible. Therefore σ(T2 ) = {0} = kT2 k and thus T2 = 0. So H1⊥ is
the eigenspace of T for the eigenvalue 0. But since all the eiegenspace for T are already in H1 we have
H1⊥ = 0.

We have previously developed a functional calculus to define f (T ) for general bounded operators and
were able to use functions f which are analytic in an open set containing σ(T ). For a compact self-adjoint
operator T we can define f (T ) for every function defined on the spectrum of A.
Theorem 3.46. (Functional Calculus) Let T be a compact self-adjoint operator and f a bounded complex-
valued function defined on σ(T ). To such an f we can define f (T ) such that

1. If f ≡ 1 then f (T ) = 1.

2. If f (x) = x then f (T ) = T .

3. The map f 7→ f (T ) is an isomorphism of the ring of bounded functions on σ(A) into the algebra
L(H).

4. The isomorphism is isometric:

kf (T )k = sup |f (µ)k (3.151)
µ∈σ(T )

5. If f is real-valued, f (T ) is symmetric.

6. If f is positive on σ(T ), then f (T ) is positive.

Proof. The proof is shorter than the statement. From Hlbert-Schmidt theorem weP can pick an o.n.b. basis
{ηk } of eigenvectors such that T ηk = µk ηk . If ξ ∈ H has the decompostion ξ = xi ηi then define f (T )
by
f (T )ξ = f (µk )xk ηk . (3.152)
The theorem is now obvious.

We also have
64 CHAPTER 3. SPECTRAL THEORY

Theorem 3.47. Suppose T1 , · · · Tn are bounded operators in L(H) such that

1. Tk is self-adjoint for k = 1, · · · , n.

2. T1 is compact.

3. Tk Tl = Tk Tl for all k, l.

Then there exists an orthonormal basis {ξi } of H such that ξi is an eigenvector for all Tk .

Proof. Let Si be the eigenspaces of T1 for the eigenvalue µi . Then by Hilbert-Schmidt theorem we have

H = S1 ⊕ S2 ⊕ · · · (3.153)

Then the subspace Si is invariant under Tk since if ξ ∈ Si m, then T ξ = µi ξ then

T1 Tk ξ = Tk T1 ξ = µi Tk ξ (3.154)

and thus Tk ξ ∈ Si . The restriction of T2 on Sn is finite-dimensional and thus we can choose a basis of Si
such that the basis elements are eigenvectors for T1 and T2 and we can repeat the arguments for T3 , etc....
Since the Sn are finite dimensional the procedure will stop after a finite number of steps.

Corollary 3.48. If T is a normal operator then there exists an o.n.b {ξi } of H such that
∞
X
Tξ = µn (ξ, ξn )ξn (3.155)
n=1

Proof. We write
T + T∗ T − T∗
T = + . (3.156)
2 } | {z
| {z 2 }
≡R ≡J

Then T and J are compact, T is self-adjoint and iJ is self-adjoint.

Finally we consider non self-adjoint operators.

Theorem 3.49. Let Let T ∈ L(H) compact. Then there exists 2 sequences {ηi }N N
i=1 and {ζi }i=1 in H (with
N finite or ∞) and a sequence {νi } of non-negative numbers such that
N
X
T = νi (· , αi )βi (3.157)
i=1

Proof. If T is compact then T ∗ T is compact, self-adjoint, and positive. So there exists an orthonormal
sequence {ηi } such that
T ∗ T ηi = µi ηi (3.158)
with µi ≥ 0 and limi µi = 0. Moreover T ∗ T restricted on the orthogonal complement to the span of the ηi
is equal to 0. Now let
√
νi = µi (3.159)
3.4. APPLICATIONS 65

and set
1
ζi = T ηi . (3.160)
νi
We have then
1 1 1
(ζi , ζj ) = (T ηi , T ηj ) = (ηi , T ∗ T ηj ) = µj (ηi , ηj ) = δij (3.161)
νi νj νi νj νi νj

and so the βi are orthonormal. For any ξ ∈ H we have

N
X
ξ = (ξ, ηi )ηi + ξ ⊥ . (3.162)
i=1

where T ∗ T ξ ⊥ = 0. This implies that

0 = (ξ ⊥ , T ∗ T ξ ⊥ ) = (T ξ ⊥ , T ξ ⊥ ) = kT ξ ⊥ k2 (3.163)

and so T ξ ⊥ = 0. So we have
N
X N
X
Tξ = (ξ, ηi )T ηi = νi (ξ, ηi )ζi , (3.164)
i=1 i=1

as claimed.

3.4 Applications
3.5 Exercises
2
Exercise 21. 1. Show that the Fourier transform T maps the space of function of the form p(x)e−πx
where P is a polynomial of degree ≤ n into itself.

2. Find the eigenfunctions of T .

2
Hint: With our conventions the Fourier transform of e−πx is itself.
Exercise 22. Prove Lemma 3.18
Exercise 23. Show that if T is a normal operator (i.e., T ∗ T = T T ∗ ) then r(T ) = kT k.
Exercise 24. For T ∈ L(V ) let us define τ (T ) by

τ (T ) = max Re (µ) (3.165)

µ∈σ(T )

i.e. the spectrum of T is contained in the half-plane {Re (z) ≤ τ (T )}. Show that we have

ln kenT k
τ (T ) = lim . (3.166)
n→∞ n
66 CHAPTER 3. SPECTRAL THEORY

Exercise 25. Suppose T is bounded operator on a Hilbert space H such that T ξk = λk ξk with ξk an
orthonormal basis of H. Show that T is compact if and only if λk → 0.
Rt
Exercise 26. Show that the operator T f (t) = 0 f (s) ds on C[0, 1] is compact and has no eigenvalues.
Give an example of an operator T on lp which is compact and has no eigenvalue.
Exercise 27. Consider the integral equation
Z 1
f (t) − µ k(t, s)f (s) ds = g(t) (3.167)
0
R1
with a continuous k(t, s) and denote T f (t) = 0
k(t, s)f (s) ds.
1
1. Let r(T ) be the spectral radius of T on C[0, 1]. Show that if |µ| < r(T ) then (3.167) has a unique
solution which can be written in the form
Z 1
f (t) = g(t) + r(t, s)g(s) ds (3.168)
0

(the kernel (r(t, s) is called the resolvent kernel). Hint: Neumann series.
2. Compute the resolvent kernel and the solution for the integral equation

1 1 t−s
Z
f (t) = e f (s) ds + g(s) . (3.169)
2 0
Pn
Exercise 28. Consider a kernel k(t, s) = j=1 aj (t)b)j(s). Without loss of generality we may assume
that the aj and bj are linearly independent. Show that if the equation
Z 1
f (t) − µ k(t, s)f (s) ds = g(t) (3.170)
0

has a solution then it must be of the form

n
X Z 1
f (t) = g(t) + µ cj aj (t), , cj = bj (s)xj (s) ds (3.171)
j=1 0

and the constants cj must satisfy the linear equations

n
X
cj − µ ajk ck = yi , (3.172)
j=1

with Z 1 Z 1
ajk = bj (s)ak (s) ds , yj = bj (s)y(s) ds (3.173)
0 0
Exercise 29. Consider the equation
Z 1
f (t) − µ (s + t)x(t)dt = g(s) (3.174)
0
3.5. EXERCISES 67

1. Use the previous exercise to solve this equation if µ2 + 12µ − 12 6= 0.

2. Find the eigenvalues and eigenfunctions.
Exercise 30. Let H be a Hilbert space and T a self-adjoint compact operator with eigenvalues µk with
eigenvectors ξk . Consider the equation
(λ1 − T )ξ = η (3.175)
for a given λ 6= 0 and η ∈ H given. Show that this equation has a solution if and only if η is orthogonal to
N (λ1 − T ) and that the set of all solution general solution is given by

ξ = ξ0 + ζ (3.176)

where ζ ∈ N (λ1 − T ) and

1 1 X µn
ξ0 = η+ (η , ξn )ξn (3.177)
λ λ µn − λ
µn 6=λ

Hint: Consider separately the case where λ is an eigenvalue or not. Do not forget the show that ξ0 is
well-defined.
68 CHAPTER 3. SPECTRAL THEORY
Bibliography

[1] Peter Lax: Functional Analysis, Wiley, ISBN: 978-0471556046

[2] Michael Reed and Barry Simon, Methods of Modern Mathematical Physics, Vol I: Functional Analysis,
Academic Press, ISBN: 978-0125850506
[3] Martin Schechter Principles of Functional Analysis, Graduate Studies in Mathematics, American
Mathematical Society, ISBN: 978-0821828953

69