Distributed-Parameter

Port-Hamiltonian
Systems

Hans Zwart
Department of Applied Mathematics
University of Twente
P.O. Box 217
7500 AE Enschede
The Netherlands

and

Birgit Jacob
Institut f r Mathematik
u
Universitt Paderborn
a
Warburger Strae 100
D-33098 Paderborn
Germany

Contents
1. Introduction
1.1. Motivating examples . . . . . . . . . .
1.2. Class of PDEs . . . . . . . . . . . . .
1.3. Dirac structures and port-Hamiltonian
1.4. Overview . . . . . . . . . . . . . . . .
1.5. Exercises . . . . . . . . . . . . . . . .

. . . . .
. . . . .
systems
. . . . .
. . . . .

2. Homogeneous dierential equation

2.1. Introduction . . . . . . . . . . . . . . . . . . . .
2.2. Semigroup and innitesimal generator . . . . .
2.3. Homogeneous solutions to the port-Hamiltonian
2.4. Technical lemmas . . . . . . . . . . . . . . . .
2.5. Properties of semigroups and their generators .
2.6. Exercises . . . . . . . . . . . . . . . . . . . . .
2.7. Notes and references . . . . . . . . . . . . . . .
3. Boundary Control Systems
3.1. Inhomogeneous dierential equations .
3.2. Boundary control systems . . . . . . .
3.3. Port-Hamiltonian systems as boundary
3.4. Outputs . . . . . . . . . . . . . . . . .
3.5. Some proofs . . . . . . . . . . . . . . .
3.6. Exercises . . . . . . . . . . . . . . . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

. . . . .
. . . . .
system
. . . . .
. . . . .
. . . . .
. . . . .

. . . . . . . . . .
. . . . . . . . . .
control systems
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

1
1
6
9
17
17

.
.
.
.
.
.
.

19
19
20
26
33
34
39
42

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

43
43
46
50
52
55
57

4. Transfer Functions
4.1. Basic denition and properties . . . . . . . . . .
4.2. Transfer functions for port-Hamiltonian systems
4.3. Exercises . . . . . . . . . . . . . . . . . . . . . .
4.4. Notes and references . . . . . . . . . . . . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

61
62
66
71
72

5. Well-posedness
5.1. Introduction . . . . . . . . . . . . . . . . . . .
5.2. Well-posedness for port-Hamiltonian systems
5.3. The operator P1 H is diagonal. . . . . . . . . .
5.4. Proof of Theorem 5.2.6. . . . . . . . . . . . .
5.5. Well-posedness of the vibrating string. . . . .
5.6. Technical lemmas . . . . . . . . . . . . . . .

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

73
73
75
79
84
86
89

.
.
.
.
.
.

.
.
.
.
.
.

Contents
5.7. Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.8. Notes and references . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
6. Stability and Stabilizability
6.1. Introduction . . . . . . . . . . . . . . . . . . . . . .
6.2. Exponential stability of port-Hamiltonian systems
6.3. Examples . . . . . . . . . . . . . . . . . . . . . . .
6.4. Exercises . . . . . . . . . . . . . . . . . . . . . . .
6.5. Notes and references . . . . . . . . . . . . . . . . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

93
93
95
102
103
103

7. Systems with Dissipation

7.1. Introduction . . . . . . . . .
7.2. General class of system with
7.3. General result . . . . . . . .
7.4. Exercises . . . . . . . . . .
7.5. Notes and references . . . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

105
105
106
114
116
116

A. Mathematical Background
A.1. Complex analysis . . . . . . . . . . . . . . . . . . . .
A.2. Normed linear spaces . . . . . . . . . . . . . . . . . .
A.2.1. General theory . . . . . . . . . . . . . . . . .
A.2.2. Hilbert spaces . . . . . . . . . . . . . . . . . .
A.3. Operators on normed linear spaces . . . . . . . . . .
A.3.1. General theory . . . . . . . . . . . . . . . . .
A.3.2. Operators on Hilbert spaces . . . . . . . . . .
A.4. Spectral theory . . . . . . . . . . . . . . . . . . . . .
A.4.1. General spectral theory . . . . . . . . . . . .
A.4.2. Spectral theory for compact normal operators
A.5. Integration and dierentiation theory . . . . . . . .
A.5.1. Integration theory . . . . . . . . . . . . . . .
A.5.2. Dierentiation theory . . . . . . . . . . . . .
A.6. Frequency-domain spaces . . . . . . . . . . . . . . .
A.6.1. Laplace and Fourier transforms . . . . . . . .
A.6.2. Frequency-domain spaces . . . . . . . . . . .
A.6.3. The Hardy spaces . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

117
117
122
123
127
132
132
149
159
159
166
171
171
180
183
183
187
190

. . . . . . .
dissipation.
. . . . . . .
. . . . . . .
. . . . . . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

Bibliography

197

Index

198

ii

List of Figures
1.1. Transmission line . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2. Dirac structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.3. Composition of Dirac structures . . . . . . . . . . . . . . . . . . . . . . . . 15
2.1. Coupled vibrating strings . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.2. Coupled transmission lines . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.1. Coupled vibrating strings with external force . . . . . . . . . . . . . . . . 59
4.1. Nyquist plot of (4.44) for R = 10 . . . . . . . . . . . . . . . . . . . . . . . 70
5.1. The system (5.40) with input (5.43) and output (5.44) . . . . . . . . . . . 83
5.2. The wave equation with input and output (5.76) and (5.77) . . . . . . . . 88
5.3. The closed loop system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
7.1. Interconnection structure. . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
A.1. The relationship between T and T . . . . . . . . . . . . . . . . . . . . . 153

iii

List of Figures

iv

Chapter 1

Introduction

In the rst part of this chapter we introduce some motivation examples and show that
these possesses a common structure. Finally, we indicate what we do in more detail in
the chapters to follow.

1.1. Motivating examples

In this section we shall by using simple examples introduce our class and indicate what
are natural (control) questions for these systems. We begin with the example of the
transmission line. This model describes the charge density and magnetic ux in a cable,
as is depictured below.
V (b)

V (a)
I(a)
a

I(b)

Figure 1.1.: Transmission line

Example 1.1.1 (Transmission line) Consider the transmission line on the spatial
interval [a, b]
(, t)
Q
(, t) =
t
L()
Q(, t)

(, t) =
.
t
C()

(1.1)

Here Q(, t) is the charge at position [a, b] and time t > 0, and (, t) is the ux
at position and time t. C is the (distributed) capacity and L is the (distributed)
inductance.

1. Introduction
The voltage and current are given by V = Q/C and I = /L, respectively. The energy
of this system is given by
E(t) =

1
2

b
a

(, t)2
Q(, t)2
+
d.
L()
C()

(1.2)

If and Q satisfy the dierential equation (1.1), then we nd the following

d
E(t) =
dt

b
a

(, t)
Q(, t) Q
(, t) +
(, t)d
L() t
C() t

=
a

(, t) Q(, t) Q(, t) (, t)

d
L() C()
C() L()

(, t) Q(, t)
d
L() C()
a
(a, t) Q(a, t) (b, t) Q(b, t)

.
=
L(a) C(a)
L(b) C(b)
=

(1.3)

We see that the change of energy can only occur via the boundary. We can also write
the expression of (1.3) using voltage and current, and we nd that
d
E(t) = V (a, t)I(a, t) V (b, t)I(b, t).
dt

(1.4)

Since voltage times current equals power and the change of energy is also power, this
equation represents a power balance. We interpret this equality by saying that the
power of the systems equals the power ow at its boundary. Hence by choosing proper
boundary conditions we may ensure that the energy stays within the system.
A natural control question would be how to stabilize this system. The power balance
(1.4) is very useful for solving this question. Suppose that the voltage at = a is set
to zero, and at the other end we put a resistor, i.e., V (b, t) = RI(b, t). We see from the
power balance (1.4) that
d
E(t) = RI(b, t)2 .
dt
This implies that the energy decays. However, will the energy converge to zero, and if
so how fast? These are stability/stabilizability question which we study in Chapter 6.
Since the power ows via the boundary, it is natural to control via the boundary. In
fact we did this already in the previous paragraph when we put V (b, t) = RI(b, t). Hence
we come up with the question which and how many of the four variables, voltage and
current at the boundary, we may choose as an input. It seems from (1.4) that we may
take all four of them as (independent) inputs. As we shall see in Chapter 3, we may
only choose at most two as inputs. A similar question can be asked for outputs. Since
an output is dictated by the system, one may wonder if all four boundary conditions are
dictated by the system, see Chapter 5 for the answer.
If the system dictates the output, there should be a unique solution for a given initial
condition. In our partial dierential equation (1.1), we have not given an initial condition, i.e., (, 0) and Q(, 0). As one may expect, we choose these initial conditions

1.1. Motivating examples

in the energy space, meaning that the initial energy, E(0), is nite. Giving only an
initial condition is not sucient for a partial dierential equation (p.d.e.) to have a
(unique) solution, one also has to impose boundary conditions. In Chapter 2 we answer
the technical question for which boundary conditions the p.d.e. possesses a (unique)
solution.
The previous example is standard for the class of systems we are studying. There
is an energy function (Hamiltonian), a power balance giving that the change of energy
(power) goes via the boundary of the spatial domain. The example of the (undamped)
vibrating string is very similar.
Example 1.1.2 (Wave equation) Consider a vibrating string of length L = b a,
held stationary at both ends and free to vibrate transversely subject to the restoring
forces due to tension in the string. The vibrations on the system can be modeled by
2w
2w
(, t) = c 2 (, t),
t2

c=

T
, t 0,

(1.5)

where [a, b] is the spatial variable, w(, t) is the vertical position of the string, T
is the Youngs modulus of the string, and is the mass density, which are assumed
to be constant along the string. This model is a simplied version of other systems
where vibrations occur, as in the case of large structures, and it is also used in acoustics.
Although the wave equation is normally presented in the form (1.5), it is not the form
we will be using. However, more importantly, it is not the right model when Youngs
modulus or the mass density are depending on the spatial coordinate. When the later
happens, the correct model is given by
1
w
2w
(, t) .
(, t) =
T ()
2
t
()

(1.6)

It is easy to see that this equals (1.5) when the physical parameter are not spatially
dependent. This system has the energy/Hamiltonian
E(t) =

1
2

()
a

w
(, t)
t

+ T ()

w
(, t)

d.

(1.7)

As we did in the previous example we can calculate the change of energy, i.e., power.
This gives (see also Exercise 1.1)
w
w
w
w
d
E(t) =
(b, t)T (b)
(b, t)
(a, t)T (a)
(a, t).
dt
t

(1.8)

Again we see that the change of energy goes via the boundary of the spatial domain.
One may notice that the position is not the (actual) variable used in the energy and the
power expression. The variables are velocity ( w ) and strain ( w ).
t

For this model one can pose similar questions as for the model from Example 1.1.1.
In particular, a control problem could be to damp out the vibrations on the string.
One approach to do this is to add damping along the spatial domain. This can also be
done by interacting with the forces and velocities at the end of the string, i.e., at the
boundary.

1. Introduction
Example 1.1.3 (Beam equations) In recent years the boundary control of exible
structures has attracted much attention with the increase of high technology applications
such as space science and robotics. In these applications the control of vibrations is
crucial. These vibrations can be modeled by beam equations. For instance, the EulerBernoulli beam equation models the transversal vibration of an elastic beam if the crosssection dimension of the beam is negligible in comparison with its length. If the crosssection dimension is not negligible, then it is necessary to consider the eect of the rotary
inertia. In that case, the transversal vibration is better described by the Rayleigh beam
equation. An improvement over these models is given by the Timoshenko beam, since it
incorporates shear and rotational inertia eects, which makes it a more precise model.
These equations are given, respectively, by
Euler-Bernoulli beam:
2
2w
2w
() 2 (, t) + 2 EI() 2 (, t) = 0, (a, b), t 0,
t

where w(, t) is the transverse displacement of the beam, () is the mass per unit
length, E() is the Youngs modulus of the beam, and I() is the area moment of
inertia of the beams cross section.
Rayleigh beam:
()

2w
2
(, t) + 2
2

2
2w
(, t) I () 2
t2
t

EI()

2w
(, t)
z 2

= 0,

where (a, b), t 0, w(, t) is the transverse displacement of the beam, ()

is the mass per unit length, I is the rotary moment of inertia of a cross section,
E() is the Youngs modulus of the beam, and I() is the area moment of inertia.
Timoshenko beam:
2w

() 2 (, t) =
K()
t

2
I () 2 (, t) =
t

w
(, t) (, t)

EI() (, t) + K()

(a, b), t 0,

w
(, t) (, t) ,

(1.9)

where w(, t) is the transverse displacement of the beam and (, t) is the rotation
angle of a lament of the beam. The coecients (), I (), E(), I(), and K()
are the mass per unit length, the rotary moment of inertia of a cross section,
Youngs modulus of elasticity, the moment of inertia of a cross section, and the
shear modulus respectively.
For the last model we show that it has similar properties as found in the previous
examples. The energy/Hamiltonian for this system is given by
E(t) =

1
2

K()
a

w
(, t) (, t)

E()I()

(, t)

+ I

+ ()

w
(, t)
t

(, t)
t

d.

(1.10)

1.1. Motivating examples

Next we want to calculate the power. For this it is better to introduce some (physical)
notation rst.
w
(, t) (, t)

w
(, t)
x2 (, t) = ()
t

x3 (, t) =
(, t)

x4 (, t) = I () (, t)
t
x1 (, t) =

shear displacement
momentum
angular displacement
angular momentum

Using this notation and the model (1.9), we nd that the power equals (Exercise 1.2)
x2 (, t)
x4 (, t)
dE
(t) = K()x1 (, t)
+ E()I()x3 (, t)
dt
()
I ()

(1.11)

Again we see that the power goes via the boundary of the spatial domain.
In the previous three examples we see that by imposing the right-boundary conditions
no energy will be lost. In other words, these system cannot loose energy internally.
However, there are many systems in which there is (internal) loss of energy. This may
be caused by internal friction by internal friction, as is the case in the following example.
Example 1.1.4 (Damped wave equation) Consider the one-dimensional wave equation of Example 1.1.2. One cause of damping is known as structural damping. Structural damping arises from internal friction in a material converting vibrational energy
into heat. In this case the vibrating string is modeled by
1
ks 2 w
w
2w
(, t) =
(, t) +
(, t) ,
T ()
t2
()

() 2 t
where ks is a positive constant.
To see that the energy decays, we calculate the power, i.e.,
given by (1.7).
d
E(t) =
dt

[a, b], t 0, (1.12)

dE
dt ,

where the energy is

w
w
w
w
(b, t)T (b)
(b, t)
(a, t)T (a)
(a, t) +
t

w
2w
w
2w
(b, t)ks
(b, t)
(a, t)ks
(a, t) ks
t
t
t
t

(1.13)
b
a

2w
(, t)
t

d.

From this equality, we see that if there is no energy ow through the boundary, the
energy will still decay.
Although this system looses energy internally, questions like is the system decaying to
zero, when no force is applied at the boundary are still valid for this model as well.
The most standard example of a model with diusion, is the model of heat distribution.

1. Introduction
Example 1.1.5 (Heat conduction) The model of heat conduction consists of only
one conservation law, that is the conservation of energy. It is given by the following
conservation law:

u
= JQ ,
(1.14)
t

where u(, t) is the energy density and JQ (, t) is the heat ux. This conservation law is
completed by two closure equations. The rst one expresses the calorimetric properties
of the material:
u
= cV (T ),
(1.15)
T
where T (, t) is the temperature distribution and cV is the heat capacity. The second
closure equation denes heat conduction property of the material (Fouriers conduction
law):
T
,
(1.16)
JQ = (T, )

where (T, ) denotes the heat conduction coecient. Assuming that the variations of
the temperature are not too large, one may assume that the heat capacity and the heat
conduction coecient are independent of the temperature, one obtains the following
partial dierential equation:
1
T
=
t
cV
If we look at the (positive) quantity E(t) =
that
T
dE
(t) = T (, t)()
(, t)
dt

()
1
2

(1.17)

b
2
a cV T (, t) d,

b
a

()
a

then it is not hard to see

T
(, t)

d.

Hence even when there is no heat ow through the boundary of the spatial domain, the
quantity E(t) will decrease. It will decrease as long as the heat ux is non-zero.
This later two examples are in nature completely dierent to the rst examples. In the
next section we show that the rst three examples have a common format. We return
to the example of the structural damped wave and the heat conduction only in Chapter
7 of these notes.

1.2. Class of PDEs

In this section we revisit the rst examples of the previous section and show that they
all lie in the same class of systems.
If we introduce the variable x1 = Q and x2 = in the rst example, see equation
(1.1), then the p.d.e. can be written as

x1 (, t)
x2 (, t)

0 1
1 0

1
C()

0
1
L()

x1 (, t)
x2 (, t)

(1.18)

1.2. Class of PDEs

The Hamiltonian is written as, see (1.2)
E(t) =
=

1
2
1
2

b
a

x1 (, t)2 x2 (, t)2
+
d
C()
L()

x1 (, t) x2 (, t)

1
C()

x1 (, t)
x2 (, t)

1
L()

d.

(1.19)

For the wave equation we can write down a similar form. We dene x1 = w (moment
tum) and x2 = w (strain). The p.d.e. (1.6) can equivalently be written as

x1 (, t)
x2 (, t)

0 1
1 0

1
()

0
T ()

x1 (, t)
x2 (, t)

(1.20)

d.

(1.21)

The energy/Hamiltonian becomes in the new variables, see (1.7),

E(t) =

1
2

1
=
2

b
a

x1 (, t)2
+ T ()x2 (, t)2 d
()

x1 (, t) x2 (, t)

1
()

0
T ()

x1 (, t)
x2 (, t)

For the model of Timoshenko beam, we have already introduced our variables in Example
1.1.3. We write the model and the energy using these new variables. Calculating the
time derivative of the variables x1 , . . . , x4 , we nd by using (1.9)

x1 (, t)

x2 (, t)

=
t x3 (, t)

x4 (, t)

x2 (,t)
()

x4 (,t)
I ()

(K()x1 (, t))
x4 (,t)
I ()

(E()I()x3 (, t)) + K()x1 (, t)

(1.22)

We can write this in a form similar to those presented in (1.18) and (1.20). However, as
we shall see, we also need a constant term. Since this is a long expression, we will not
write down the coordinates and t.

x1
0 1
x2 1 0

=
t x3 0 0
x4
0 0

0 0
0 0

0 0
1 0

0
0
0
1
0
0
0
0

0
0

1
0

K
1
0
0
0 0
0
0

K
0
0
0
0

0
0
0 EI
0 0
1

0
x1
0 x2

+
0 x3
1
x4
I

0
x1
0 x2

.
0 x3
1
x4
I

0
1
0

0 EI
0 0

(1.23)

1. Introduction
Formulating the energy/Hamiltonian in the variables x1 , . . . , x4 is easier, see (1.10)
b

1
1
x2 (, t)2 + E()I()x3 (, t)2 +
x4 (, t)2 d
K()x1 (, t)2 +
()
I ()
a

T
K() 0
0
0
x1 (, t)
x1 (, t)
1
0
0 x2 (, t)
1 b x2 (, t) 0

()
d.

=

0 E()I()
0 x3 (, t)
2 a x3 (, t) 0
1
0
0
0
x4 (, t)
x4 (, t)
I ()

E(t) =

1
2

(1.24)

We see that in the new formulation these examples have a common structure. There is
only one spatial and one time derivative, and the relation between these two derivatives
is of the form

x
(, t) = P1
[H()x(, t)] + P0 [H()x(, t)] .
(1.25)
t

T
Furthermore, we have that P1 is symmetric, i.e., P1 = P1 , P0 is anti-symmetric, i.e.,
T = P . Furthermore, they are both independent of . Finally, H is a (strictly)
P0
0
positive symmetric multiplication operator, independent of t. The energy or Hamiltonian
can be expressed by using x and H. That is

E(t) =

1
2

b
a

x(, t)T H()x(, t)d.

(1.26)

As we have seen in the examples, the change of energy (power) of these systems was only
possible via the boundary of its spatial domain. In the following theorem we show that
this is a general property for any system which is of the form (1.25) with Hamiltonian
(1.26).
Theorem 1.2.1. Consider the partial dierential equation (1.25) in which P0 , P1 are
T
T
constant matrices satisfying P1 = P1 and P0 = P0 . Furthermore, H is independent
on t and is symmetric, i.e. for all s we have that H()T = H(). For the Hamiltonian/energy given by (1.26) the following balance equation holds for all solutions of
(1.25)
b
1
dE
(1.27)
(Hx)T (, t)P1 (Hx) (, t) .
(t) =
dt
2
a
Proof: By using the partial dierential equation, we nd that
1
dE
(t) =
dt
2
=

1
2
1
2

b
a

1
x
(, t)T H()x(, t)d +
t
2

P1
a
b
a

b
a

x(, t)T H()

(Hx) (, t) + P0 (Hx) (, t)

x(, t)T H(, t) P1

x
(, t)d
t

H()x(, t)d+

(Hx) (, t) + P0 (Lx) (, t) d.

1.3. Dirac structures and port-Hamiltonian systems

Using now the fact that P1 , H() are symmetric, and P0 is anti-symmetric, we write the
last expression as
1
2

b
a

(Hx) (, t)

P1 H()x(, t) + [H()x(, t)]T P1

(Hx) (, t) d+

1 b
[H()x(, t)]T P0 H()x(, t) + [H()x(, t)]T [P0 H()x(, t)] d
2 a
1 b
(Hx)T (, t)P1 (Hx) (, t) d
=
2 a
b
1
= (Hx)T (, t)P1 (Hx) (, t) .
2
a
Hence we have proved the theorem.
The balance equation will turn out to be very important, and will guide us in many
problems. An overview of the (control) problems which we study in the coming chapters
is given in Section 1.4. First we concentrate a little bit more on the class of systems
given by (1.25). We show that we have to see it as a combination of two structure.
One given by P1 and P0 , and the other given by H. This is the subject of the following
section, in which we also explain the name Port-Hamiltonian.

1.3. Dirac structures and port-Hamiltonian systems

In this section we show that we can identify a deeper underlying structure to the p.d.e.
(1.25) and the balance equation (1.27). Therefore we look once more at the rst displayed
equation in the proof of Theorem 1.2.1. For a = and b = this equation becomes
1
dE
(t) =
dt
2

1
x
(, t)T H()x(, t)d +
t
2

x(, t)T H()

x
(, t)d.
t

(1.28)

In the expression on the right-hand side we have x and Hx. These are the same variables
t
used to describe the p.d.e. (1.25). Let us rename these variables f = x and e = Hx.
t
Furthermore, we forget the time, i.e., we see e and f only as functions of the spatial
variable. By doing so the p.d.e. becomes
f () = P1

e
() + P0 e()

(1.29)

and the right-hand side of (1.28) becomes

1
2

f ()T e()d +

1
2

e()T f ()d.

(1.30)

1. Introduction
Using the equation (1.29), we can rewrite the integrals in (1.30).
1
2

f ()T e()d +

1
2
1
2
1
=
2
1
=
2
1
=
2
=

1
2

e()T f ()d

T
e
() + P0 e() e()d+

e
e()T P1 () + P0 e() d

e
e T
() P1 e() + e()T P1 ()d

P1

e()T P0 e() + e()T P0 e()d

e()T P1 e() d,

T
T
where we have used that P1 = P1 and P0 = P0 . We remark that the above derivation
is exactly the same as the one in the proof of Theorem 1.2.1. Now under the mild
assumption that e() is zero in plus and minus innity, we conclude that

1
2

f ()T e()d +

1
2

e()T f ()d = 0.

(1.31)

for all e and f satisfying (1.29).

Based on (1.28) we call the expressions f T ed, eT f d, the power. Remember that
the change of energy is by denition the power. Hence we see from (1.31) that the power
is zero. This was already clear from the proof of Theorem 1.2.1, but we received it now
for any pair of variables which satises (1.29).
To illustrate this observation we consider three systems which have the same P1 and
P0 , but are totally dierent in their time behavior.
Example 1.3.1 We consider the following simple p.d.e.
x
x
(, t) =
(, t),
t

R.

(1.32)

We regard this equation as the equation (1.29) with P1 = 1 and P0 = 0 and f = x ,

t
1
e = x. Hence if we dene the energy E(t) = 2 x(, t)2 d and we assume that

x(, t) = 0, then we know that E(t) is zero along solutions.

Example 1.3.2 Consider the following p.d.e.
x
x
(, t) = x(, t) (, t),
t

R.

(1.33)

This equation is known as the inviscid Burgerss equation. If we dene f = x and

t
1
e = 2 x2 , then we see that this equation equals (1.29) for the same P1 and P0 as in the

10

1.3. Dirac structures and port-Hamiltonian systems

previous example. Furthermore, we know that

0=2

f (, t)e(, t)d =

x
(, t)x(, t)2 d
t

provided x(, t) = x(, t) = 0 for all t. The later integral can see as the time
1
derivative of H(t) = 3 x(, t)3 d. Hence we nd that this H is a conserved quantity,
i.e.,

dH
x
f (, t)e(, t)d = 0.
(1.34)
(t) =
(, t)x(, t)2 d =
dt
t

The above holds for every x satisfying (1.33) with x(, t) = x(, t) = 0. Since
the p.d.e. (1.33) is non-linear, proving existence of solutions is much harder than in
the previous example. However, as in the linear example we have found a conserved
quantity.
In the previous example we have chosen a dierent e, but the same f . We can also
choose a dierent f .
Example 1.3.3 Consider the discrete-time implicit equation
x(, n + 1) x(, n) =

[x(, n + 1) + x(, n)] , R, n Z.

(1.35)

In this equation, we choose f (, n) = x(, n+1)x(, n) and e(, n) = x(, n+1)+x(, n).
For this choice, we see that (1.35) is the same as (1.29) with P1 = 1 and P0 = 0.

If we choose the energy to be same as in Example 1.3.1, i.e., E(n) = x(, n)2 d,
then we nd that
E(n + 1) E(n) =

x(, n + 1)2 x(, n)2 d =

f (, n)e(, n)d = 0, (1.36)

provided x(, n) = 0 for all n Z. So for the implicit dierence equation (1.35) we
have once more a conserved quantity without knowing the solutions, or even knowing
existence.
As is become clear in the previous examples we may distinguish between an underlying
structure and the actual system. This underlying structure is named a Dirac structure
and is dened next.
Denition 1.3.4. Let E and F be two Hilbert spaces with inner product , E and
, F , respectively. Assume moreover that they are isometrically isomorphic, that is
there exists a linear mapping rF ,E : F E such that
rF ,E f1 , rF ,E f2

= f1 , f2

(1.37)

for all f1 , f2 F. The bond space B is dened as F E. On B we dene the following

symmetric pairing
f1
e1

f2
e2

= f1 , rE,F e2

+ e1 , rF ,E f2 E ,

(1.38)

11

1. Introduction
1
where rE,F = rF ,E .
Let V be a linear subspace of B, then the orthogonal subspace with respect to the
symmetric pairing (1.38) is dened as

V = {b B | b, v

= 0 for all v V}.

(1.39)

A Dirac structure is a linear subspace of the bond space D satisfying

D = D.

(1.40)

The variables e and f are called the eort and ow , respectively, and their spaces E
and F are called the eort and ow space. The bilinear product f, rE,F e F is called the
power or power product. Note that f, rE,F e F = rF ,E f, e E .
Dirac structures are depictured in Figure 1.2. Finally, we mention that by (1.39), we

Figure 1.2.: Dirac structure

have that for any element of the Dirac structure
2 f, rE,F e

= b, b

= 0.

(1.41)

This we interpret by saying that for any element of a Dirac structure the power is zero.
A Dirac structure can been seen as the largest subspace which this holds i.e., if V is a
subspace of B satisfying (1.41), then V is a Dirac structure if there does not exists a
subspace W such that V W, V = W, and the power of every element in W is zero.
Next we identify the Dirac structure associated to Examples 1.3.11.3.3.
Example 1.3.5 Choose the eort and ow space as L2 (, ), and let rF ,E = I.
Dene the following subspace of B = F E
D=

f
e

de
L2 (, ),
d
de
e() = e() = 0, and f =
.
d

B | e is absolutely continuous and

We claim that this subspace is a Dirac structure. Let b = ( f ) B and let b, d

e
d
for all d = fd D. Using our power product this implies that
e
0 = f, ed + fd , e = f, ed +

12

ded
,e .
d

(1.42)

=0

1.3. Dirac structures and port-Hamiltonian systems

In other words

ded
, e = ed , f .
d

(1.43)

This is equivalent to saying that e lies in the domain of the dual of the dierential
d
operator, d . Similar to Example A.3.64 and A.3.66 we conclude from this equation
that e is absolutely continuous and e() = e() = 0. Furthermore, by integration by
parts we see that
ded
,e =
d

ded
()e()d =
d

ed ()

de
de
()d = ed ,
.
d
d

Combining this with (1.43), we conclude that

ed ,

de
= ed , f .
d

Since this holds for a dense set of ed , we conclude that f =

( f ) D, and so D is a Dirac structure.
e

de
d .

Concluding, we see that

Now we have formally dened what is a Dirac structure, we can dene our class of
systems.
Let H be a real-valued function of x, and let D be a Dirac structure. We furthermore
assume that we have a mapping H from F to R which is Frchet dierentiable for any
e
x, i.e., see Denition A.5.25.
H(x + x) H(x) = (dH(x)) (x) + R(x, x),

(1.44)

R(x,x)
x

converges to zero when x 0. Since H takes values in R, we can by

Riesz representation theorem write the term (dH(x)) (x) as x, f F for some f . Note

that f still depends on x. Since F is isomorphically isomorf to E, we can nd an e such

F = x, rE,F e F . We denote this e by H (x). Combining this notation

that x, f

with

with equation (1.44), we nd

H(x + x) H(x) = x, rE,F

H
(x)
x

+ o().

(1.45)

The system associated with H and D is dened as

{x(, t) |

x
t (, t)
H
x (, t)

D for all t}.

(1.46)

Before we write our examples in this format, we show that the system dened above has
H as a conserved quantity along its trajectory.
H(x(t + )) H(x(t)) = H (x(t) + x(t) + r(x(t), )) H(x(t)),

13

1. Introduction
where r(x(t), ) 0 when 0. Since is small, we may ignore this term. Doing
so, and using (1.45) we nd
H
H(x(t + )) H(x(t))
= x(t), rE,F

(x(t))

+ o().

By the denition of our system, we have that the power product is zero, see (1.41) and
so, we nd that
dH(x(t))
= 0.
dt
Note that we normally write H(t) instead of H(x(t)). Thus along trajectories, H is a
conserved quantity. This conserved quantity is called the Hamiltonian. Equation (1.46)
clearly indicates that the system is dened by two objects. Namely, the Dirac structure
and the Hamiltonian.
We illustrate the above by looking once more to the Examples 1.3.1 and 1.3.2.
Example 1.3.6 From Example 1.3.5 we know which Dirac structure lies under the
p.d.e.s of Examples 1.3.1 and 1.3.2. Hence it remains to identify the Hamiltonians.
1
1
For Example 1.3.1 we easily see that H = 2 x2 d. If we dene H = 2 x2 , then it
by combining (1.46) with (1.42) gives the p.d.e. (1.32).
1
1
For Example 1.3.2 we nd H = 3 x3 d and H = 3 x3 .
In contrast to our examples in Section 1.1, the examples 1.3.11.3.3 do not have a
boundary. However, as seen in e.g. Example 1.1.1 the boundary is very useful for control
purposes. So we have to re-think and adjust the theory as developed until now. To
illustrate this, we consider Example 1.3.3 on a compact spatial interval.
Example 1.3.7 Consider the p.d.e.
x
x
(, t) =
(, t),
t

(a, b).

As in Theorem 1.2.1,we take the energy equal to E(t) =

b

dE
(t) =
dt

1
2

b
2
a x(, t) d.

(1.47)
We nd that

x
1
(, t)x(, t)d =
x(b, t)2 x(a, t)2 .
t
2

(1.48)

x
1
(, t)x(, t)d
x(b, t)2 x(a, t)2 = 0.
t
2

(1.49)

Or equivalently,
dE
(t) =
dt

b
a

As before, we want to see the middle expression as a power product. Hence if we

introduce f = x and e = x, then the integral induces a power product between f and
t
e. However, we still have the boundary terms. We introduce the boundary port variables
1
1
f = 2 [x(b) x(a)], and e = 2 [x(b) + x(a)]. With these variables, we see that we
may write (1.49) as
b
dE
f ()e()d f e = 0.
(1.50)
=
dt
a

14

1.3. Dirac structures and port-Hamiltonian systems

We can see this as a system like (1.46) by dening
E = F = L2 (a, b) R
and
rF ,E =

I 0
0 1

The Dirac structure is given by

f
f
de
L2 (a, b),
(1.51)
D = { B | e is absolutely continuous and
e
d
e
1
1
de
, f = [e(b) e(a)] and e = [e(b) + e(a)]}.
f=
d
2
2
Again we have a Dirac structure, but now it contains boundary variables. We can
formulate our system on this Dirac structure as
x
t

f
{x(, t) | H D}.

(1.52)

x
e

1
where D is given by (1.51) and H = 2 x2 .

From (1.52) we see that our system is dened via a Dirac structure, an Hamiltonian
and port variables. This motivated the name Port-Hamiltonian Systems.
One of the advantages of considering a system as a Dirac structure with a Hamiltonian
is that coupling of systems is now very easy. Suppose we have two Dirac structures, D1

and D2 as depictured in Figure 1.3. We couple the structures by f2 = f2 (the ow out

f2

f1
e1

D1

f2

e2

e2

f1
D2

e1

Figure 1.3.: Composition of Dirac structures

of the rst system equals the incoming ow
structure dened by

f1

f
D = 1 | there exist f2 , e2 s.t.

e1

e1

of the other system) and e2 = e2 . Then the

f1
f1

f2
D1 and f2
e1

e1

e2
e2

D2 .

(1.53)

15

1. Introduction
has zero power. To see this, we assume that for both systems f, e + e, f denotes the
power. We take as power for D

f1 , e1 + e1 , f1 + f1 , e1 + e1 , f1

We nd for this power product

f1 , e1 + e1 , f1 + f1 , e1 + e1 , f1

= f1 , e1 + e1 , f1 + f2 , e2 + e2 , f2

f2 , e2 e2 , f2 + f1 , e1 + e1 , f1

=

f1
f2

f1

f2
The last expressions are zero since

e1
e2

f1
f2
e1
e2

e1

e2

e1
e2

e1

e2

D1 and

f1
f2
e1

e2

f1
f2

,
,

f1
f2

+
.

D2 , respectively. Thus

we see that the total power of the interconnected Dirac structure is zero. Although this
is not sucient to show that D dened by (1.53) is a Dirac structure, it indicates the
promising direction. It only remains to show that D is maximal. For many coupled
Dirac structures this holds. If the systems has the Hamiltonian H1 and H2 respectively,
then the Hamiltonian of the coupled system is H1 + H2 . Of course we can extend this
to the coupling of more than two systems. We show a physical example next.
Example 1.3.8 (Suspension system) Consider a simplied version of a suspension
system described by two strings connected in parallel through a distributed spring. This
system can be modeled by
2u
2u
=c2 2 + (v u)
t2

2v
2v
=c2 2 + (u v) (, ), t 0,
t2

(1.54)

where c and are positive constants and u(, t) and v(, t) describe the displacement,
respectively, of both strings. The use of this model has potential applications in isolation of objects from outside disturbances. As an example in engineering, rubber and
rubber-like materials are used to absorb vibration or shield structures from vibration.
As an approximation, these materials can be modeled as a distributed spring. We show
that this system can be described as the interconnection of three subsystems, i.e., two
vibrating strings and one distributed spring. Seeing the system as an interconnection of
subsystems allows us to have some modularity in the modeling process, and because of
this modularity, the modeling process can be performed in an iterative manner, gradually
rening the model by adding other subsystems.

16

1.4. Overview

1.4. Overview
Now we know the class of systems we will be working with we can give more details on
the content of the coming chapters.
In Chapter 2 we study for which homogenous boundary conditions the p.d.e. (1.25)
possesses a unique solution which has non-increasing energy. This we do by applying
the general theory of innite-dimensional systems. If these boundary conditions are nonzero, then in Chapter 3 we show that this p.d.e. still has well-dened solutions. Hence
this enable us to apply a control (input) at the boundary to the system described by
(1.25). Under this same conditions, we show that boundary output is also possible. The
mapping between the input and the output can be described by the transfer function.
This is the subject of Chapter 4. Till Chapter 5 we have only considered systems which
are non-increasing in energy if no control is applied. Furthermore, the control has been
restricted to smooth functions. In Chapter 5, we extend our class of systems in both
directions. We show that many more boundary conditions are possible, and furthermore,
we show that if the homogenous system is well-posed, then the same hold for the system
when L2 -input are applied. Chapter 6 we can solve our rst control problem. We can
identify a large class of boundary feedback which stabilize the system. The stability is
exponential, meaning that the energy decays exponentially fast to zero. In 7 we can
treat a larger class of system. There we return to the examples 1.1.4 and 1.1.5. In this
section, we really need the underlying Dirac structure. Hence till Chapter 7, the Dirac
structure is underlying our system, and apart for using to dene boundary port, we shall
not use it very prominently. This changes in Chapter 7.

1.5. Exercises
1.1. In this exercise we check some integrals which appeared in our examples.
a) Check equation (1.8).
b) Check the equality (1.13).
1.2. Prove equation (1.11).
1.3. Show that equation (1.51) in Example 1.3.7 denes a Dirac structure.

17

1. Introduction

18

Chapter 2

Homogeneous dierential equation

2.1. Introduction
In this chapter we study the partial dierential equation (1.25). In particular, we characterize boundary conditions such that this p.d.e. has a unique solution, and such that
the energy decays along the solutions. In order to clarify the approach we take, let us
consider the most simplied version of (1.25)
x
x
(, t) = (, t),
t

[0, 1],

t 0,

(2.1)

where is a positive constant. As initial condition we take x0 ().

If x0 is continuously dierentiable, then it is easy to see that a solution is given by,
see Exercise 2.1.
x0 ( + t) [0, 1], + t < 1
(2.2)
x(, t) =
x0 (1)
[0, 1], + t > 1
However, this is not the only solution of (2.1). Another solution is given by
x(, t) =

x0 ( + t)
g( + t)

[0, 1], + t < 1

[0, 1], + t > 1

(2.3)

where g is an arbitrary continuous dierentiable function on (1, ) satisfying g(1) =

x0 (1).
Hence we have that the p.d.e. does not possess a unique solution. The reason for this
is that we did not impose a boundary condition. If we impose the boundary condition
x(1, t) = 0 for all t, then the unique solution is given by, see also (2.2),
x(, t) =

x0 ( + t)
0

[0, 1], + t < 1

[0, 1], + t > 1

(2.4)

Note that we assume that x0 (1) satises the boundary condition as well. The rule of
thumb is that for a rst order p.d.e. one need one boundary condition. However, this is

19

2. Homogeneous dierential equation

only a rule of thumb. If we impose to the p.d.e. (2.1) the boundary condition x(0, t) = 0,
then this p.d.e. does not possess any solution if x0 = 0, see Exercise 2.2.
We return to the p.d.e. (2.1) with boundary condition x(1, t) = 0. The solution is
given by (2.4). We see that for any initial condition, even the ones which are only
integrable, the function x(, t) of (2.4) is well-dened. This function depends on x0 . In
particular, we can dene mappings
x0 x(, t),

t 0.

(2.5)

This mapping has the following properties

It is linear, i.e., if x0 is written as c1 f0 + c2 g0 , then x can be written as c1 f + c2 g,
where f0 , g0 are mapped to f and g, respectively.
If we take x1 equal to x(, ), then the (composed) mapping
x0 x(, ) = x1 x(, t)

equals the mapping

x0 x(, t + ).
The reason for these properties lies in the linearity and time-invariance of the p.d.e.
Next we choose a (particular) class of initial conditions. It is easy to see that (2.1) is of
1
1 1
the form (1.25), and so the energy associated to it equals 2 0 |x()|2 d, or 1 0 |x()|2 d.
2
We take the class of functions with nite energy to be our class of initial conditions.
Thus we have dened for our simple example the space of initial conditions, and
we have seen some nice properties of the solution. In the sequel we omit the spatial
argument, when we write down an initial condition, or a solution, and we see (2.5) as a
mapping in the energy space.
In the following section we rst have to consider some abstract theory. The general
result obtained there enables us to show that for certain boundary conditions the p.d.e.
(1.25) possesses a unique solution.

2.2. Semigroup and innitesimal generator

In this section, we recap some abstract dierential theory. We denote by X an abstract
, X . For the simple
Hilbert space, with inner product , X and norm X =
p.d.e. considered in the previous section, we saw some nice properties of the mapping
from the initial condition to the solution at time t. These properties are formalized in
the following denition.
Denition 2.2.1. Let X be a Hilbert space. The operator valued function t T (t),
t 0 is a strongly continuous semigroup if the following holds
1. For all t 0, T (t) is a bounded linear operator on X, i.e., T (t) L(X);
2. T (0) = I;

20

2.2. Semigroup and innitesimal generator

3. T (t + ) = T (t)T ( ) for all t, 0.
4. For all x0 X, we have that T (t)x0 x0

converges to zero, when t 0.

We sometimes abbreviate strongly continuous semigroup to C0 -semigroup and most

times it will be denoted by (T (t))t0 .

We call X the state space, and its elements states. To obtain a feeling for these dening
properties assume that T (t) denotes the mapping of initial condition to solution at time
t of some linear, time-invariant, dierential equation. We remark that this will always
hold for any semigroup, see Lemma 2.2.6. Under this assumption, we can understand
these dening properties of a strongly continuous semigroup much better.
1. That T (t) is a bounded operator means that the solution at time t has not left
the space of initial conditions, i.e., the state space. The linearity implies that the
solution corresponding to the initial condition x0 + x0 equals x(t) + x(t), where

x(t) and x(t) are the solution corresponding to x0 and x0 , respectively. This is

logical, because we assumed that the underlying dierential equation is linear.

2. This is trivial; the solution at time zero must be equal to the initial condition.
3. Let x( ) be the state at time . If we take this as our new initial condition and
proceed for t seconds, then by the time-invariance this must equal x(t+ ). Since we
assume that T (s) is the mapping from x0 to x(s), we see that the time-invariance
of the underlying dierential equation implies property 3. of Denition 2.2.1.
Property 3. is known as a group property, and since it only holds for positive time,
it motivates the name semigroup.
4. This property tells you that if you go backward in time to zero, then x(t) approaches the initial condition. This sounds very logical, but need not to hold for
all operators satisfying 1.3.
Property 4. is known as strong continuity.
The easiest example of a strongly continuous semigroup is the exponential of a matrix.
That is, let A be an n n matrix, the matrix-valued function T (t) = eAt satises the
properties of Denition 2.2.1 on the Hilbert space Rn , see Exercise 2.3. Clearly the
exponential of a matrix is also dened for t < 0. If the semigroup can be extended to
all t R, then we say that T (t) is a group. We present the formal denition next.
Denition 2.2.2. Let X be a Hilbert space. The operator valued function t T (t),
t R is a strongly continuous group, or C0 -group, if the following holds
1. For all t R, T (t) is a bounded linear operator on X;
2. T (0) = I;
3. T (t + ) = T (t)T ( ) for all t, R.

21

2. Homogeneous dierential equation

4. For all x0 X, we have that T (t)x0 x0

converges to zero, when t 0.

It is easy to see that the exponential of a matrix is a group. However, only a few
semigroups are actually a group. In the study of p.d.e.s you encounter semigroups more
often than groups. The reason for that is that if you have a group, you may go from
the initial condition backward in time. For our simple p.d.e. of Section 2.1 this is not
possible, as it is shown at the end of the following example.
Example 2.2.3 In this example we show that the mapping dened by (2.5) denes a
strongly continuous semigroup. As state space we choose L2 (0, 1). We see that its norm
corresponds with the energy associated to this system.
Based on (2.4) and (2.5) we dene the following (candidate) semigroup on L2 (0, 1)
(T (t)x0 ) () =

x0 ( + t)
0

[0, 1], + t < 1

[0, 1], + t > 1

(2.6)

This is clearly a linear mapping. It is also bounded since

T (t)x0

=
0

| (T (t)x0 ) ()|2 d

max{1t,0}

=
0
max{1,t}

=
t
1

|x0 ( + t)|2 d

|x0 ()|2 d

|x0 ()|2 d = x0 2 .

(2.7)

From this we conclude that T (t) is bounded with bound less or equal to one.
Using (2.6), we see that for all x0 there holds T (0)x0 = x0 . Thus Property 2. of
Denition 2.2.1 holds.
We take an arbitrary function x0 L2 (0, 1) and call T ( )x0 = x1 , then for [0, 1],
we have
x1 ( + t) + t < 1
0
+ t > 1

x0 ( + t + ) z + t + < 1

= 0
+ t + > 1

0
+ t > 1

(T (t)x1 ) () =

Since the third case is already covered by the second one, we have that
(T (t)x1 ) () =

22

x0 ( + t + )
0

+ t + < 1
+ t + > 1

2.2. Semigroup and innitesimal generator

By (2.6) this equals (T (t + )x0 ) (), and so we have proved the third property of Denition 2.2.1. It remains to show the fourth property, that is the strong continuity. Since
we have to take the limit for t 0, we may assume that t < 1. Then
T (t)x0 x0

1t

=
0

|x0 ( + t) x0 ()|2 d +

1t

|x0 ()|2 d.

L2 (0, 1),

Since x0
the last term converges to zero when t 0. For the rst term some
more work is required.
First we assume that x0 is a continuous function. Then for every (0, 1), the
function x0 ( + t) converges to x0 () if t 0. Furthermore, we have that |x( + t)|
max[0,1] |x0 ()|. Using Lebesgue dominated convergence theorem, we conclude that
1t

lim
t0

|x0 ( + t) x0 ()|2 d = 0.

Hence for continuous functions we have proved that property 4. of Denition 2.2.1 holds.
This property remains to be shown for an arbitrary function in L2 (0, 1).
Let x0 L2 (0, 1), and let > 0. We can nd a continuous function x L2 (0, 1) such
that x0 x . Next we choose t > 0 such that T (t)x x for all t [0, t ].
By the previous paragraph this is possible. Combining this we nd for that t [0, t ]
T (t)x0 x0

T (t)x0 T (t)x + T (t)x x + x x0

T (t)(x0 x ) + T (t)x x + x x0

x x0 + T (t)x x + x x0

3,

where we used (2.7). Since this holds for all > 0, we have that Property 4. holds.
Having checked all dening properties for a C0 -semigroup, we conclude that (T (t))t0
given by (2.6) is a strongly continuous semigroup.
It is now easy to see that (T (t))t0 cannot be extended to a group. If there would be a
possibility to dene T (t) for negative t, then we must have that T (t)T (t) = T (t+t) =
T (0) = I for all t > 0. However, from (2.6) it is clear that T (2/) = 0, and so there
exists no operator Q such that QT (2/) = I, and thus (T (t))t0 cannot be extended to
a group.
So we have seen some examples of strongly continuous semigroups. It is easily shown
that the semigroup dened by (2.6) is strongly continuous for every t. This holds for any
semigroup, as is shown in Theorem 2.5.1. In that theorem more properties of strongly
continuous semigroups are listed.
Given the semigroup eAt t0 with A being a square matrix, one may wonder how
to obtain A. The easiest way to do this is by dierentiating eAt and evaluating this at
t = 0. We could try to do this with any semigroup. However, we only have that an
arbitrary semigroup is continuous, see property 4, and so it may be hard (impossible)
to dierentiate at zero. The trivial solution is that we only dierentiate T (t)x0 when it
is possible, as is shown next.

23

2. Homogeneous dierential equation

Denition 2.2.4. Let (T (t))t0 be a C0 -semigroup on the Hilbert space X. If the
following limit exists
T (t)x0 x0
,
(2.8)
lim
t0
t
then we say that x0 is an element of the domain of A, shortly x0 D(A), and we dene
Ax0 as
T (t)x0 x0
Ax0 = lim
.
(2.9)
t0
t
We call A innitesimal generator of the strongly continuous semigroup (T (t))t0 .

We may wonder what is the operator A for the semigroup of Example 2.2.3. We derive
this next.
Example 2.2.5 Consider the C0 -semigroup dened by (2.6). For this semigroup, we
would like to obtain A, and see how it is related to the original p.d.e. which we started
with (2.1).
For the semigroup of Example 2.2.3, we want to calculate (2.8). We consider rst the
limit for a xed [0, 1). Since < 1, there exists a small time interval [0, t0 ) such that
+ t < 1 for all t in this interval. Evaluating (2.8) at and assuming that t is in the
prescribed interval, we have
lim
t0

x0 ( + t) x0 ()
(T (t)x0 ) () x0 ()
= lim
.
t0
t
t

The later limit exists, when x0 is dierentiable, and for these functions the limit equals
dx0 (). So we nd an answer for < 1. For = 1, the limit (2.8) becomes
d
lim
t0

0 x0 (1)
.
t

We see that this limit will never exist, except if x0 (1) = 0.

Hence we nd that the domain of A will consists of all functions which are dierentiable
and which are zero at = 1. Furthermore,
Ax0 =

dx0
.
d

(2.10)

Since A has to map into L2 (0, 1), we see that the domain consists of functions in L2 (0, 1)
which are dierentiable and whose derivative lies in L2 (0, 1). This space is known as the
Sobolev space H 1 (0, 1). With this notation, we can write down the domain of A
D(A) = {x0 L2 (0, 1) | x0 H 1 (0, 1) and x0 (1) = 0}.

(2.11)

As in our example of Section 2.1, it turns out that every semigroup is related to a
dierential equation. This we state next. The proof of this lemma can be found in
Theorem 2.5.2.

24

2.2. Semigroup and innitesimal generator

Lemma 2.2.6. Let A be the innitesimal generator of the strongly continuous semigroup (T (t))t0 . Then for every x0 D(A), we have that T (t)x0 D(A) and
d
T (t)x0 = AT (t)x0 .
dt

(2.12)

So combining this with property 2. of Denition 2.2.1, we see that for x0 D(A),
x(t) := T (t)x0 is a solution of the (abstract) dierential equation
x(t) = Ax(t),

x(0) = x0 .

(2.13)

Although T (t)x0 only satises (2.13) if x0 D(A), we call T (t)x0 the solution for any
x0 .
It remains to formulate the abstract dierential equation for our semigroup of Examples 2.2.3 and 2.2.5.
Example 2.2.7 We have calculated A in (2.10), and so we can now easily write down
the abstract dierential equation (2.13). Since we have time and spatial dependence, we
use the partial derivatives. Doing so (2.13) becomes
x
x
= ,
t

which is our p.d.e. of Section 2.1. Note that the boundary condition at = 1 is not
explicitly visible. It is hidden in the domain of A.
We know now how to get an innitesimal generator for a semigroup, but normally, we
want to go into the other direction. Given a dierential equation, we want to nd the
solution, i.e., the semigroup. There exists a general theorem for showing this, but since
we shall not use it in the sequel, we dont include it here. We concentrate on a special
class of semigroups and groups, and hence generators.
Denition 2.2.8. A strongly continuous semigroup (T (t))t0 is called a contraction
semigroup if T (t)x0 X x0 X for all x0 X and all t 0.
A strongly continuous group is called a unitary group if T (t)x0 X = x0 X for all
x0 X and all t R.

If (T (t))t0 is a contraction semigroup, then the function f (t) := T (t)x0 2 must

X
have a non-positive derivative at t = 0, provided this derivative exists. Using the fact
that f (t) = T (t)x0 , T (t)x0 X , and that (2.13) holds for x0 D(A), it is easy to show
that the derivative of f equals for x0 D(A)
f(t) = AT (t)x0 , T (t)x0

+ T (t)x0 , AT (t)x0

X.

Hence if A is the innitesimal generator of a contraction semigroup, then

Ax0 , x0

+ x0 , Ax0

= f(0) 0.

It is not hard to show that if (T (t))t0 is a contraction semigroup, then the same holds
for (T (t) )t0 . Hence a similar inequality as derived above holds for A as well. Both
conditions are sucient as well.

25

2. Homogeneous dierential equation

Theorem 2.2.9. An operator A dened on the Hilbert space X is the innitesimal
generator of a contraction semigroup on X if and only if the following conditions hold
1. for all x0 D(A) we have that Ax0 , x0 + x0 , Ax0 0;
2. for all x0 D(A ) we have that A x0 , x0 + x0 , A x0 0.
For unitary groups there is a similar theorem.
Theorem 2.2.10. An operator A dened on the Hilbert space X is the innitesimal
generator of a unitary group on X if and only if A = A .

2.3. Homogeneous solutions to the port-Hamiltonian system

In this section, we apply the general result presented in the previous section to our p.d.e.,
i.e., we consider

x
(, t) = P1
[H()x(, t)] + P0 [H()x(, t)] .
t

(2.14)

with the boundary condition

WB

H(b)x(b, t)
H(a)x(a, t)

= 0.

(2.15)

In order to apply the theory of the previous section, we do not regard x(, ) as a function
of place and time, but as a function of time, which takes values in a function space, i.e.,
we see x(, t) as the function x(t, ) evaluated at . With a little bit of misuse of notation,
we write x(t, ) = (x(t)) (). Hence we forget the spatial dependence, and we write the
p.d.e. as the (abstract) ordinary dierential equation

dx
(t) = P1
[Hx(t)] + P0 [Hx(t)] .
dt

(2.16)

Hence we consider the operator

Ax := P1

d
[Hx] + P0 [Hx]
d

(2.17)

on a domain which includes the boundary conditions. The domain should be a part of the
state space X, which we identify next. For our class of p.d.e.s we have a natural energy
function, see (1.26). Hence it is quite natural to consider only states which have a nite
b
energy. That is we take as our state space all functions for which a x()T H()x()d
is nite. We assume that for every [a, b], H() a symmetric matrix and there exist
m, M , independent of , with 0 < m M < and mI H() M I. Under these
b
assumptions it is easy to see that the integral a x()T H()x()d is nite if only if x is
square integrable over [a, b]. We take as our state space
X = L2 ((a, b); Rn )

26

(2.18)

2.3. Homogeneous solutions to the port-Hamiltonian system

with inner product

1 b
f ()T H()g()d.
(2.19)
X
2 a
This implies that the squared norm of a state x equals the energy of this state.
So we have found our operator A and our state space. As mentioned above, the domain
of A will include the boundary conditions.
It turns out that formulating the boundary conditions directly in x at = a and = b
is not the best choice. It is better to formulate them in the boundary eort and boundary
ow , which are dened as
f, g

1
e = [H(b)x(b) + H(a)x(a)]
2

and

1
f = [P1 H(b)x(b) P1 H(a)x(a)] , (2.20)
2

respectively.
We show some properties of this transformation.
Lemma 2.3.1. Let P1 be symmetric and invertible, then the matrix R0 dened as
1
R0 =
2

P1 P1
I
I

(2.21)

is invertible, and satises

P1
0
0 P1

where

T
= R0 R0 ,

0 I
I 0

(2.22)

(2.23)

All possible matrices R which satises (2.22) are given by the formula
R = U R0 ,
with U satisfying U T U = .
Proof: We have that
1

P1 I
P1 I

0 I
I 0

1
=
2

P1 P1
I
I

P1
0
0 P1

1
Thus using the fact that P1 is symmetric, we have that R0 := 2 P1 P1 satises
I
I
(2.22). Since P1 is invertible, the invertibility of R0 follows from equation (2.22).
Let R be another solution of (2.22). Hence

RT R =

P1
0
0 P1

T
= R0 R0 .

This can be written in the equivalent form

T
1
R0 RT RR0 = .
1
Calling RR0 = U , we have that U T U = and R = U R0 , which proves the assertion.

27

2. Homogeneous dierential equation

Combining (2.20) and (2.21) we see that
f
e

= R0

(Hx) (b)
(Hx) (a)

(2.24)

Since the matrix R0 is invertible, we can write any condition which is formulated in
(Hx)(b) and (Hx)(a) into an equivalent condition which is formulated in f and e .
Furthermore, we see from (2.20) that the following holds
T
(Hx) (b)T P1 (Hx) (b) (Hx) (a)T P1 (Hx) (a) = f e + eT f .

(2.25)

Using (2.24), we write the boundary condition (2.15) (equivalently) as

WB

f
e

= 0,

(2.26)

where WB = WB R0 .
Thus we have formulated our state space X, see (2.18) and (2.19), and our operator,
A, see (2.17). The domain of this operator is given by, see (2.15) and (2.26),

D(A) = {x L2 ((a, b); Rn ) | Hx H 1 ((a, b); Rn ), WB

f
e

= 0}.

(2.27)

Here H 1 ((a, b); Rn ) are all functions from (a, b) to Rn which are square integrable and
have a derivative which is again square integrable.
The following theorem shows that this operator generates a contraction semigroup
precisely when the power (2.25) is negative, see (1.27).
Theorem 2.3.2. Consider the operator A dened in (2.17) and (2.27), where we assume
the following
P1 is an invertible, symmetric real n n matrix;
P0 is an anti-symmetric real n n matrix;
For all [a, b] the n n matrix H() is real, symmetric, and mI H() M I,
for some M, m > 0 independent of ;
WB is a full rank real matrix of size n 2n.
Then A is the innitesimal generator of a contraction semigroup on X if and only if
T
WB WB 0.
Furthermore, A is the innitesimal generator of a unitary group on X if and only if
T
WB satises WB WB = 0.
Proof: The proof is divided in several steps. In the rst step we simplify the expression
Ax, x + x, Ax X . We shall give the proof for the contraction semigroup in full detail.
The proof for the unitary group follows easily from it, see also Exercise 2.4. We write
WB = S(I + V, I V ), see Lemma 2.4.1.

28

2.3. Homogeneous solutions to the port-Hamiltonian system

Step 1. For the dierential operator (2.17) we have
Ax, x

+ x, Ax

1
2

P1
a
b

1
2

(Hx) () + P0 (Lx) ()

x()T H() P1

H()x()d+

(Hx) () + P0 (Hx) () d.

Using now the fact that P1 , H() are symmetric, and P0 is anti-symmetric, we write the
last expression as
1
2

b
a

d
(Hx) ()
d

P1 H()x() + [H()x()]T P1

d
(Hx) () d+
d

1 b
[H()x()]T P0 H()x() + [H()x(, t)]T [P0 H()x()] d
2 a
1 b d
=
(Hx)T ()P1 (Hx) () d
2 a d
1
= (Hx)T (b)P1 (Hx) (b) (Hx)T (a)P1 (Hx) (a) .
2
Combining this with (2.25), we see that
Ax, x

+ x, Ax

+ x, Ax

1 T
f e + eT f .

(2.28)

f lies in the kernel of W . Hence by using Lemma 2.4.2

B
e
IV
for some Rn . Substituting this in the above, we
IV

By assumption, the vector

we know that f equals

e
nd
Ax, x

1 T
f e + eT f

2
1
= T (I V T )(I V ) + T (I V T )(I V )
2
=T (I + V T V ).

(2.29)

By the assumption and Lemma 2.4.1, we have that this is less or equal to zero. Hence
we have proved the rst condition in Theorem 2.2.9.
Step 2. Let g L2 ((a, b), Rn ) be given. If for every f H 1 ((a, b); Rn ) which is zero at
= a and = b, the following equality hold for some g L2 ((a, b); Rn )

b
a

g()T

df
()d =
d

g ()T f ()d,

then g H 1 ((a, b); Rn ) and dg = .

g
d
Step 3. In this step we determine the adjoint operator of A. By denition, y X lies in
the domain of A if there exists a y X, such that

y, Ax

= y, x

(2.30)

29

2. Homogeneous dierential equation

for all x D(A).
Let x D(A), then (2.30) becomes
1
2

b
a

y()T H() P1

1
d
[Hx]() + P0 Hx() d =
d
2

b
a

y ()T H()x()d.

(2.31)

Since the half is on both sides, we neglect it and write the the left-hand side of this
equation as
b
a

y()T H() P1

d
[Hx]() + P0 Hx() d
d

b
d
y()T H()P0 Hx()d
[Hx]()d +
d
a
a
b
b
d
[Hx]()d
[P0 H()y()]T Hx()d, (2.32)
=
[P1 H()y()]T
d
a
a
b

y()T H()P1

where we have used that H and P1 are symmetric, and that P0 is anti-symmetric. The
last integral is already of the form y2 , x X . Hence it remains to write the second last
integral in a similar form. Since we are assuming that y D(A ) this is possible. In
particular, there exists a y1 X such that
b
a

[P1 H()y()]T

d
[Hx]()d =
d

y1 ()T H()x()d

(2.33)

for all x D(A). The set of functions x H 1 ((a, b); Rn ) which are zero at = a and
= b forms a subset of D(A), and so by step 2, and (2.33) we conclude that
P1 (Hy)() H 1 ((a, b); R)n and y1 =

d
[P1 Hy] .
d

Since P1 is constant and invertible, we have that Hy H 1 ((a, b); R)n . Integrating by
part we nd that for x D(A),
b

[P1 (Hy)()]T

d
[Hx]()d =
d

b
a

d
[P1 H()y()]T [Hx]()d
d

+ [P1 (Hy)()]T (Hx)()

b
a

(2.34)

The boundary term can be written as

b

30

(Hy)(b)
(Hy)(a)

[P1 (Hy)()] (Hx)()

(Hy)(b)
(Hy)(a)

(Hy)(b)
(Hy)(a)

P1
0
0 P1

(Hx)(b)
(Hx)(a)

T
R0 R0
T
T
R0

(Hx)(b)
(Hx)(a)

f
e

(2.35)

2.3. Homogeneous solutions to the port-Hamiltonian system

where we used (2.22) and (2.24).
Combining (2.32), (2.34), and (2.35) we nd that
b
a

y()T H() P1

d
[Hx]() + P0 Hx() d =
d
b

b
a

d
[P1 H()y()]T (Hx)()d
d

[P0 H()y()]T (Hx)()d

(Hy)(b)
(Hy)(a)

f
e

T
R0

(2.36)

The right-hand side of this equation must equal the inner product of x with some function
y . The rst two terms on the right-hand side are already in this form, and hence it

remains to write the last term in the requested form. However, since this last term
only depends on the boundary variables, this is not possible. Hence this term has to
disappear, i.e., has to be zero for all x D(A). Combining (2.27) and Lemma 2.4.2 we

see that f can be written as IV l, l Rn . We dene

e
I+V

f
e

(Hy)(b)
(Hy)(a)

= R0

(2.37)

Using this notation, the last term of (2.36) equals

(Hy)(b)
(Hy)(a)

T
T
R0

f
e

f
e

I V
I V

l.

The rst expression is zero for all x D(A), i.e., all f in the kernel of WB if and
e
n . By taking the transpose of this last
only if the second expression is zero for all l R

expression, we see that this holds if and only if f ker(I V T , I V T ).

e

Combining this with equation (2.36) we see that

A y =

d
d
[P1 Hy] P0 Hy = P1 [Hy] P0 Hy
d
d

(2.38)

and its domain equals

D(A ) = {y L2 ((a, b); Rn ) | Hy H 1 ((a, b); Rn ),
(I V T , I V T )

where

f
e

f
e

= 0},

(2.39)

is given by (2.37).

Step 4. In this step we show that A y, y X + y, A y 0 for all y D(A ). Since the
expression of A , see Step 3, is minus the expression for A, we can proceed as in Step 1.
By doing so, we nd
A y, y

+ y, A y =

1
(Hy)T (b)P1 (Hy) (b) (Hy)T (a)P1 (Hy) (a) .
2

(2.40)

31

2. Homogeneous dierential equation

Using (2.37), we can rewrite this as
A y, y

+ y, A y =

1 T
f e eT f .

Since f ker(I V T , I V T ), or equivalently,

e

conclude by Lemma 2.4.2 that

f
e

f
e

I VT
I V T

(2.41)

ker(I + V T , I V T ), we

for some Rn . Substituting this in (2.41) gives

A y, y

1 T
(I V )(I V T ) + T (I V )(I V T )
2
=T [I + V V T ].

+ y, A y =

From our condition on V , see the beginning of this proof, we see that the above expression
is negative. Hence using Theorem 2.2.9 we conclude that A generates a contraction
semigroup.
Step 5. It remains to show that A generates a unitary group when WB WB = 0. By
Lemma 2.4.1 we see that this condition on WB is equivalent to V being unitary. Now
we show that the domain of A equals the domain of A. Comparing (2.27) and (2.39)
we have that the domain are equal if and only if the kernel of WB equals the kernel of
(I V T , I V T ). Since V is unitary, we have
ker

I V T

I VT

= ker(V T

I +V

I V

) = ker(V T S 1 WB ) = ker WB ,

where in the last equality we used that S is invertible. Comparing (2.17), (2.27) with
(2.38), (2.39), and using the above equality, we conclude that A = A . Applying
Theorem 2.2.10 we see that A generates a unitary group.
We apply this Theorem to our standard example from the introduction, see also Example 2.2.5 and 2.2.7.
Example 2.3.3 Consider the homogeneous p.d.e. on the spatial interval [0, 1].
x
x
(, t) =
(, t),
t

x(, 0) = x0 (),
0 = x(1, t),

[0, 1], t 0
[0, 1]
t 0.

We see that the rst equation can be written in the from (2.14) by choosing P1 = 1,
H = 1 and P0 = 0. Using this and equation (2.20) the boundary variables are given by
1
f = [x(1) x(0)],
2

32

1
e = [x(1) + x(0)].
2

(2.42)

2.4. Technical lemmas

The boundary condition becomes in these variables
1
0 = x(1, t) = [f (t) + e (t)] = WB
2

f (t)
e (t)

(2.43)

1
1
with WB = ( 2 2 ).
T
Since WB WB = 1, we conclude from Theorem 2.3.2 that the operator associated
to the p.d.e. generates a contraction semigroup on L2 (0, 1). The expression for this
contraction semigroup is given in Example 2.2.3.

2.4. Technical lemmas

This section contains two technical lemmas on representation of matrices. They are
important for the proof, but not for the understanding of the examples.
Lemma 2.4.1. Let W be a n 2n matrix and let = 0 I . Then W has rank n
I 0
and W W T 0 if and only if there exist a matrix V Rnn and an invertible matrix
S Rnn such that
W = S I +V I V
(2.44)
with V V T I, or equivalently V T V I.
Furthermore, W W T > 0 if and only if V V T < I and W W T = 0 if and only if
V V T = I, i.e., V is unitary.
Proof: If W is of the form (2.44), then we nd
W W T = S

I +V

I V

I +VT
I VT

S T = S[2I 2V V T ]S T ,

which is non-negative, since V V T I.

Now we prove that if W is of full rank and is such that W W T 0, then relation
(2.44) holds. Writing W as W = W1 W2 , we see that W W T 0 is equivalent to
T
T
W1 W2 + W2 W1 0. Hence
(W1 + W2 )(W1 + W2 )T (W1 W2 )(W1 W2 )T 0.

(2.45)

If x ker((W1 + W2 )T ), then the above inequality implies that x ker((W1 W2 )T ).

T
T
Thus x ker(W1 ) ker(W2 ). Since W has full rank, this implies that x = 0. Hence
W1 + W2 is invertible.
Using (2.45) once more, we see that
(W1 + W2 )1 (W1 W2 )(W1 W2 )T (W1 + W2 )T I
and thus V := (W1 + W2 )1 (W1 W2 ) satises V V T I. Summarizing, we have
W1 W2

=
=

1
W1 + W2 + W1 W2 W1 + W2 W1 + W2
2
1
(W1 + W2 ) I + V I V .
2

33

2. Homogeneous dierential equation

Dening S := 1 (W1 + W2 ), we have shown the representation (2.44).
2
If instead of inequality, we have equality for W , then it is easy to show that we have
equality in the equation for V as well. Thus V is unitary.
Lemma 2.4.2. Suppose that the n 2n matrix W can be written in the format of
equation (2.44), i.e., W = S(I + V, I V ) with S and V square matrices, and S
IV
invertible. Then the kernel of W equals the range of IV .
If V is unitary, then the kernel of W equals the range of W T .
Proof: Let ( x1 ) be in the range of
x2
x1
x2

IV
IV

. By the equality (2.44), we have that

= S

I +V

I V

x1
x2

= S

I +V

I V

I V
I V

l = 0.

IV
Hence we see that the range of IV lies in the kernel of W . It is easy to show that
W has rank n, and so the kernel of W has dimension n. Thus if we can show that the
IV
2n n matrix IV has full rank, then we have proved the rst assertion. If this
matrix would not have full rank, then there should be a non-trivial element in its kernel.
It is easy to see that the kernel consists of zero only, and so we have proved the the rst
part of the lemma.
Suppose now that V is unitary, then

I V
I V

I + V T
I V T

V = W T S T V.

Since the range of W T equals the range of W T S T V , we have proved the second
assertion.

2.5. Properties of semigroups and their generators

This section contains some nice properties of semigroup and generators. Since the proofs
are rather long, we decided to put them separately. However, form time to time we shall
refer to one of these properties.
Theorem 2.5.1. A strongly continuous semigroup (T (t))t0 on the Hilbert space X
has the following properties:
a. T (t) is bounded on every nite subinterval of [0, );
b. T (t) is strongly continuous for all t [0, );
c. For all x X we have that

1
t

t
0 T (s)xds

x as t 0+ ;

d. If 0 = inf ( 1 log T (t) ), then 0 = lim ( 1 log T (t) ) < ;

t
t
t>0

34

2.5. Properties of semigroups and their generators

e. > 0 , there exists a constant M such that t 0, T (t) M et .
This constant 0 is called the growth bound of the semigroup.
Proof: a. First we show that T (t) is bounded on some neighborhood of the origin,
that is, there exist > 0 and M > 1 depending on such that
T (t) M

for t [0, ].

If this does not hold, then there exists a sequence {tn }, tn 0+ such that T (tn )
n. Hence, by the Uniform Boundedness Theorem A.3.19, there exists one x such that
{ T (tn )x } is unbounded; but this contradicts the strong continuity at the origin. If we
set t = m + with 0 , then
T (t) T ()

T ( ) M 1+m M M t/ = M et ,

where = 1 log M .
b. For xed t > 0, s 0 we have
T (t + s)x T (t)x T (t) T (s)x x M et T (s)x x .
Hence we may conclude that
lim

s0+

T (t + s)x T (t)x = 0.

Moreover, for t > 0 and s 0 suciently small, we have

T (t s)x T (t)x T (t s) x T (s)x .
Thus lim T (t + s)z T (t)x = 0, and T (t)x is continuous.
s0

c. Let x X and > 0. By the strong continuity of (T (t))t0 we can choose a > 0
such that T (s)x x for all s [0, ]. For t [0, ] we have that
1
t

t
0

T (s)xds x

1
t
1
t

t
0
t
0

[T (s)x x]ds
T (s)x x ds

1
t

ds = .
0

d. Let t0 > 0 be a xed number and M = sup T (t) ; then for every t t0 there exists
t[0,t0 ]

n N such that nt0 t < (n + 1)t0 . Consequently,

log T (t)
t

log T n (t0 )T (t nt0 )

t
n log T (t0 )
log M
+
t
t
log T (t0 ) nt0 log M

+
.
t0
t
t

35

2. Homogeneous dierential equation

log T (t0 )
+ logtM
t0
log T (t0 ) tt0
log M
if log
t0
t +
t

The latter is smaller than or equal to

is smaller than or equal to

lim sup
t

if log T (t0 ) is positive, and it

T (t0 ) is negative. Thus

log T (t)
log T (t0 )

< ,
t
t0

and since t0 is arbitrary, we have that

lim sup
t

log T (t)
log T (t)
log T (t)
inf
lim inf
.
t
t>0
t
t
t

Thus

log T (t)
log T (t)
= lim
< .
t
t>0
t
t
e. If > 0 , there exists a t0 such that
0 = inf

log T (t)
<
t

for t t0 ;

T (t) et

for t t0 .

that is,
But
T (t) M0

for 0 t t0 ,

and so with
M = M0 ,

for the case that > 0,

and
M = et0 M0 for the case that < 0,
we obtain the stated result.
Theorem 2.5.2. Let (T (t))t0 be a strongly continuous semigroup on a Hilbert space
X with innitesimal generator A. Then the following results hold:
a. For x0 D(A), T (t)x0 D(A) t 0;
b.

d
(T (t)x0 ) = AT (t)x0 = T (t)Ax0 for x0 D(A), t > 0;
dt

c.

dn
(T (t)x0 ) = An T (t)x0 = T (t)An x0 for x0 D(An ), t > 0;
dtn

d. T (t)x0 x0 =
e.

t
0 T (s)xds

dense in X;

t
0 T (s)Ax0 ds

D(A) and A

for x0 D(A);
t
0 T (s)xds

f. A is a closed linear operator;

36

= T (t)x x for all x X, and D(A) is

2.5. Properties of semigroups and their generators

g.

D(An ) is dense in Z.

n=1

Proof: a,b,c. First we prove a and b. Let s > 0 and consider

(T (s) I)x0
T (s) I
T (t + s)x0 T (t)x0
= T (t)
=
T (t)x0 .
s
s
s
If x0 D(A), the middle limit exists as s 0+ , and hence the other limits also
exist. In particular, T (t)x0 D(A) and the strong right derivative of T (t)x0 equals
AT (t)x0 = T (t)Ax0 .
For t > 0 and s suciently small, we have
(T (s) I)x0
T (t s)x0 T (t)x0
= T (t s)
.
s
s

Hence the strong left derivative exists and equals T (t)Ax0 . Part c follows by induction
on this result.
d. Take any x X and x0 D(A). Then
t

x , T (t)x0 x0 =

d
x , T (u)x0 du,
du

and hence
z , T (t)x0 x0

z , T (u)Ax0 du

for x0 D(A)

x ,

T (u)Ax0 du .
0

x was arbitrary and so this proves d.

e. We rst show that D(A) is dense in Z. Consider the following for any x X
T (s) I
s

T (u)xdu =
0

1
s

T (s + u)xdu

1
s

T (u)xdu.
0

These integrals are well dened, since T (t) is strongly continuous (Lemma A.5.5 and
Example A.5.15). Letting = s + u in the second integral, we have
T (s) I
s

T (u)xdu =
0

t+s

1
s
1
s

T ()xd

t+s

1
s
1
s

T (u)xdu
0
t

T ()xd

T ()xd +
s

t
s

1
s

T (u)xdu

T (u)xdu
0

s
0
s
0

(T (t + u) T (u))xdu
T (u)(T (t) I)xdu.

37

2. Homogeneous dierential equation

Now, as s 0+ , the right-hand side tends to (T (t) I)x (see Theorem 2.5.1.c). Hence
t

T (u)xdu D(A) and A

T (u)xdu = (T (t) I)x.

Furthermore, 1 0 T (u)xdu x as t 0+ , and hence for any x X, there exists a

t
sequence in D(A) that tends to x. This shows that D(A) = X.
f. To prove that A is closed, we let {nn }nN be a sequence in D(A) converging to x
such that Axn converges to y. Then T (s)Axn T (s)y M es Axn y and so
T (s)Axn T (s)y uniformly on [0, t]. Now, since xn D(A), we have that
t

T (t)xn xn =

T (s)Axn ds.
0

Using the Lebesgue dominated convergence Theorem A.5.21, we see that

t

T (t)x x =

T (s)yds,
0

and so
lim
t0

T (t)x x
1
= lim
t0 t
t

T (s)yds = y.
0

Hence x D(A) and Ax = y, which proves that A is closed.

g. Let C0 ([0, )) be the class of all real-valued functions on [0, ) having continuous
derivatives of all orders and having compact support contained in the open right half-line

(0, ). If C0 ([0, )), then so does (r) , the rth derivative of , and (u)T (u)x is
a continuous vector-valued function from [0, ) to X. Let X0 be the set of all elements
of the form

g=

x X, C0 ([0, )).

(u)T (u)xdu

These are well dened by Lemma A.5.5. We shall show that X0 D(Ar ) for r 1 and
that X0 is dense in X. For suciently small s, we have
T (s) I
g =
s
=

1
s
1
s

(u)[T (u + s)z T (u)x]du

[(u s) (u)]T (u)xdu

1
s

(u)T (u)xdu.
0

But (us)(u) (u) as s 0+ , uniformly with respect to u, and the last expression
s
is zero for suciently small s, since the support is contained in (0, ). Thus g D(A)

and Ag = 0 (u)T (u)xdu. Repeating this argument, we see that g D(Ar ) for all
r > 0, and

Ar g = (1)r
0

38

(r) (u)T (u)xdu

2.6. Exercises
which shows that X0

r=1

D(Ar ). Suppose now that the closure of X0 is not X. Then

there must exist a x0 X such that

x0 , g = 0 g X0 and
Thus

x0 ,

x0 = 1.

(u)T (u)xdu =
0

(u) x0 , T (u)x du = 0
0

C0 ([0, )) and x X. But

exists a C0 ([0, )) such that

and so X0 = X.

x0 , T (u)x0 is continuous with x0 = 1. Hence there

0 (u) x0 , T (u)x0 du = 0. This is a contradiction,

2.6. Exercises
2.1. Check equation (2.2).
2.2. In this exercise you will prove that the partial dierential equation
x
x
(, t) =
(, t),
t

[0, 1],

t 0,

with non-zero initial condition x(, 0) = x0 () and boundary condition

x(0, t) = 0
does not possess a solution.
a) Assume that the p.d.e. possesses a solution, show that for any continuously
1
dierentiable f satisfying f () = 0 for 1 the function q(t) = 0 f ( +
t)x(, t)d has derivative zero for t 0.

b) For the function dened in the previous item, show that q(1) = 0, independently of the value of f in the interval [0, 1).
c) Conclude from the previous two items that
continuously dierentiable functions f .

1
0 f ()x0 ()d

is zero for all

d) Prove that for any non-zero initial condition the p.d.e. with the chosen boundary condition doe not possess a solution in positive time.
2.3. Let A be a real n n matrix, and dene T (t) as eAt .
a) Show that (T (t))t0 is a strongly continuous semigroup on Rn .
b) Show that (T (t))tR is a strongly continuous group on Rn .
2.4. In this exercise we show that A generates a unitary group if and only if A and A
generate a contraction semigroup.

39

2. Homogeneous dierential equation

a) Prove that if (T (t))tR is a strongly continuous group satisfying T (t) 1
for all t R, then (T (t))tR is a unitary group.

b) Prove Theorem 2.2.10.

2.5. Consider dierential operator associated to the p.d.e.

x
x
(, t) = (, t),
t

[0, 1],

t 0,

(2.46)

with > 0 generates a strongly continuous semigroup on L2 (0, 1) for any of the
following boundary conditions.
a) The state at the right-hand side is set to zero, i.e., x(1, t) = 0.
b) The states at both ends are equal, i.e., x(1, t) = x(0, t).
c) Determine for = 1, the boundary conditions for the semigroup associated
to the p.d.e. (2.46) is a unitary group.
2.6. Consider dierential operator associated to the p.d.e.
x
x
(, t) = (, t),
t

[0, 1],

t 0,

(2.47)

generates a strongly continuous semigroup on L2 (0, 1) for any of the following

boundary conditions.
a) The state at the left-hand side is set to zero, i.e., x(0, t) = 0.
b) The states at both ends are equal, i.e., x(1, t) = x(0, t).
c) Determine the boundary conditions for the semigroup associated to the p.d.e.
(2.47) is a unitary group.
2.7. Consider the transmission line of Example 1.1.1. To this transmission line we put
a resistor at the right-end, i.e., V (b, t) = RI(b, t), and at the left-end we put the
voltage equal to zero. Show that the operator associated to this p.d.e. generates a
contraction semigroup on the energy space.
2.8. Consider the vibrating string of Example 1.1.2.
a) Formulate this in our general form (2.14). Determine P1 , P0 and H.

b) Reformulate the condition on H, see Theorem 2.3.2 in conditions on T and

. Are these conditions restrictive from a physical point of view?
c) Show that by imposing the no-force boundary conditions, i.e., w (b, t) =

w
(a, t) = 0 the system generates a unitary group on the energy space.

d) If the endpoints of string are hold at a constant position, i.e., w(a, t) =

w(b, t) = p, p R independent of time, does the operator associated to
the p.d.e. generate a strongly continuous semigroup on the energy space?

40

2.6. Exercises
2.9. Consider the Timoshenko beam from Example 1.1.3. Show that the operator
associated to this p.d.e. generates a C0 -semigroup on the energy space, when
a)
b)

w
w
t (a, t) = t (b, t) = 0,

w
t (a, t) = 0, t (a, t)
Q (b, t), Q 0.
t

t (a, t) = t (b, t) = 0.
0, w (b, t) = (b, t) +
t

and
=

(b, t), and

(b, t)

2.10. In the theory developed in this chapter, we considered the p.d.e.s of the spatial
interval [0, 1]. However, the theory is independent of this spatial interval. In this
exercise, we show that if we have proved a theorem for the spatial interval [0, 1],
then one can easily formulate the result for the general interval [a, b].
a) Assume that the spatial coordinate lies in the interval [a, b]. Introduce the
new spatial coordinate as
a
.
=
ba
Reformulate the p.d.e. (2.14) in the new spatial coordinate.
b) What are the new H, P0 , when P1 remains the same?

c) Determine the boundary eort and ow in the new coordinate.

d) How do the boundary conditions (2.15) and (2.26) change when using the
new spatial variable?

11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00

2.11. Consider coupled vibrating strings as given in the gure below. We assume that
II

11
00
11
00
11
00
11
00

I
III

Figure 2.1.: Coupled vibrating strings

all the length of the strings are equal. The model for every vibrating string is given
by (1.6) with physical parameters, I , TI , II , etc. Furthermore, we assume that the
three strings are connected via a (massless) bar, as shown in Figure 2.1. This bar
can only move in the vertical direction. This implies that the velocity of string I
at its right-hand side equals those of the other two strings at their left-hand side.
Furthermore, the force of string I at its right-end side equals the sum of the forces
of the other two at their left-hand side, i.e.,
TI (b)

wI
wII
wIII
(b) = TII (a)
(a) + TIII (a)
(a).

41

2. Homogeneous dierential equation

As depictured, the strings are attached to a wall.
a) Identify the boundary conditions for the system given in Figure 2.1.
b) Formulate the coupled strings as depictured in Figure 2.1 as a Port-Hamiltonian
system (2.14) and (2.15). Furthermore, determine the energy space X.
c) Show that the dierential operator associated to the above system generates
a contraction semigroup on the energy space X.
2.12. Consider transmission lines in a network as depictured in Figure 2.2. In the cou-

Trans. line II

V
Trans. line I

V
Trans. line IV

Trans. line III

Figure 2.2.: Coupled transmission lines

pling parts K, we have that Kirchho laws holds. Hence charge owing out of the
transmission line I, enters II and III, etc. Thus for the coupling between I and II
and III, there holds
VI (b) = VII (a) = VIII (a)
and
II (b) = III (a) + IIII (a).
a) Identify the boundary conditions for the system induced by the second coupling in Figure 2.2.
b) Formulate the coupled transmission lines as depictured in Figure 2.2 as a
Port-Hamiltonian system (2.14). Furthermore, identify the boundary eort
and ow, and the energy space X.
c) If the voltage at both ends is set to zero, i.e., VI (a) = 0, VIV (b) = 0, then
show that the dierential operator associated to the above system generates
a contraction semigroup on the energy space X.
d) If left end side is coupled to the right-hand side, i.e., the voltage and current
at both ends are equal, then show that the dierential operator associated to
the above system generates a unitary group on the energy space X.

2.7. Notes and references

More on semigroups can be found in [5]. The result of Section 2.3 are taken from [9].

42

Chapter 3

Boundary Control Systems

In this chapter we add a control function to the partial dierential equation (1.25), see
also (2.14). In particular, we are interested in boundary controls and we will show that
these systems have well-dened solutions. We explain our ideas by means of an example,
the controlled transport equation, which is given by
x
x
(, t) =
(, t),
t

x(, 0) = x0 (),
x(1, t) = u(t),

[0, 1], t 0
[0, 1]

(3.1)

t0

for a control function u. In Chapter 2 we have solved the partial dierential equation
for the specic choice u = 0. In this chapter we show that the solution of (3.1) is given
by
x0 ( + t)
+t1
x(, t) =
u( + t 1) + t > 1
and that in a similar manner the partial dierential equation (1.25) with a boundary
control can be treated.
In Section 3.1 we rst have to study some abstract theory, which enables us to show
that for certain boundary controls the partial dierential equation (1.25) possesses a
unique solution.

3.1. Inhomogeneous dierential equations

In Chapter 2 we have studied homogeneous (abstract) ordinary dierential equations of
the form
x(t) = Ax(t), x(0) = x0 .

(3.2)
Under the assumption that A generates a strongly continuous semigroup (T (t))t0 we
showed that the solution of (3.2) is given by
x(t) = T (t)x0 .

43

3. Boundary Control Systems

In order to be able to add control operators to equation (3.2) we have to study inhomogeneous (abstract) ordinary dierential equations rst. Such an inhomogeneous
equation is given by
x(t) = Ax(t) + f (t), x(0) = x0 .

(3.3)
We assume that f is a continuous function with values in the Hilbert space X. Later,
when we deal with input and control functions f (t) has usually the form Bu(t).
First we have to dene what we mean by a solution of (3.3), and we begin with the
notion of a classical solution. The function x(t) is called a classical solution of (3.3)
on an interval [0, ] if x(t) is a continuous function on [0, ] whose derivative is again
continuous on [0, ], x(t) D(A) for all t [0, ] and x(t) satises (3.3) for all t [0, ].
Assume that f C([0, ]; X) and that x is a classical solution of (3.3) on [0, ]. Then
formally we have
d
[T (t s)x(s)] = T (t s)x(s) AT (t s)x(s)

ds
= T (t s)[Ax(s) + f (s)] AT (t s)x(s)
= T (t s)f (s),

which implies
t

t
0

d
[T (t s)x(s)] ds = T (t t)x(t) T (t 0)x(0)
ds
0
= x(t) T (t)x0 .

T (t s)f (s) ds =

Equivalently,
t

T (t s)f (s)ds.

x(t) = T (t)x0 +
0

This equation is known as the variation of constant formula. The next lemma shows
that the formal argument which we used to derive this formula can be made precise. For
the proof we refer to Section 3.5.
Lemma 3.1.1. Assume that f C([0, ]; X) and that x is a classical solution of (3.3)
on [0, ]. Then Ax() is an element of C([0, ]; X), and
t

x(t) = T (t)x0 +
0

T (t s)f (s)ds.

(3.4)

This lemma tells us how a classical solution will look like, but does not tells us anything
about existence and uniqueness of solutions. This is the subject of the following theorem,
whose proof is given in Section 3.5
Theorem 3.1.2. If A is the innitesimal generator of a C0 -semigroup (T (t))t0 on a
Hilbert space X, f C 1 ([0, ]; X) and x0 D(A), then (3.4) is continuously dierentiable on [0, ] and it is the unique classical solution of (3.3).

44

3.1. Inhomogeneous dierential equations

By taking f = 0 and x0 D(A), it is clear that (3.4) will not be the classical solution
of (3.3) in general. Although (3.4) is only a classical solution for special initial states
x0 and smooth functions f , we see that this equation is well-dened for every t 0
and every x0 X and f L2 ([0, ]; X). Furthermore, for x1 D(A ), one can show
that x1 , x(t) is continuous dierentiable, and satises (3.4) weakly. Therefore, for any
x0 X and every f L2 ([0, ]; X) we call (3.4) the solution of (3.3). Since we sometimes
want to distinguish between this solution and the classical solution, we use the name
mild solution for the variation of constant formula (3.4).
Most times the inhomogeneous dierential equation (3.3) is of the form
x(t) = Ax(t) + Bu(t),

x(0) = x0 ,

(3.5)

where B is a linear bounded operator from the input space U to the state space X. The
spaces U and X are Hilbert spaces. u is considered as the (control) input and B is the
control operator. This is a special case of (3.3) via f (t) = Bu(t). Hence the (mild)
solution of (3.5) is given by
t

x(t) = T (t)x0 +
0

T (t s)Bu(s) ds

(3.6)

for every x0 X and every u L2 ([0, ]; X). We apply this to our simple example of
the transport equation.
Example 3.1.3 We study the following controlled partial dierential equation
x
x
(, t) =
(, t) + u(t)
t

x(1, t) = 0.
We can write this partial dierential equation as
x(t) = Ax(t) + Bu(t)

with, see also Examples 2.2.7 and 2.3.3,

Ax =
D(A) =

dx
,
d
x L2 (0, 1) | x H 1 (0, 1) and x(1) = 0 ,

Bu = 1 u.

In Chapter 2 we have shown that A generates a C0 -semigroup on L2 (0, 1). Using the
formula for the semigroup generated by A, see Examples 2.2.3 and 2.2.5 we get
x(, t) = x0 ( + t)[0,1] ( + t) +
where

t
0

[0,1] ( + t )u( ) d,

[0,1] () = 1 if [0, 1] and [0,1] () = 0 otherwise.

45

3. Boundary Control Systems

In this example we applied a control within the spatial domain. More logical for
this example is to apply a control at the boundary. However, when doing so, we cannot
rewrite this system in our standard form (3.5), see Example 3.2.3. This is general the case
when controlling a p.d.e. via its boundary. Thus systems with control at the boundary
form a new class of systems, and are introduced next.

3.2. Boundary control systems

We rst explain the idea by means of the controlled transport equation (3.1). Consider
the following system
x
x
(, t) =
(, t),
t

x(, 0) = x0 (),
x(1, t) = u(t),

[0, 1], t 0
[0, 1]

(3.7)

t 0.

for an input u L2 (0, ).

Boundary control problems like the one above occur frequently in our applications,
but unfortunately they do not t into our standard formulation (3.5). However, for
suciently smooth inputs it is possible to reformulate such problems so that they do
lead to an associated system in the standard form (3.5). In order to nd the associated
system for the controlled transport equation we use the following trick. Assume that x is
a classical solution of the p.d.e. (3.7) and that u is continuously dierentiable. Dening
v(, t) = x(, t) u(t),
we obtain the following partial dierential equation for v
v
v
(, t) =
(, t) u(t),

v(1, t) = 0,
t 0.

[0, 1], t 0

This partial dierential equation for v can be written in the standard form as
v(t) = Av(t) + B u(t)

for u = u. Hence by applying a simple trick, we can reformulate a p.d.e. with boundary

control into a p.d.e. with internal control. The price we have to pay is that u has to be
smooth.
The trick applied to (3.7) can be extended to abstract boundary control systems:
x(t) = Ax(t),

Bx(t) = u(t),

x(0) = x0 ,

(3.8)

where A : D(A) X X, u(t) U , U is a Hilbert space, and the boundary operator

B : D(B) X U satises D(A) D(B).
In order to reformulate equation (3.8) into an abstract form (3.5), we need to impose
extra conditions on the system.

46

3.2. Boundary control systems

Denition 3.2.1. The control system (3.8) is a boundary control system if the following
hold:
a. The operator A : D(A) X with D(A) = D(A) ker(B) and
Ax = Ax

for x D(A)

(3.9)

is the innitesimal generator of a C0 -semigroup on X;

b. There exists a B L(U, X) such that for all u U , Bu D(A), the operator AB
is an element of L(U, X) and
BBu = u,

u U.

(3.10)

Part b. of the denition is equivalent to the fact that the range of the operator B
equals U . Note that Part a. of the denition guarantees that the system possesses a
unique solution for the choice u(t) = 0, i.e., the homogeneous equation is well-posed.
Part b. allows us to choose every value in U for u(t). In other words, the values of inputs
are not restricted, which is a logical condition for inputs.
We say that the function x(t) is a classical solution of the boundary control system
of Denition 3.2.1 if x(t) is a continuously dierentiable function, x(t) D(A) for all t,
and x(t) satises (3.8) for all t.
For a boundary control system, we can apply a similar trick as the one applied in the
beginning of this section. This is the subject of the following theorem. It turns out that
v(t) = x(t) Bu(t) is a solution of the abstract dierential equation
v(t) = Av(t) B u(t) + ABu(t),

v(0) = v0 .

(3.11)

Since A is the innitesimal generator of a C0 -semigroup and B and AB are bounded

linear operators, we have from Theorem 3.1.2 that equation (3.11) has a unique classical
solution for v0 D(A). Furthermore, we can prove the following relation between the
(classical) solutions of (3.8) and (3.11).
Theorem 3.2.2. Consider the boundary control system (3.8) and the abstract Cauchy
equation (3.11). Assume that u C 2 ([0, ]; U ) for all > 0. Then, if v0 = x0 Bu(0)
D(A), the classical solutions of (3.8) and (3.11) are related by
v(t) = x(t) Bu(t).

(3.12)

Furthermore, the classical solution of (3.8) is unique.

Proof: Suppose that v(t) is a classical solution of (3.11). Then v(t) D(A) D(A)
D(B), Bu(t) D(B), and so
Bx(t) = B[v(t) + Bu(t)] = Bv(t) + BBu(t) = u(t),

47

3. Boundary Control Systems

where we have used that v(t) D(A) ker B and equation (3.10). Furthermore, from
(3.12) we have
x(t) = v(t) + B u(t)

= Av(t) B u(t) + ABu(t) + B u(t)

by (3.11)

= A(v(t) + Bu(t))

by (3.9)

= Ax(t)

by (3.12).

= Av(t) + ABu(t)

Thus, if v(t) is a classical solution of (3.11), then x(t) dened by (3.12) is a classical
solution of (3.8).
The other implication is proved similarly. The uniqueness of the classical solutions of
(3.8) follows from the uniqueness of the classical solutions of (3.11).
The (mild) solution of (3.11) is given by
t

v(t) = T (t)v(0) +
0

T (t )[ABu( ) B u( )] d

(3.13)

for every v(0) X and every u H 1 (0, ); U ). Therefore, the function

t

x(t) = T (t)(x0 Bu(0)) +

T (t )[ABu( ) B u( )] d + Bu(t)

(3.14)

is called the mild solution of the abstract boundary control system (3.8) for every x0 X
and every u H 1 ((0, ); U ).
As an example we study again the controlled transport equation from the beginning
of this section.
Example 3.2.3 We consider the following system
x
x
(, t) =
(, t),
t

x(, 0) = x0 (),
x(1, t) = u(t),

[0, 1], t 0
[0, 1]
t 0.

for an input u H 1 (0, ). In order to write this example in the form (3.8) we choose
X = L2 (0, 1) and
dx
,
d
Bx = x(1),
Ax =

D(A) = x L2 (0, 1) | x H 1 (0, 1) ,

D(B) = D(A).

These two operators satisfy the assumption of a boundary control system. More precisely: the operators A and B are linear, A restricted to the domain D(A) ker B
generates a C0 -semigroup, see Example 2.2.7. Furthermore, the range of B is C = U

48

3.2. Boundary control systems

and the choice B = 1 implies BBu = u. Using AB = 0, we conclude from equation
(3.14) that the solution is given by
x(t) = v(t) + Bu(t)
t

T (t )[ABu( ) B u( )] d + Bu(t)

= T (t)v(0) +
0
t

= T (t)v(0)

T (t )u( ) d + u(t).

Using the precise form of the shift-semigroup, see Example 2.2.3, we can write the
solution of the boundary controlled partial dierential equation as
x(, t) = v0 ( + t)[0,1] ( + t)

t
0

[0,1] ( + t )u( ) d + u(t).

If + t > 1, we have
x(, t) = [u( )]|t
+t1 + u(t) = u( + t 1),
and if + t 1, then
x(, t) = v0 ( + t) [u( )]|t + u(t) = v0 ( + t) + u(0) = x0 ( + t).
0
Or equivalently,
x(, t) =

x0 ( + t)
+t1
u( + t 1) + t > 1

(3.15)

which proves our claim made on the rst page of this chapter.
Now it is not hard to show that the controlled p.d.e. form this example cannot be
written as the abstract control system
x(t) = Ax(t) + Bu(t),

(3.16)

with B a bounded operator. Since for u = 0, we have that the p.d.e. becomes the
homogeneous equation of Examples 2.2.5 and 2.2.7, we have that A can only be A
restricted to D(A) ker B. Hence the semigroup is the shift semigroup.
If our controlled p.d.e. would be of the form (3.16), then by Theorem 3.1.2, we would
have that x(t) D(A), whenever x0 D(A) and f C 1 ((0, ; X). Choosing x0 D(A)
and u = 1, we see by (3.15), that x(1, t) = u(t). Since this is unequal to zero, we
have that x(t) D(A). Concluding we nd that the boundary controlled p.d e. of this
example cannot be written in the form (3.16).
The controlled transport equation is a simple example of our general class of portHamiltonian systems. This example could be written as a boundary control system. In
the following section, we show that this holds in general for a port-Hamiltonian system.

49

3. Boundary Control Systems

3.3. Port-Hamiltonian systems as boundary control systems

In this section we add a boundary control to our Hamiltonian system and we show
that the assumptions of a boundary control system are satised. The port-Hamiltonian
system with control is given by

x
(, t) = P1 [Hx(t)] + P0 [Hx(t)]
t

f (t)
u(t) = WB,1
e (t)
0 = WB,2

(3.17)
(3.18)

f (t)
.
e (t)

(3.19)

We rst recall the dening operators of the port-Hamiltonian system. As in Section 2.3
we assume that
P1 is an invertible, symmetric real n n-matrix;
P0 is an anti-symmetric real n n-matrix;
H() is a symmetric, invertible n n-matrix for every [a, b] and mI H()
M I for some m, M > 0 independent of ;
WB :=

WB,1
WB,2

is a full rank real matrix of size n 2n.

We recall that the boundary eort and ow are given by, see (2.24)
f (t)
e (t)

= R0

H(b)x(b, t)
H(a)x(a, t)

where R0 is the invertible n n-matrix dened in (2.21).

We can write the Hamiltonian system as a boundary control system
x(t) = Ax(t),

x(0) = x0 ,

Bx(t) = u(t),
by dening
Ax = P1
D(A) =

[Hx] + P0 [Hx],

x L2 ((a, b); Rn ) | Hx H 1 ((a, b); Rn ), WB,2

Bx = WB,1
D(B) = D(A).

50

f
,
e

(3.20)
f
e

=0 ,

(3.21)
(3.22)
(3.23)

3.3. Port-Hamiltonian systems as boundary control systems

As in Section 2.3 we have chosen the Hilbert space X = L2 ((a, b); Rn ) equipped with
the inner product
1 b
f ()T H()g() d
f, g X :=
2 a
as our state space. The input space U equals Rk , where k is the number of rows of WB,1 .
We are now in a position to show that the controlled port-Hamiltonian system is indeed
a boundary control system.
Theorem 3.3.1. If
0 I
I 0

WB,1
WB,2

T
T
WB,1 WB,2 0,

(3.24)

then the system (3.17)(3.19) is a boundary control system on X. Furthermore, the

operator

Ax = P1 [Hx] + P0 [Hx]
(3.25)

with the domain

D(A) =

Hx H 1 ((a, b); Rn |

f
e

ker

WB,1
WB,2

(3.26)

generates a contraction semigroup on X.

Proof: We begin with the simple observation that
f
e

WB,1
WB,2

ker

is equivalent to
WB

f
e

WB,1
WB,2

f
e

= 0.

From Theorem 2.3.2 follows that the operator A dened in (3.25) and (3.26) is the
innitesimal generator of a contraction semigroup on X. Moreover, by (3.21) and (3.22)
we have that D(A) = D(A) ker B. Hence part a. of Denition 3.2.1 is satised.
The n 2n-matrix WB is of full rank n and R0 is an invertible matrix. Thus there
exists a 2n n-matrix S such that
W B R0 S =

WB,1
R0 S =
WB,2

Ik 0
,
0 0

where Ik is the identity matrix on Rk . We write S =

operator B L(U, X) by
(Bu)() := H()1 S11

(3.27)

S11 S12
S21 S22

b
a
+ S21
ba
ba

and we dene the

u.

51

3. Boundary Control Systems

The denition of B implies that Bu is a square integrable function and that Hx
H 1 ((a, b); Rn ). Furthermore, from (3.27) it follows that WB,2 R0 S11 = 0. Combining
S21
this with the denition of the boundary eort and ow, we obtain that Bu D(A).
Furthermore, B and AB are linear bounded operators from U to X and using (3.27)
once more we see that
S1
BBu = WB,1 R0
u = u.
S2
Thus the Hamiltonian system is indeed a boundary control system.
As an example we once more study the controlled transport equation.
Example 3.3.2 We consider the system
x
x
(, t) =
(, t),
t

x(, 0) = x0 (),

[0, 1], t 0

(3.28)

[0, 1].

This system can be written in form (3.17) by choosing n = 1, P0 = 0, P1 = 1 and H = 1.

Therefore, we have
1
R0 =
2

1 1
1 1

and

f (t)
e (t)

1
=
2

x(1, t) x(0, t)
.
x(1, t) + x(0, t)

Since n = 1, we can either apply one control or no control at all. The control free case
has been treated in Chapter 2, and so we choose one control. By using the boundary
variables, the control is written as, see (3.18)
1
u(t) = a b
2

x(1, t) x(0, t)
x(1, t) + x(0, t)

1
= [(a + b)x(1, t) + (b a)x(0, t)] .
2

(3.29)

Note that WB = (a, b) has full rank if and only if a2 + b2 = 0.

By Theorem 3.3.1 gives that the p.d.e. (3.28) together with (3.29) is a boundary
control system if a2 + b2 = 0 and 2ab 0, see (3.24). Thus possible boundary controls
are for example

2
u(t) = x(1, t),
(a = b =
),
2
u(t) = 3x(1, t) x(0, t),
(a = 2, b = 2 2).
For the control u(t) = x(1, t) + 3x(0, t) we dont know the answer.

3.4. Outputs
In the previous sections we have added a control function to our systems. In this section
additionally an output is added. We follow the line laid out in the previous sections. We

52

3.4. Outputs
start by assuming that the output equation can be represented via a bounded operator.
Since this is normally not the case for observations at the boundary, we have to consider
boundary observation separately. For this we directly formulate the result for our portHamiltonian systems.
We start with the control system 3.5 to which we add a output equation.
x(t) = Ax(t) + Bu(t),

x(0) = x0 ,

y(t) = Cx(t) + Du(t),

(3.30)
(3.31)

where B is a linear bounded operator from the input space U to the state space X, C
is a linear bounded operator from X to the output space Y , and D is a linear operator
from U to Y . All Spaces U , X and Y are Hilbert spaces.
In Section 3.1 we showed that the solution of (3.30) is given by
t

T (t s)Bu(s) ds.

x(t) = T (t)x0 +
0

This function is well-dened for every x0 X and every u L2 ([0, ]; U ). The output
equation (3.31) only contains bounded operators. Hence there is no diculty in solving
this equation. We summarize the answer in the following theorem
Theorem 3.4.1. Consider the abstract equation (3.30)(3.31), with A the innitesimal
generator of the C0 -semigroup (T (t))t0 , and B, C, and D bounded.
The solution of (3.30)(3.31) is given by the variation of constant formula (3.6)
t

x(t) = T (t)x0 +
0

T (t s)Bu(s) ds,
t

y(t) = CT (t)x0 + C
0

T (t s)Bu(s) ds + Du(t)

for every x0 X and every u L2 ([0, ]; U ).

As said in the beginning of this section, our main focus lies on boundary observation for port-Hamiltonian systems. We develop conditions on the boundary observation
guaranteeing that a certain balance equation is satised, which are important for Chapter 4 and 5. The standard Hamiltonian system with boundary control and boundary
observation is given by

[Hx(t)] + P0 [Hx(t)]

f (t)
u(t) = WB
e (t)
x(t) = P1

y(t) = WC

f (t)
.
e (t)

(3.32)
(3.33)
(3.34)

It is assumed that (3.32)(3.33) satisfy the assumption of a port-Hamiltonian system, see

T
Section 3.3. Further we assume that WB WB 0, where = 0 I . This guarantees
I 0

53

3. Boundary Control Systems

that (3.32)(3.33) is a boundary control system, see Theorem 3.3.1. Note that we have
taken WB = WB,1 or equivalently WB,2 = 0. In other words, we are using the maximal
number of control.
The output equation is formulated very similar to the input equation. As for the input
we assume that we use full measurements, i.e., WC be a full rank matrix of size n 2n.
Since we do not want to measure quantities that we already have chosen as inputs, see
(3.33), we assume that WB is of full rank, or equivalently this matrix is invertible.
WC
Combining this assumption with the fact that is invertible, we see that the product
WB
T
W T WC ) is invertible as well. Its inverse is dened as
WC ( B
PWB ,WC =

WB
T
WB
WC

T
WC

T
WB WB
T
WC WB

T
WB WC
T
WC WC

(3.35)

Theorem 3.4.2. Consider the system (3.32)(3.34) with WB a full rank n 2n matrix
T
satisfying WB WB 0, and WC a full rank n 2n matrix such that WB is invertible.
WC

f (0)
, the
e (0)
H 1 ((a, b); Rn ). The

For every u C 2 (0, ; Rn ), Hx(0) H 1 ((a, b); Rn ), and u(0) = WB

system (3.32)(3.34) has a unique (classical) solution, with Hx(t)

output y() is continuous, and the following balance equation is satised:
d
x(t)
dt

2
X

1 T
u (t) y T (t) PWB ,WC
2

u(t)
.
y(t)

(3.36)

Proof: See Exercise 3.1.

As an example we return to the controlled transport equation of Example 3.3.2.
Example 3.4.3 We consider the system
x
x
(, t) =
(, t),
t

x(, 0) = x0 (),
u(t) = x(1, t),

[0, 1], t 0

(3.37)

[0, 1]
t 0.

In Example 3.3.2 we have already seen that this system can be written in form (3.32)

(3.33) by choosing n = 1, P0 = 0, P1 = 1, H = 1 and WB = ( 22 22 ).

Now we add the output equation
1
y(t) = c d
2

x(1, t) x(0, t)
x(1, t) + x(0, t)

1
= (c + d)x(1, t) + (c d)x(0, t)).
2

(3.38)

Since WC = (c d ) must have full rank, we nd that c2 + d2 = 0. Furthermore, since

WB
must be invertible, we nd that c = d.
WC
Hence all conditions of Theorem 3.4.2 satisfying for this system whenever c = d.
are
For the particular choice c = d = 22 , that is y(t) = x(0, t), we nd PWB ,WC =
1 0 , or equivalently
0 1
d
x(t)
dt

54

2
X

1
|u(t)|2 |y(t)|2 .
2

3.5. Some proofs

3.5. Some proofs

This section contains the proofs of Lemma 3.1.1 and Theorem 3.1.2.
Proof: (of Lemma 3.1.1) From (3.3), we have that Ax(t) = x(t) f (t) and x

C([0, ]; X) shows that Ax() C([0, ]; X).

We now prove (3.4). Let t be an arbitrary, but xed, element of (0, ] and consider
the function T (t s)x(s) for s [0, t). We shall show that this function is dierentiable
in s. Let h be suciently small and consider
T (t s h)x(s + h) T (t s)x(s)
h
T (t s h)x(s + h) T (t s h)x(s)
+
=
h
T (t s h)x(s) T (t s)x(s)
.
h
If h converges to zero, then the last term converges to AT (ts)x(s), since x(s) D(A).
Thus it remains to show that the rst term converges. We have the following equality
T (t s h)x(s + h) T (t s h)x(s)
T (t s)x(s)

h
x(s + h) x(s)
= T (t s h)
T (t s h)x(s) +

h
T (t s h)x(s) T (t s)x(s).

The uniform boundedness of T (t) on any compact interval and the strong continuity
allow us to conclude from the last equality that
lim T (t s h)

h0

x(s + h) x(s)
T (t s)x(s) = 0.

So we have proved that

d
[T (t s)x(s)] = T (t s)x(s) AT (t s)x(s)

ds
= T (t s)[Ax(s) + f (s)] AT (t s)x(s)
= T (t s)f (s).

This implies
t
0

d
[T (t s)x(s)] ds = T (t t)x(t) T (t 0)x(0)
0 ds
= x(t) T (t)x0 .

T (t s)f (s) ds =

Thus a classical solution to (3.3) necessarily has the form (3.4).

55

3. Boundary Control Systems

Proof: (of Theorem 3.1.2) Uniqueness: If z1 and z2 are two dierent solutions, then
their dierence (t) = z1 (t) z2 (t) satises the dierential equation
d
= A,
dt

(0) = 0

and so we need to show that its only solution is (t) 0. To do this, dene y(s) =
dy
T (t s)(s) for a xed t and 0 s t. Clearly,
= 0 and so y(s) = constant
ds
= T (t)(0) = 0. However, y(t) = (t) shows that (t) = 0.
t

Existence: Clearly, all we need to show now is that v(t) = 0 T (ts)f (s)ds is an element
of C 1 ([0, ]; X) D(A) and satises dierential equation (3.3). Now
s

T (t s)[f (0) +

v(t) =
0

f()d]ds

0
t

=
0

T (t s)f (0)ds +

T (t s)f()dsd,

where we have used Fubinis Theorem A.5.22. From Theorem 2.5.2.e, it follows that
t

T (t )z z = A

T (t s)zds

for all z Z.

t
t
t

Hence v(t) D(A), and 0 A T (t s)f()ds d = 0 T (t )f() f() d < .

Thus, since A is closed, by Theorem A.5.23 we have that
t

Av(t) = [T (t) I]f (0) +

t

= T (t)f (0) +
0

[T (t ) I]f()d

T (t )f()d f (t).

Now, since the convolution product is commutative, i.e.,

s)ds, we have that

t
0 g(t s)h(s)ds

t
0 g(s)h(t

v(t) =
0

T (s)f (t s)ds

and so
dv
(t) = T (t)f (0) +
dt

t
0
t

= T (t)f (0) +
0

T (s)f(t s)ds
T (t s)f(s)ds,

once again using commutativity of the convolution product. It follows that

uous and
dv
(t) = Av(t) + f (t).
dt

56

dv
dt

is contin-

3.6. Exercises

3.6. Exercises
3.1. Prove Theorem 3.4.2. Hint: Use the calculations of the proof of Theorem 2.3.2.
3.2. Consider the transmission line of Example 1.1.1 for any of the following boundary
condition:
a) At the left-end the voltage equal u1 (t) and at the right-end the voltage equals
the input u2 (t).
b) At the left-end we put the voltage equal to zero and at the right-end the
voltage equals the input u(t).
c) At the left-end we put the voltage equal to u(t) and at the right-end the
voltage equals R times the current, for some R > 0.
Show that these systems can be written as a boundary control system.
3.3. Consider the vibrating string of Example 1.1.2 with the boundary conditions
x
(0, t) = 0

x
(1, t) = u(t)

and

t 0.

Reformulate this system as a boundary control system.

3.4. Consider the Timonshenko beam from Example 1.1.3 for any of the following
boundary condition:
a)
b)

w
t (a, t) = u1 (t),
w
t (a, t) = u(t),
Q (b, t), Q
t

w
t (b, t)

t (a, t)

0.

t (a, t) = t (b, t) = 0.

w
t (b, t) = (b, t) + (b, t),

= u2 (t), and
= 0,

and

(b, t)

Reformulate these systems as boundary control systems.

3.5. We consider again the transmission line of Exercise 3.2 with boundary conditions
a). Dene an output to the system such that
d
x(t)
dt

2
X

= u(t)T y(t).

3.6. In the formulation of port-Hamiltonian systems as boundary control systems, we

have the possibility that some boundary conditions are set to zero, see (3.19).
However, when we add an output, this possibility was excluded, see (3.32)(3.34).
Furthermore, we assumed that we had n outputs. In this exercise we show that
this did not pose a restriction to the theory.
T
a) Show that if WB,2 = 0, i.e., WB = WB,1 , and WB WB 0, then (3.17) with
control (3.18) is a well-dened boundary control system. Given the expression
for the domain of the innitesimal generator A.

57

3. Boundary Control Systems

T
b) Let WB be a n2n matrix of full rank satisfying WB WB 0. We decompose
u = WB f as
e
u1
WB,1
f
u=
=
.
u2
WB,2
e

By using the previous item, show that putting u2 = 0 is allowed. Furthermore,

show that it leads to the same boundary control system as in (3.17)(3.19).
Consider the following system

x
(, t) = P1 [Hx(t)] + P0 [Hx(t)]
t

f (t)
u(t) = WB,1
e (t)
f (t)
e (t)

0 = WB,2

f (t)
e (t)

y(t) = WC
WB,1
WB,2

where we assume that WB =

T
WB WB

is a full rank matrix of size n 2n satisfying

0. Furthermore, WC is a full-rank matrix of size k 2n satisfying

rank

WB
WC

= n + k.

c) Explain why the above rank conditions are logical.

d) Show that the above system is well-dened. In particular show that for
smooth inputs and initial conditions, we have a unique solution, with the
state and output trajectory continuous.
3.7. Consider the vibrating string of Exercise 3.3 with the boundary conditions
x
(0, t) = 0

and

x
(1, t) = u(t)

t 0.

and we measure the velocity at = 0.

a) Prove that this is a well-dened input-output system.
Hint: Use the previous exercise.
b) Does a balance equation like (3.36) hold?
c) Repeat the above two questions for the measurement y(t) =

w
t (1, t).

3.8. Consider the coupled strings of Exercise 2.11. Now we apply a force u(t), to the
bar in the middle, see Figure 3.1. This implies that the force balance in the middle
becomes
wII
wIII
wI
(b) = TII (a)
(a) + TIII (a)
(a) + u(t).
TI (b)

58

3.6. Exercises

II

I
III

11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00

u(t)

11
00
11
00
11
00
11
00
11
00

Figure 3.1.: Coupled vibrating strings with external force

a) Formulate the coupled vibrating strings with external force as a boundary
control system.
b) Additionally, we measure the velocity of the bar in the middle. Reformate
the system with this output as (3.32)(3.34).
c) For the input and output dened above, determine the power balance in terms
of the input and output, see (3.36).

59

3. Boundary Control Systems

60

Chapter 4

Transfer Functions
In this chapter we discuss the concept of transfer functions. Let us rst recapitulate the
concept for nite-dimensional systems. Consider the ordinary dierential equation
y (t) + 3y(t) 7y(t) = u(t) + 2u(t),

(4.1)

where the dot denotes the derivative with respect to time. In many textbooks one derives
the transfer function by taking the Laplace transform of this dierential equation under
the assumption that the initial conditions are zero. Since the following rules hold for the
Laplace transform
f(t) F (s) f (0)

f (t) s2 F (s) sf (0) f(0),

we have that after Laplace transformation the dierential equation becomes the algebraic
equation:
s2 Y (s) + 3sY (s) 7Y (s) = sU (s) + 2U (s).
(4.2)
This implies that
Y (s) =

s2

s + 2
U (s).
+ 3s 7

(4.3)

The rational function in front of U (s) is called the transfer function associated with the
dierential equation (4.1).
This is a standard technique, but there are some diculties with it if we want to extend
it to partial dierential equations. One of the diculties is that one has to assume that u
and y are Laplace transformable. Since u is chosen, this is not a strong assumption, but
once u is chosen, y is dictated by the dierential equation, and it is not known a priory
whether it is Laplace transformable. Furthermore, the Laplace transform only exists
in some right half-plane of the complex plane. This implies that we have the equality
(4.3) for those s in the right-half plane for which the Laplace transform of u and y both
exist. The right-half plane in which the Laplace transform exists is named the region of
convergence. Even for the simple dierential equation (4.1) equality (4.3) does not hold

61

4. Transfer Functions
everywhere. Taking into account the region of convergence of both u and y, we nd that
(4.3) only holds for those s which lies right of the poles, i.e., the zeros of s2 + 3s 7.
To overcome all these diculties we dene the transfer function in a dierent way.
We shall look for solutions of the dierential equation which are exponentials. Let us
illustrate this for the simple dierential equation of (4.1). Given s C, we look for a
solution pair of the form (u(t), y(t)) = (est , ys est ). If for an s such a solution exists, and
it is unique, then we call ys the transfer function of (4.1) in the point s. Substituting
this pair into our dierential equation, we have
s2 ys est + 3sys est ys est = sest + 2est .

(4.4)

We recognize the common term est which is never zero, and so we may divide by it.
After this division, we obtain
s2 ys + 3sys ys = s + 2.

(4.5)

This is uniquely solvable for ys if and only if s2 + 3s 7 = 0.

We see that we have obtained, without running into mathematical diculties, the
transfer function. Since for p.d.e.s or for abstract dierential equations the concept of
a solution is well-dened, we may dene transfer function via exponential functions.

4.1. Basic denition and properties

In this section we start with a very general denition of a transfer function, which even
applies to systems not described by a p.d.e, but via e.g. a dierence dierential equation
or an integral equation. To formulate this denition, we rst have to introduce a general
system. In a general system, we have a time axis, T, which is assumed to be a subset of
R. Furthermore, we distinguish three spaces, U , Y , and R. U and Y are the input- and
output space, respectively, whereas R contains the rest of the variables. In our examples,
R will become the state space. A system S is a subset of (U R Y )T , i.e., a subset
of all functions from T to U R Y .
Denition 4.1.1. Let S be a system, let s be an element of C, and let u0 U . We say
that (u0 est , r(t), y(t)) is an exponential solution in S if there exist r0 R, y0 Y , such
that (u0 est , r0 est , y0 est ) = (u0 est , r(t), y(t)), t T.
If for every u0 U the output trajectory, y0 est , corresponding to an exponential
solution is unique, then we call the mapping u0 y0 the transfer function at s. We
denote this mapping by G(s). Let C be the set consisting of all s for which the
transfer function at s exists. The mapping s G(s) is dened as the transfer
function of the system S.

Recall that a system is said to be linear if U , R, and Y are linear spaces, and if
(u1 + u2 , r1 + r2 , y1 + y2 ) S whenever (u1 , r1 , y1 ) and (u2 , r2 , y2 ) are in S.
The system is time-invariant, when T is an interval of the form (t0 , ), t0 , and
(u( + ), r( + ), y( + )) is in S for all > 0, whenever (u, r, y) S.

62

4.1. Basic denition and properties

Lemma 4.1.2. Consider the abstract system S which is linear and time-invariant on
the time axis [0, ). If S is such that (0, 0, y(t)) S implies that y(t) is the zero
function, then exponential u and r implies exponential y, i.e., if (u0 est , r0 est , y(t)) S,
then y(t) = y0 est for some y0 Y .
Additionally, assume that for a given s C there exists an exponential solution for
all u0 U , and assume that (0, r0 est , y(t)) S implies that r0 = 0. Then the transfer
function at s exists, and is a linear mapping.
Proof: Let (u0 est r0 est , y(t)) be an element of S, and let 0. Combining the linearity
and time-invariance, we see that
(0, 0, y(t + ) es y(t)) = (u0 es(t+ ) , r0 es(t+ ) , y(t + )) (u0 es est , r0 es est , es y(t))
is an element of S. By assumption, this implies that y(t+ ) = es y(t) for all . Choosing
t = 0, gives the desired result.
Now assume that for a given s C there exists an exponential solution for all u0 U .
First we show that the exponential solution is unique. When (u0 est , r0 est , y0 est ) and
(u0 est , r0 est , y0 est ) are both in S, then by the linearity (0, (r0 r0 )est , (y0 y0 )est ) S.

By our assumption this implies that r0 = r0 and y0 = y0 .

From this we see that we can dene a mapping u0 y0 . It remains to show that this
mapping is linear. Let (u10 est , r10 est , y10 est ) and (u20 est , r20 est , y20 est ) be two exponential solutions. By the linearity, it is easy to see that
((u10 + u20 )est , (r10 + r20 )est , (y10 + y20 )est ) S
Hence this implies that u10 + u20 is mapped to y10 + y20 . In other words, the
mapping is linear.
In our applications r will be the state x. It turns out that for the class of systems
we are considering, the conditions in the above lemma are very weak, see Exercise 4.1.
Thus the transfer function exists, and is a linear operator.
We begin by showing that the transfer function for the system
x(t) = Ax(t) + Bu(t)

(4.6)

y(t) = Cx(t) + Du(t),

(4.7)

where B, C and D are bounded operators and A generates a strongly continuous semigroup, is given by the familiar formula G(s) = C(sI A)1 B + D, for s in the resolvent
set, (A), of A.
Theorem 4.1.3. Consider the state linear system (4.6)(4.7), with B, C, and D bounded
operators. As solutions of this system we take the mild solutions, see Theorem 3.4.1.
If (u(t), x(t), y(t)) is an exponential solution, then it is a classical solution as well.
Furthermore, for s (A), the transfer function exists and is given by
G(s) = C(sI A)1 B + D.

(4.8)

63

4. Transfer Functions
Proof: The mild solution of (4.6) is given by
t

x(t) = T (t)x0 +
0

T (t )Bu( )d

(4.9)

T (t )Bu0 es d.

(4.10)

For an exponential solution this becomes

t

x0 est = T (t)x0 +
0

By taking Laplace transforms, this equality becomes

1
x0
= (I A)1 x0 + (I A)1 Bu0
s
s

(4.11)

x0 ( s)(I A)1 x0 = (I A)1 Bu0 .

(4.12)

for C with Re() > max{Re(s), 0 }, where 0 is the growth bound of (T (t))t0 .
Since for xed the right-hand side of (4.11) is an element of D(A), the same holds
for the left-hand side. In other words, x0 D(A). By Theorem 3.1.2, we conclude that
x(t) is a classical solution.
Equation (4.11) can be rewritten as

Multiplying this by (I A), we obtain (sI A)x0 = Bu0 . Thus for s (A),
x0 = (sI A)1 Bu0 .

(4.13)

Using the output equation of our system, we have that

y0 est = y(t) = Cx(t) + Du(t) = Cx0 est + Du0 est = C(sI A)1 Bu0 est + Du0 est .
Hence y0 = C(sI A)1 Bu0 + Du0 which proves (4.8).
The above theorem shows that for state-space systems of the form (4.6)(4.7) the
transfer function exists, and is given by the formula well-known from nite-dimensional
system theory. Our main class of systems, the port-Hamiltonian systems have their
control and observation at the boundary, and are not of the form (4.6)(4.7). As proved
in Theorem 3.3.1, they form a subclass of the boundary control systems. In the following theorem, we formulate transfer functions for boundary control systems. As in the
previous theorem, exponential solutions are always classical solutions.
Theorem 4.1.4. Consider the system
x(t) = Ax(t),

u(t) = Bx(t)

x(0) = x0
(4.14)

y(t) = Cx(t)
where (A, B) satises the conditions for boundary control system, see Denition 3.2.1
and C is a bounded linear operator from D(A) to Y , with Y a Hilbert space.

64

4.1. Basic denition and properties

As the class of solutions we take the mild solutions, see equation (3.14).
If (u(t), x(t), y(t)) is an exponential solution, then it is a classical solution as well.
Furthermore, for s (A), the transfer function exists and is given by
G(s) = C(sI A)1 [AB sB] + CB

(4.15)

for s (A).
For s (A) and u0 U , G(s) can also be found as the (unique) solution of
sx0 = Ax0
u0 = Bx0

(4.16)

G(s)u0 = Cx0 ,
with x0 D(A).
Proof: The proof is very similar to that of Theorem 4.1.3. By (3.14) we know that the
mild solution is given by
t

x(t) = T (t)(x0 Bu(0)) +

T (t ) [ABu( ) B u( )] d + Bu(t).

Assuming that (u(t), x(t), y(t)) is an exponential solution, the above equation becomes
est x0 = T (t)(x0 Bu0 ) +

t
0

T (t ) [ABes u0 Bses u0 ] d + Best u0 .

Taking Laplace transforms, we nd for C with Re() > max{Re(s), 0 }, where 0

is the growth bound of the semigroup (T (t))t0 ,
su0
u0
u0
x0
= (I A)1 (x0 Bu0 ) + (I A)1 AB
B
.
+B
s
s
s
s
Or equivalently,
su0
x0 Bu0
u0
= (I A)1 (x0 Bu0 ) + (I A)1 AB
B
.
s
s
s
This implies that x0 Bu0 D(A), and
(I A)[x0 Bu0 ] = [x0 Bu0 ]( s) + ABu0 Bsu0 .
Subtracting the term [x0 Bu0 ] from both sides, this becomes
(sI A)[x0 Bu0 ] = ABu0 Bsu0 .

(4.17)

x0 = (sI A)1 [AB sB] u0 + Bu0 .

(4.18)

For s (A), this gives

65

4. Transfer Functions
Since x0 Bu0 D(A) and since u C 2 ([0, ); U ) we have by Theorem 3.2.2 that x(t)
is a classical solution of (4.14). In particular, for all t 0, x(t) D(A). By assumption
the domain of C contains the domain of A. Hence y0 est = y(t) = Cx(t), holds point-wise
in t. Taking t = 0, gives y0 = Cx0 . Combining this equality with (4.18), we obtain
(4.15).
Since x(t) = x0 est is a classical solution for u(t) = u0 est , we can substitute this in the
dierential equation (4.14). By doing so we nd
sx0 est = Ax0 est
u0 est = Bx0 est
y0 est = Cx0 est
Removing the exponential term, we nd equation (4.16). The uniqueness of x0 follows
by (4.18).
We close this section by calculating the transfer function for the simple Example 3.4.3.
Example 4.1.5 Consider the system
x
x
(, t) =
(, t),
t

x(, 0) = x0 (),
u(t) = x(1, t),
y(t) = x(0, t),

[0, 1], t 0

(4.19)

[0, 1]
t 0.

t 0.

If we dene Cx = x(0), then it is easy to see that all assumptions in Theorem 4.1.4 are
satised, see Theorem 3.3.1. Hence we can calculate the transfer function, we do this
via the equation (4.16). For the system (4.19) this becomes
dx0
x0
() =
()

d
= x0 (1)

sx0 () =
u0

G(s)u0 = x0 (0).
The above dierential equation has the solution x0 () = es . Using the other two
equations, we see that G(s) = es .

4.2. Transfer functions for port-Hamiltonian systems

In this section we apply the results found in the previous section to our class of portHamiltonian systems. Since we have obtained Theorem 4.1.4 describing transfer functions for general boundary control system, the application to port-Hamiltonian system is

66

4.2. Transfer functions for port-Hamiltonian systems

straightforward. We obtain the transfer function for the system dened in (3.32)(3.34).
That is, the system is given by

[Hx(t)] + P0 [Hx(t)]

f (t)
u(t) = WB
e (t)

(4.20)

x(t) = P1

y(t) = WC

(4.21)

f (t)
.
e (t)

(4.22)

It is assumed that (4.20)(4.21) satisfy the assumption of a port-Hamiltonian system, see

T
Section 3.3. Further we assume that WB WB 0, where = 0 I . This guarantees
I 0
that (4.20)(4.21) is a boundary control system, see Theorem 3.3.1.
As for the input we assume that we use full measurements, i.e., WC be a full rank
matrix of size n 2n and WB is of full rank, or equivalently this matrix is invertible.
WC
Theorem 4.2.1. Consider the system (4.20)(4.22). This system has the transfer function G(s), which is determined by
d
[Hx0 ] + P0 [Hx0 ]
d
f
= WB ,0
e,0

(4.23)

sx0 = P1
u0

G(s)u0 = WC

(4.24)

f,0
,
e,0

(4.25)

where
f,0
e,0

(Hx0 ) (b)
(Hx0 ) (a)

= R0

(4.26)

Furthermore, the transfer function satises the following equality

Re (s) x0

with PWB ,WC the inverse of

1
2
WB
WC

uT
0

uT G(s)
0

T
( WB

T
WC

PWB ,WC

u0
G(s)u0

(4.27)

).

Proof: The proof is a direct combination of Theorems 3.4.2 and 4.1.4. By the rst
theorem, we know that the system (4.20)(4.22) is a well-dened boundary control system and that the output equation is well-dened in the domain of the system operator
A. Hence all conditions of Theorem 4.1.4 are satised, and the dening relation for the
transfer function, equation (4.15), becomes (4.23)(4.25).
The transfer function is by denition related to the exponential solution (u0 est , x0 est ,
G(s)u0 est ). Substituting this solution in (3.36) gives (4.27).

67

4. Transfer Functions
Looking at (4.23)(4.25) we see that the calculation of the transfer function is equivalent to solving an ordinary dierential equation. If H is constant, i.e., independent of
, this is easy. However, in general it can be very hard to solve this ordinary dierential
equation by hand, see Exercise 4.4.
In the above theorem we assumed that we had n controls and n measurements. However, this needs not to hold. If the system has some of its boundary conditions set to
zero, and/or less than n measurements, then one can take two approaches for obtaining
the transfer function. As is shown in Exercise 3.6 this system satises all the conditions
of Theorem 4.1.4, and hence one can use that theorem to obtain the dierential equation determining the transfer function. Another approach is to regard the zero boundary
conditions as inputs set to zero, and to add extra measurements such that we have n
controls and n measurements. The transfer function one is looking for is now a sub-block
of the n n transfer function. We explain this in more detail by means of an example.
However, before we do this we remark that (4.27) equals the balance equation (1.27)
which lies at the hart of our class of systems.
Example 4.2.2 Consider the transmission line of Example 1.1.1 for which we assume
that we control the voltages at both ends, and measure the currents at the same points.
Furthermore, we assume that the spatial interval equals [0, 1]. Hence the model becomes
(, t)
Q
(, t) =
t
L()
Q(, t)

(, t) =
t
C()
u(t) =

y(t) =

Q(1,t)
C(1)
Q(0,t)
C(0)
(1,t)
L(1)
(0,t)
L(0)

(4.28)

(4.29)

(4.30)

For the transfer function, this p.d.e. is replaced by the ordinary dierential equation
d 0 ()
d L()
d Q0 ()
s0 () =
d C()

sQ0 () =

(4.31)

u0 =

u10
u20

Q0 (1)
C(1)
Q0 (0)
C(0)

y0 =

y10
y20

0 (1)
L(1)
0 (0)
L(0)

(4.32)

(4.33)

Since we want to illustrate transfer functions, and their properties, we make the simplifying assumption that C() = L() = 1 for all [0, 1]. With this assumption, it is

68

4.2. Transfer functions for port-Hamiltonian systems

easy to see that the solution of (4.31) is given by
Q0 () = es + es ,

0 () = es + es ,

(4.34)

where , are complex constants. Using (4.32) we can related these constant to u0 ,

1
s es
e

1 es
1
es

u0 .

(4.35)

Combining this with (4.33) gives

y0 =

es es
2
2
es + es

1
s es
e

u0 .

Thus the transfer function is given by

G(s) =

1
tanh(s) sinh(s)
1
tanh(s)
sinh(s)

(4.36)

Using now the balance equation (1.4), we nd

Re (s) x0

= Re (V0 (0)I0 (0)) Re (V0 (1)I0 (1))

= Re (u20 G21 (s)u10 + u20 G22 (s)u20 )

Re (u10 G11 (s)u10 + u10 G12 (s)u20 ).

(4.37)

By taking u10 = 0, we conclude that the real part of G22 is positive for Re(s) > 0.
Combined with the fact that G22 is analytic for s C+ := {s C | Re (s) > 0},
0
we have that G22 is positive real. This can also be checked by direct calculation on
G22 (s) = tanh(s).
Consider next the system dened by the p.d.e. (4.28) with input u(t) = Q(1, t), output
I(1, t) and boundary condition Q(0, t) = 0. We can proceed like we did above, but we
see that we have already obtained the transfer function by putting u20 = 0 in (4.32) and
only look at y10 in (4.33). Hence the transfer function of this single input single output
system is tanh(s).
The transfer functions (4.36) and tanh(s) have their poles on the imaginary axis,
and so one cannot draw a Bode or Nyquist plot. In order to show these concept known
from classical control theory can also be used for system described by p.d.e.s we add a
damping such that we obtain a system with no poles on the imaginary axis.
We consider the p.d.e. (4.28) with the following conditions
V (1, t) = RI(1, t)

(4.38)

u(t) = V (0, t)

(4.39)

y(t) = I(0, t).

(4.40)

69

4. Transfer Functions
Again we take the simplifying assumption that C() = L() = 1, [0, 1]. Calculation
the transfer function leads to the o.d.e. (4.31) with the boundary conditions
V0 (1) = RI0 (1)

(4.41)

u0 = V0 (0)

(4.42)

y0 = I0 (0).

(4.43)

The o.d.e. has as solution (4.34). The equations (4.41)(4.43) imply

es + es = R es + es
u0 = +

y0 = + .
Solving this equation gives the following transfer function,
G(s) =

cosh(s) + R sinh(s)
.
sinh(s) + R cosh(s)

(4.44)

The Nyquist plot of this is a perfect circle, see Figure 4.1 Again using the balance

Im G(i)

50

10

Re G(i)

Figure 4.1.: Nyquist plot of (4.44) for R = 10

equation (1.4), we nd that for this system
Re (s) x0

= Re (V0 (0)I0 (0)) Re (V0 (1)I0 (1))

= Re (u0 G(s)u0 ) Re (I0 (1)RI0 (1)).

Hence for Re(s) > 0, we have

Re (G(s))u2 = Re (s) x0
0
and so G is positive real.

70

+ RI0 (1)2 ,

(4.45)

4.3. Exercises

4.3. Exercises
4.1. Show that the assumption of Lemma 4.1.2 holds for the system (4.6)(4.7), whenever s (A).
4.2. Show that the assumption of Lemma 4.1.2 holds for the system (4.14), whenever
s (A).
4.3. Determine the transfer function of the system
x
x
(, t) =
(, t),
t

x(, 0) = x0 (),
u(t) = (b)x(b, t),
y(t) = (a)x(a, t),

[a, b], t 0

(4.46)

[0, 1]
t 0.

t 0,

where is a (strictly) positive continuous function not depending on t.

4.4. Consider the system of the transmission line given by (4.28)(4.30).
a) Show that even when the physical parameter C and L are spatial dependent,
the equality (4.37) still holds.
b) Choose L() = e , and C() = 1, and determine the transfer function.
Hint: You may use a computer package like Maple or Mathematica.
4.5. Our standard port-Hamiltonian system is dened on the spatial interval [a, b].
In Exercise 2.10 we have shown that it can easily be transformed to a portHamiltonian system on the spatial interval [0, 1]. How does the transfer function
change?
4.6. Show that the boundary control system (4.14) denes a linear and time-invariant
system, see below Denition 4.1.1.
4.7. Show that the Nyquist plot of transfer function (4.44) is a circle. Furthermore,
show that G restricted to the imaginary axis is periodic, and determine the period.
4.8. Consider the vibrating string of Example 1.1.2. We assume that the mass density
and Youngs modulus are constant. We control this system by controlling the
velocity at = b and the strain at = a, i.e., u(t) =
same quantities, but at the opposite ends, i.e., y(t) =

w
(t,b)
t
w
(t,a)

w
(t,a)
t
w
(t,b)

. We observe the
.

Determine the transfer function of this system.

4.9. Since we have dened transfer functions via a dierent way, it may be good to
check some well-known property of it. Let S1 and S2 be two linear and timeinvariant systems, with input-output pair u1 , y1 and u2 , y2 , respectively. Assume
that for a given s C both systems have a transfer function.

71

4. Transfer Functions
a) Show that the series connection, i.e., u2 = y2 has the transfer function G(s) =
G2 (s)G1 (s).
b) Show that the parallel connection, i.e., u1 = u2 = u, and y = y1 + y2 has the
transfer function G1 (s) + G2 (s).
c) Show that the feedback connection, i.e., u1 = u y2 , y = y1 has the transfer
function G1 (s) [I + G2 (s)]1 provided I + G2 (s) is invertible.
4.10. Consider the coupled strings of Exercise 3.8. As input we apply a force to the bar
in the middle, and as output we measure the velocity of this bar. Assuming that
all physical parameters are not depending on , determine the transfer function.

4.4. Notes and references

The ideas for dening the transfer function in the way we did is old, but has hardly
been investigated for distributed parameter system. [32] was the rst paper where this
approach has been used for innite-dimensional systems. In that paper the concept we
named transfer function was called a characteristic function.
One may nd the exponential solution in Polderman and Willems [20], where all
solutions of this type are called the exponential behavior.

72

Chapter 5

Well-posedness

5.1. Introduction
Consider the abstract linear dierential equation
x(t) = Ax(t) + Bu(t),

x(0) = x0

y(t) = Cx(t) + Du(t).

(5.1)
(5.2)

where x is assumed to take values in the Hilbert space X, u in the Hilbert space U , and
y is the Hilbert space Y . The operator A is assumed to be the innitesimal generator
of a C0 -semigroup, and B, C, and D are bounded linear operators. Under these assumptions we know that the abstract dierential equation (5.1)(5.2) possesses for every
u L2 ((0, tf ); U ) a unique (mild) solution, see Theorem 3.4.1. Existence of solutions
for an arbitrary initial condition x0 X and input u L2 ((0, tf ); U ), such that x is
continuous and y L2 ((0, tf ); Y ) will be called well-posedness. Hence if B, C, and D
are bounded linear operators, then the system (5.1)(5.2) is well-posed if and only if A
is the innitesimal generator of a C0 -semigroup.
As we have seen in Chapter 3 our class of port-Hamiltonian systems cannot be written
in the format (5.1)(5.2) with B, C and D bounded. However, we know that for every
initial condition and every smooth input function we have a mild solution of the state
dierential equation, see (3.14). Furthermore, for smooth initial conditions and smooth
inputs, the output equation is well-dened, see Theorem 3.4.2. Note that this was only
T
obtained under the condition WB WB 0. This inequality is equivalent to the fact
that we have a contraction semigroup, see Theorem 2.3.2. That a larger class of inputs
might be possible, can be seen in the following example.
Consider the controlled transport equation on the interval [0, 1] with scalar control
and observation on the boundary
x
x
(, t) =
(, t),
t

u(t) = x(1, t),

y(t) = x(0, t).

x(, 0) = x0 (),

[0, 1]

(5.3)
(5.4)
(5.5)

73

5. Well-posedness
From Example 3.2.3 we know that the mild solution of (5.3)(5.4) is given by
x(, t) =

x0 ( + t)
u( + t 1)

+t1
+ t > 1.

(5.6)

We see that for every t 0 the function x(, t) X = L2 (0, 1), whenever u L2 (0, tf ).
Furthermore, x(, t) is a continuous function in t, i.e., x(, t) x((, t + h) converges to
zero when h converges to zero, see Exercise 5.1. Hence the mild solution (5.6) can be
extended from controls in H 1 (0, tf ) to L2 (0, tf ). If x0 and u are smooth, then we clearly
see that y(t) is well-dened for every t 0 and it is given by
y(t) =

x0 (t)
u(t 1)

0t1
t > 1.

(5.7)

However, when x0 L2 (0, 1) and u L2 (0, tf ), the expression (5.7) still gives that y is
well-dened as an L2 -function.
Summarizing, we see that we can dene a (mild) solution for (5.3)(5.5) for all x0 X
and all u L2 (0, tf ). This solution gives a state trajectory in the state space which is
continuous, and an output trajectory which is square integrable on every compact time
interval. Hence this system is well-posed.
Suppose next that we are applying a feedback of the form u(t) = 2y(t), this leads to
the new p.d.e.
x
x
(, t) =
(, t),
t

x(1, t) = 2x(0, t).

x(, 0) = x0 (),

[0, 1]

(5.8)
(5.9)

We would like to know whether the system has a unique (mild) solution, and so we try
to apply Theorem 2.3.2. From Example 2.3.3 we have that the boundary eort and ow
are given by
1
1
e = [x(1) x(0)].
f = [x(1) + x(0)],
2
2
Hence the boundary condition (5.9) can be written as f 3e = 0, i.e., WB = (1, 3).
T
Calculation WB WB gives 6. Since this is negative, we know by Theorem 2.3.2)
that the operator associated to the p.d.e. (5.8)(5.9) does not generate a contraction
semigroup. However, it does generate a C0 -semigroup. It is not hard to see that the
solution of (5.8)(5.9) is given by
x(, t) = 2n+1 x0 ( ),

(5.10)

where + t = n + , [0, 1], n N, see Exercise 5.2.

Concluding, we see from this very simple port-Hamiltonian system that the class of
controls can be extended, and that it is possible to have a semigroup, even when the
T
condition WB WB 0 is not satised. The aim of this chapter is to show that these
results hold for any port-Hamiltonian system in our class.

74

5.2. Well-posedness for port-Hamiltonian systems

5.2. Well-posedness for port-Hamiltonian systems

In this section we dene formally what we mean by well-posedness. Although wellposedness can be dened very generally, see the Notes and reference section, we do it
for our class of systems introduced in Sections 3.3 and 3.4.

x
(, t) = P1 [Hx(t)] + P0 [Hx(t)]
t

f (t)
u(t) = WB,1
e (t)
0 = WB,2
y(t) = WC

(5.11)
(5.12)

f (t)
e (t)

(5.13)

f (t)
.
e (t)

(5.14)

The precise assumptions on this system will be given in Theorem 5.2.6. However, as
before, we assume that H is a symmetric n n matrix satisfying 0 < mI H() M I.
WB,1
Furthermore, WB := WB,2 is a full rank matrix of size n 2n. We assume that WB,1

is a k 2n matrix. Our state space is the Hilbert space X = L2 ((a, b); Rn ) with the
inner product
1 b
f ()T H()g()d .
(5.15)
f, g X =
2 a
The following lemma follows easily from Theorems 3.3.1 and 3.4.2. The proof is left
as an exercise, see Exercise 5.3.
Lemma 5.2.1. Let tf be a positive real number. Assume that the operator A dened

as P1 H + P0 H with domain,
D(A) = {x0 X | Hx0 H 1 ((a, b); Rn ), WB

f,0
e,0

= 0}

is the innitesimal generator of a C0 -semigroup on X. Then the system (5.11)(5.13) is

a boundary control system.
In particular, for every Hx0 H 1 ((a, b); Rn ) and every u() C 2 ([0, tf ); Rk ) with
f,0
,0
u(0) = WB,1 f,0 , and 0 = WB,2 e,0 , there exists a unique classical solution of
e
(5.11)(5.13) on [0, tf ]. Furthermore, the output (5.14) is well-dened and y() is continuous on [0, tf ].

This lemma tells us that under the existence assumption of a semigroup, we have a
solution for smooth inputs and initial conditions. The concept of well-posedness implies
that we have solutions for every initial condition and every square integrable input.
Denition 5.2.2. Consider the system (5.11)(5.14) and let k be the dimension of u.
This system is well-posed if there exists a tf > 0 and mf 0 such that the following
holds:

75

5. Well-posedness

1. The operator A dened as P1 H + P0 H with domain,

D(A) = {x0 X | Hx0 H 1 ((a, b); Rn ), WB

f,0
e,0

= 0}

(5.16)

is the innitesimal generator of a C0 -semigroup on X.

2. The following inequality holds for all Hx0 H 1 ((a, b); Rn ) and u C 2 ([0, tf ); Rk )
f,0
e,0

with u(0) = WB,1

x(tf )

2
X

f,0
e,0

, and 0 = WB,2
tf

+
0

y(t) 2 dt mf

x0

2
X

tf

u(t) 2 dt .

(5.17)

If the system is well-posed, then we can dene a (mild) solution of (5.11)(5.14) for
all x0 X and u L2 ((0, tf ), Rk ) such that x(t) is a continuous function in X and y is
square integrable.
Theorem 5.2.3. If the system (5.11)(5.14) is well-posed, then the mild solution (3.14)
can be extended to hold for all u L2 ((0, tf ); Rk ). Furthermore, the state trajectory
remains continuous, and the output is square integrable.
For the (extended) solution, the inequality (5.17) holds.
Furthermore, if the system is well-posed for some tf > 0, then it is well-posed for all
tf > 0.
Proof: See Section 5.6.
As we have seen in Chapter 4, the system (5.11)(5.14) has a transfer function. From
Theorem 4.1.4 we conclude that this function exists in the resolvent set of A. Since A
generates a C0 -semigroup, this resolvent set contains a right half-plane. So the transfer
function exists on some right-half plane. Furthermore, the transfer function is bounded
for Re(s) .

Lemma 5.2.4. Let G be the transfer function of a well-posed system, then

lim sup G(s) mf ,

(5.18)

Re(s)

where mf is the constant from equation (5.17).

Proof: If (5.18) does not hold, then we can nd an s such that |e2stf | > mf and

G(s) > mf . Next choose u0 U such that u0 = 1 and G(s)u0 = G(s) . The
exponential solution (u0 est , x0 est , G(s)u0 est ) satises
x(tf )

2
X

tf

y(t) 2 dt =

x(tf )

2
X

= |e2stf | x0
> mf

x0

This contradicts (5.17), and so (5.18) must hold.

76

tf

G(s)u0 est 2 dt

0
2
X
2
X

+ G(s)

tf

est u0 2 dt

0
tf

+
0

u(t) 2 dt .

5.2. Well-posedness for port-Hamiltonian systems

Although the transfer function is bounded on some right half-plane, this does not
imply that it will converge along the real axis. If this happens, we call the transfer
function regular.
Denition 5.2.5. Let G(s) be the transfer function of (5.11)(5.14). The system
(5.11)(5.14) is regular when limsR,s G(s) exists. If the system (5.11)(5.14) is
regular, then the feed-through term D is dened as D = limsR,s G(s).

Now we have all the ingredients to formulate the main result of this chapter.
Theorem 5.2.6. Consider the partial dierential equation (5.11)(5.14) on the spatial
interval [a, b], with x(, t) taking values in Rn . Let X be the Hilbert space L2 ((a, b); Rn )
with inner product (5.15). Furthermore, assume that
T
P1 is real-valued, invertible, and symmetric, i.e., P1 = P1 ,

H() is a (real) symmetric matrix satisfying 0 < mI H() M I, [a, b].

The multiplication operator P1 H can be written as
P1 H() = S 1 ()()S(),

(5.19)

with () a diagonal multiplication operator, and both () and S() are continuously dierentiable,
WB :=
rank

WB,1
WB,2
WB,1
WB,2
WC

is a n 2n matrix with rank n,

= n + rank (WC ).

If the homogeneous p.d.e., i.e., u 0, generates a C0 -semigroup on X, then the system (5.11)(5.14) is well-posed, and the corresponding transfer function G is regular.
Furthermore, we have that limRe(s) G(s) = lims,sR G(s).
This theorem tells us that if A dened in Lemma 5.2.1 generates a C0 -semigroup, then
the system is well-posed. In particular, we have a mild solution for all square integrable
T
inputs. From Chapter 2 we know that if WB WB 0, then A generates a (contraction)
semigroup, and so in this situation the system is well-posed.
In the coming sections we prove this result. Here we comment on the conditions.
The rst two conditions are very standard, and are assumed to be satised for all
our port-Hamiltonian systems until now.
Note that we do not have a condition on P0 . In fact the term P0 H may be replaced
by any bounded operator on X, see Lemma 5.4.1.
The third condition is not very strong, and will almost always be satised if H() is
continuously dierentiable. Note that contains the eigenvalues of P1 H, whereas
S 1 contains the eigenvectors.

77

5. Well-posedness
The fourth condition tells us that we have n boundary conditions, when we put
the input to zero. This very logical, since we have an nth order p.d.e.
The last condition tells that we are not measuring quantities that are set to zero,
or set to be an input. This condition is not important for the proof, and will
normally follow from correct modeling.
As mentioned, the third condition is not very strong. We prove some properties of
next.
Lemma 5.2.7. Let P1 and H satisfy the conditions of Theorem 5.2.6. Then can be
written as
()
0
() =
,
(5.20)
0
()
where () is a diagonal (real) matrix, with (strictly) positive functions on the diagonal,
and () is a diagonal (real) matrix, with (strictly) negative functions on the diagonal.
Proof: Let [a, b] be xed. Since H() > mI, we can take the square root of it. By
1
1
the law of inertia, we know that the inertia of H() 2 P1 H() 2 equals the inertia of P1 .
1
1
This implies that the inertia of H(z) 2 P1 H() 2 is independent of . Furthermore, since
1
1
P1 is invertible, we conclude that the number of negative eigenvalues of H(z) 2 P1 H() 2
equals the number of negative eigenvalues of P1 . A similar statement holds for the
positive eigenvalues.
1
1
A simple calculation gives that the eigenvalues of H() 2 P1 H() 2 are equal to the
eigenvalues of P1 H(). Concluding, we see that for all [a, b] zero is not an eigenvalue
of P1 H(), and that the number of negative and positive eigenvalues of P1 H() is independent of . We can regroup the eigenvalues such that rst are positive. By doing so,
we obtain (5.20).
We illustrate the conditions in the theorem by proving that they are easily satised
for the example of the wave equation. From Example 1.1.2 together with (1.20) we know
that the model of the wave equation written in the port-Hamiltonian form is given by

x1 (, t)
x2 (, t)

1
()

0 1
1 0

where x1 = w is the momentum and x2 =

t
Hence we have that
P1 =

0 1
1 0

0
T ()

x1 (, t)
x2 (, t)

(5.21)

is the strain.

H() =

1
()

0
T ()

Being physical constants, the Youngs modulus T and the mass density are positive.
Hence it is clear that P1 and H satisfy the rst two conditions of Theorem 5.2.6. We
shall show that (5.19) holds.
P1 H() =

78

0
1
()

T ()
0

(5.22)

5.3. The operator P1 H is diagonal.

satises the third condition, i.e., relation (5.19). The eigenvalues of this matrix are
with () = T () . The corresponding eigenvectors are
()
()

and

1
()

()
1
()

Hence

P1 H = S 1 S =

0
0

(5.23)

1
2
1
2

(5.24)

where we have omitted the dependence on . The idea of the proof is the following.
Since S is invertible, well-posedness will not change if we perform a basis transformation,
x = Sx. After this basis transformation, the p.d.e. becomes

dS 1 ()
(, t) = () (, t) + S()
x
()(, t)
x
t

dS 1 ()
()1 (, t)
x
=
()(, t).
x
+ S()
()2 (, t)
x

(5.25)

We see that we have a very nice set of simple p.d.e.s, just two simple delay line, but
1
they are corrupted by the term S() dS d () ()(, t). We rst assume that this term
x
is not present, and so we study the well-posedness of the collection of delay lines

x1 (, t)

x2 (, t)

()1 (, t)
x
()2 (, t)
x

(5.26)

Although it seems that these p.d.e.s are uncoupled, they are coupled via the boundary
conditions. In Section 5.3 we investigate when a p.d.e. like the one given in (5.26) with
control and observation at the boundary is well-posed. In Section 5.4 we return to the
original p.d.e., and show that ignoring bonded terms, like we did in (5.25) and (5.26)
does not inuence the well-posedness of the system. Since a basis transformation does
not eect it either, we have proved Theorem 5.2.6.

5.3. The operator P1 H is diagonal.

In this section, we prove Theorem 5.2.6 if P1 H is diagonal, i.e., when S = I. For this
we need the following two lemmas.
Lemma 5.3.1. Let () be a positive continuous dierentiable function on the interval
[a, b]. With this function we dene the scalar system

w
(, t) =
(()w(, t)) ,
t

w(, 0) = w0 ()

[a, b]

(5.27)

The value at b we choose as input

u(t) = (b)w(b, t)

(5.28)

79

5. Well-posedness
and as output we choose the value on the other end
y(t) = (a)w(a, t).

(5.29)

The system (5.27)(5.29) is a well-posed system on the state space L2 (a, b). Its transfer
function is given by
G(s) = ep(b)s ,
(5.30)
where p is dened as

p() =

()1 d

[a, b].

(5.31)

This transfer function satises

lim

Re(s)

G(s) = 0.

(5.32)

Proof: It is easy to see that the system (5.27)(5.29) is a very simple version of the

general Port-Hamiltonian system (3.32)(3.34) with P1 = 1, H = , WB = ( 22 , 22 ) and

T
WC = ( 22 , 22 ), see Exercise 5.4. Since WB WB = 1 > 0, we conclude by Theorem
3.4.2 that (5.27)(5.29) has a well-dened solution provided the initial condition and the
input are smooth. For this class, the balance equation (3.36) holds. However, this is the
same as (1.27). Using that form of the power balance, we obtain

d
dt

b
a

w(, t)()w(, t)d = [[()w(, t)] ()w(, t)]b

a
= |u(t)|2 |y(t)|2 ,

where we used (5.28) and (5.29). Thus for all tf > 0 we have that
b

w(, 0)()w(, 0)d

w(, tf )()w(, tf )d
a

tf

=
0

|u( )|2 d

tf
0

|y( )|2 d.

(5.33)

Since is strictly positive, we have that the energy norm a w(, t)()w(, t)d is
equivalent to the L2 (a, b)-norm, and so on a dense set an inequality like (5.17) is satised.
Thus the system is well-posed.
As we have seen in Chapter 4, the transfer function G(s) is constructed by nding
for s C and for all u0 a triple (us (t), ws (, t), y(t)) = (u0 est , w0 ()est , y0 est ) satisfying
(5.27)(5.29). Substituting a triple of this form in the p.d.e., gives
sw0 () =

(()w0 ()) ,

u0 = (b)w0 (b),

y0 = (a)w0 (a).

Thus w0 () = u0 ()1 exp(s(p()p(b))), and y0 = u0 exp(sp(b)). This proves (5.30).

The property (5.32) follows directly from (5.30) and the fact that p(b) > 0.

80

5.3. The operator P1 H is diagonal.

For a p.d.e. with negative coecient, we obtain a similar result.
Lemma 5.3.2. Let () be a negative continuous function on the interval [a, b]. With
this function we dene the scalar system

w
(, t) =
(()w(, t)) ,
t
z

w(, 0) = w0 ()

[a, b]

(5.34)

The value at a we choose as input

u(t) = (a)w(a, t)

(5.35)

and as output we choose the value on the other end

y(t) = (b)w(b, t).

(5.36)

The system (5.34)(5.36) is a well-posed system on the state space L2 (a, b). Its transfer
function is given by
G(s) = en(b)s ,
(5.37)
where

n() =

()1 d,

[a, b].

(5.38)

This transfer function satises

lim

Re(s)

G(s) = 0.

(5.39)

We use these two lemmas to prove Theorem 5.2.6 when P1 H is diagonal and the input
space has dimension n.
Consider the following diagonal hyperbolic system on the spatial interval [a, b]

x+ (, t)
x (, t)

()
0
0
()

x+ (, t)
x (, t)

(5.40)

where () is a diagonal (real) matrix, with positive functions on the diagonal, and ()
is a diagonal (real) matrix, with negative functions on the diagonal. Furthermore, we
assume that and are continuously dierentiable.
With this p.d.e. we associate the following boundary control and observation
us (t) :=

(b)x+ (b, t)
(a)x (a, t)

(5.41)

ys (t) :=

(a)x+ (a, t)
(b)x (b, t)

(5.42)

Theorem 5.3.3. Consider the p.d.e. (5.40) with us and ys as dened in (5.41) and
(5.42), respectively.

81

5. Well-posedness
The system dened by (5.40)(5.42) is well-posed and regular. Furthermore, its
transfer function converges to zero for Re(s) .
To the p.d.e. (5.40) we dene a new set of inputs and outputs. The new input u(t)
is written as
u(t) = Kus (t) + Qys (t),
(5.43)
where K and Q are two square matrices, with [K, Q] of rank n. The new output
is written as
y(t) = O1 us (t) + O2 ys (t).
(5.44)
where O1 and O2 are some matrices. For the system (5.40) with input u(t) and
output y(t), we have the following possibilities:
1. If K is invertible, then the system (5.40), (5.43), and (5.44) is well-posed and
regular. Furthermore, its transfer function converges to O1 K 1 for Re(s)

2. If K is not invertible, then the operator AK dened as

AK

g+ ()
g ()

()
0
0
()

g+ ()
g ()

(5.45)

with domain
D(AK ) =

g+ ()
g ()

H 1 ((a, b), Rn ) |
K

(b)g+ (b)
(a)g (a)

+Q

(a)g+ (a)
(b)g (b)

=0

(5.46)

does not generate a C0 -semigroup on L2 ((a, b); Rn ).

Note that the last item implies that the homogeneous p.d.e. does not have a well-dened
solution, when K is not invertible.
Proof: The rst item is a direct consequence of Lemma 5.3.1 and 5.3.2 by noticing
that the system (5.40)(5.42) is built out of copies of the system (5.27)(5.29) and the
system (5.34)(5.36). Furthermore, these sub-systems do not interact with each other.
For the proof of the rst part of the second assertion, with K invertible, we rewrite
the new input, as us (t) = K 1 u(t) K 1 Qys (t). This can be seen as a feedback
interconnection on the system (5.40)(5.42), as is depicted in Figure 5.1. The system
contains one feedback loop with gain matrix K 1 Q. By Theorem 5.6.1, we have that if
I + Gs (s)K 1 Q is invertible on some right half-plane and if this inverse is bounded on a
right-half plane, then the closed loop system is well-posed. Since limRe(s) Gs (s) = 0,
we see that this holds for every K 1 and Q. So under the assumption that K is invertible,
we nd that (5.40) with input and output given by (5.43) and (5.44) is well-posed. The
regularity follows easily. By regarding the loops in Figure 5.1, we see that the feed-though
term is O1 K 1 .

82

5.3. The operator P1 H is diagonal.

y(t)

O1

u(t)

K 1

ys (t)

Gs (s)

us (t)

O2

Figure 5.1.: The system (5.40) with input (5.43) and output (5.44)
So it remains to show that there is no C0 -semigroup when K is non-invertible. Since
K is singular, there exists a non-zero v Rn such that v T K = 0. Since [K, Q] has full
rank, we know that q T := v T Q = 0. So at least one of the components of q is unequal
to zero. For the sake of the argument, we assume that this holds for the rst one.
If AK would be the innitesimal generator of a C0 -semigroup, then for all x0 D(AK )
the abstract dierential equation
x(t) = AK x(t),

x(0) = x0

(5.47)

would have classical solution, i.e., for all t > 0, x(t) is dierentiable, it is an element of
D(AK ), and it satises (5.47). Hence by (5.46), we have that x(t) is an element of H 1 .
Since we are working in a one dimensional spatial domain, we have that functions in H 1
are continuous. So we have that for every t, x(t) is a continuous function of satisfying
the boundary conditions in (5.46).
So if AK would generate a C0 -semigroup, then for every x0 D(AK ) there would be a
x+ (,t)
function x(, t) := x (,t) which is a (mild) solution to the p.d.e. (5.40), and satises
for all t > 0 the boundary condition
K

(b)x+ (b, t)
(a)x (a, t)

+Q

(a)x+ (a, t)
(b)x (b, t)

= 0.

Using the vectors v and q, we see that this x(, t) must satisfy
0 = qT

(a)x+ (a, t)
(b)x (b, t)

t > 0.

(5.48)

Now we construct an initial condition in D(AK ), for which this equality does not hold.
Note that we have chosen the rst component of q unequal to zero.
The initial condition x0 is chosen to have all components zero except for the rst one.
For this rst component we choose an arbitrary function in H 1 (a, b) which is zero at a
and b, but nonzero everywhere else on the open set (a, b). It is clear that this initial
condition is in the domain of AK . Now we solve (5.40).
Standard p.d.e. theory gives that the solution of (5.40) can be written as
x+,m (, t) = f+,m(pm () + t)m ()1 ,
1

x, (, t) = f, (n () + t) ()

(5.49)
(5.50)

83

5. Well-posedness
where m and are the m-th and the -th diagonal element of and , respectively.

Furthermore, pm () = a m ()1 d, nl () = a l ()1 d, see also Exercises 5.4 and

5.5. The functions f+ , f need to be determined from the boundary and initial conditions.
Using the initial condition we have that f+,m (pm ()) = m ()x0,+,m () and f, (n ()) =
()x0,, (). Since pm > 0, and n < 0, we see that the initial condition determines f+
on a (small) positive interval, and f on a small negative interval. By our choice of the
initial condition, we nd that
[0, p1 (b)),

f+,1 () = 1 ()x0,+,1 ()

[0, pm (b)),

f+,m () = 0

[n (b), 0),

f, () = 0

m 2,

(5.51)

1.

The solution x(, t) must also satisfy (5.48), thus for all t > 0 we have that
0 = qT

f+ (t)
f (n(b) + t)

(5.52)

Combining this with (5.51), we nd

0 = q1 f+,1 (p1 ()) = q1 x0,+,1 ()1 ()
1
on some interval [a, ]. Since q1 and 1 are unequal to zero, we nd that x0 must be
zero on some interval. This is in contradiction with our choice of the initial condition.
Thus AK cannot be the innitesimal generator of a C0 -semigroup.

5.4. Proof of Theorem 5.2.6.

In this section we use the results of the previous section to prove Theorem 5.2.6. We
begin with a useful lemma, which proof can be found in Section 5.6.
Lemma 5.4.1. The system (5.11)(5.14) is well-posed if and only if the system
x

(, t) = P1
(Hx) (, t),
t

(5.53)

with inputs, outputs given by (5.12)(5.14) is well-posed.

Let G(s) denote the transfer function of (5.11)(5.14) and G0 (s) the transfer function
of (5.53) with (5.12)(5.14). Then
lim G(s) = lim G0 (s),

(5.54)

and
lim

Re(s)

84

G(s) =

lim

Re(s)

G0 (s),

(5.55)

5.4. Proof of Theorem 5.2.6.

This lemma tells us that we may ignore any bounded linear term involving x.
Now we have all the results needed to prove Theorem 5.2.6.
By the third assumption of Theorem 5.2.6, the matrices P1 and H satisfy the equation
(5.19):
P1 H() = S 1 ()()S().
With this we introduce the new state vector
x(, t) = S()x(, t),

[a, b]

(5.56)

Under this basis transformation, the p.d.e. (5.11) becomes

dS 1 ()
x

(, t) = () (, t) + S()
x
()(, t)+
x
t

d
S()P0 ()S()1 x(, t),

x(, 0) = S()x0 () = x0 ().

(5.57)

The relations (5.12)(5.14) become

1
1
0 =M11 P1 S 1 (b)(b)(b, t) + M12 P1 S 1 (a)(a)(a, t)
x
x

=M11 (b)(b, t) + M12 (a)(a, t)

x
x

(5.58)

1
1
u(t) =M21 P1 S 1 (b)(b)(b, t) + M22 P1 S 1 (a)(a)(a, t)
x
x
21 (b)(b, t) + M22 (a)(a, t)

=M
x
x

(5.59)

y(t)

1
=C1 P1 S 1 (b)(b)(b, t)
x

+
1 (b)(b, t) + C2 (a)(a, t).

=C
x
x

We introduce M =

Mj1 Mj2

M11 M12

M21 M22

Mj1 Mj2

C1 C2

(5.60)

with

and

C=

1
C2 P1 S 1 (a)(a)(a, t)
x

C1 C2

1
P1 S(b)1
0
1
0
P1 S(a)1

1
0
P1 S(b)1
1
0
P1 S(a)1

j = 1, 2.

1
P1 S(b)1
0
1
0
P1 S(a)1

has full rank, we see that the rank conditions in

Theorem 5.2.6 imply similar rank conditions for M and C.

Using Lemma 5.4.1 we see that we only have to prove the result for the p.d.e.

Since the matrix

(, t) =
() (, t).
x
t

(5.61)

with boundary conditions, inputs, and outputs as described in (5.58)(5.60).

It is clear that if condition (5.58) is not present, then Theorem 5.3.3 gives that the
above system is well-posed and regular if and only if the homogeneous p.d.e. generates a
C0 -semigroup on L2 ((a, b); Rn ). Since the state transformation (5.56) denes a bounded

85

5. Well-posedness
mapping on L2 ((a, b); Rn ), we have proved Theorem 5.2.6 provided there is no condition
(5.12).
Thus it remains to prove Theorem 5.2.6 if we have put part of the boundary conditions
to zero. Or equivalently, to prove that the system (5.58)(5.61) is well-posed and regular
if and only if the homogeneous p.d.e. generates a C0 -semigroup.
We replace (5.58) by

v(t) = M11 (b)(b, t) + M12 (a)(a, t),

x
x

(5.62)

where we regard v as a new input. Hence we have the system (5.61) with the new
extended input
v(t)
u(t)

M11

M21

(b)(t, b) +
x

M12

M22

(a)(t, a).
x

(5.63)

and the output (5.60). By doing so, we have obtained a system without a condition
(5.58). For this system we know that it is well-posed and regular if and only if the
homogeneous equation generates a C0 -semigroup.
Assume that the system (5.61), (5.63) and (5.60) is well-posed, then we may choose
any (locally) square input. In particular, we may choose v 0. Thus the system
(5.57)(5.61) is well-posed and regular as well.
Assume next that the p.d.e. with the extended input in (5.63) set to zero, does not
generate a C0 -semigroup. Since this gives the same homogeneous p.d.e. as (5.61) with
(5.58) and u in (5.59) set to zero, we know that this p.d.e. does not generate a C0 semigroup. This nally proves Theorem 5.2.6.

5.5. Well-posedness of the vibrating string.

In this section we illustrate the usefulness of Theorem 5.2.6 by applying it to the vibrating
string of Example 1.1.2.
By equation (1.20) we know that for the vibrating string there holds
P1 =

0 1
1 0

H() =

1
()

0
T ()

Since we want to illustrate the theory and proofs derived in the previous sections, we
do not directly check if for a (to-be-given) set of boundary conditions the semigroup
condition is satised. Instead of that, we rewrite the system in its diagonal form, and
check the conditions using Theorem 5.3.3. As we have seen in Section 5.4, the proof of
Theorem 5.2.6 follows after a basis transformation directly from Theorem 5.3.3. Hence
we start by diagonalizing P1 H. Although all the results hardly change, we assume for
simplicity of notation that Youngs modulus T and the mass density are constant.
Being physical constants, they are naturally positive.
From equation (5.24 we know that the operator P1 H is diagonalizable:
P1 H = S 1 S =

86

0
0

1
2
1
2

(5.64)

5.5. Well-posedness of the vibrating string.

where is positive and satises 2 = c = T .

Hence the state transformation under which the p.d.e. becomes diagonal is
x=

1
2

x.

Since we assumed that > 0, we see that x1 , x2 correspond to x+ , and x in equation

(5.40), respectively and , to and , respectively. Hence we have that the input and
output us and ys dened for the diagonal system (5.40) by the equations (5.41)(5.42)
are expressed in the original coordinates by
1
2
1
ys (t) =
2

x1 (b, t) + x2 (b, t)
x1 (a, t) x2 (a, t)

us (t) =

,
.

x1 (a, t) + x2 (a, t)
x1 (b, t) x2 (b, t)

(5.65)
(5.66)

This pair of boundary input and output variables consists in complementary linear combinations of the momentum x1 and the strain x2 at the boundaries: however they lack
an obvious physical interpretation. One could consider another choice of boundary input
and outputs, for instance the velocity and the strain at the boundary points and choose
as input
x1
(b, t)
(5.67)
u1 (t) =
x2 (a, t)
and as output

x1
(a, t)

y1 (t) =

x2 (b, t)

(5.68)

We may apply Theorem 5.3.3 to check whether this system is well-posed, and to nd
the feed-through. Expressing the input-output pair (u1 , y1 ) in (us , ys ) gives
u1 (t) =
y1 (t) =
Hence
K=

O1 =

0
1

0
0

1
T

us (t) +
us (t) +

Q=

1
T

O2 =

0
1
T
1

ys (t),

0
1

0
0

1
T
1

0
0
1

(5.69)

ys (t).

(5.70)

(5.71)

(5.72)

Since K is invertible, the system with the input-output pair (u1 , y1 ) is well-posed and
0
regular, and the feed-through term is given by O1 K 1 = 1 0 .

87

5. Well-posedness
w
, the
w
(t,b)
t

Since the states are dened as x1 = w and x2 =

t
are easily formulated using w. Namely, u1 (t) =

w
(t,a)

control and observation

and y1 (t) =

w
(t,a)
t
w
(t,b)
z

respectively. Hence we observe the velocity and strain at opposite ends.

Next we show that if we would control the velocity and strain at the same end, this
does not give a well-posed system. The control and observation are given by
w
t (b, t)
w
(b, t)

u2 (t) =

x1
(b, t)

(5.73)

x2 (b, t)

and as output
w
t (a, t)
w
(a, t)

y2 (t) =

x1
(a, t)

x2 (a, t)

(5.74)

ys (t).

(5.75)

It is easy to see that

u2 (t) =

0
0

1
0

1
0

us (t) +

Clearly the matrix in front of us is not invertible, and hence we conclude by Theorem
5.3.3 that the wave equation with the homogeneous boundary conditions u2 = 0 does
not generate a C0 -semigroup. Hence this system is not well-posed.
Until now we have been controlling velocity and strain at the end points. However,
for the wave equation, it seems very naturally to control the position, i.e., w(, t). So we
consider the wave equation (1.5) with the following control and observation.
u3 (t) =
y3 (t) =

w(b, t)

(5.76)

w
(a, t)

w(a, t)
w
z (b, t)

(5.77)

Since the rst control and rst observation cannot be written as linear combination of
our boundary eort and ow, we nd that this system is not of the form (5.11)(5.14).
However, we still can investigate the well-posedness of the system. For this we realize
that the rst element in u3 is the time derivative of the rst element of u1 . So we can
see the wave equation with the input (5.76) and output (5.77) as the following series
connection.
u3 (t)

s0
(0 1)

u1 (t)

G1 (s)

y1 (t)

s1 0
0 1

y3 (t)

Figure 5.2.: The wave equation with input and output (5.76) and (5.77)

88

5.6. Technical lemmas

From this it is clear that the transfer function, G3 (s) of the system with input u3 and
output y3 is given by
s1 0
0 1

G3 (s) =

G1 (s)

s 0
0 1

.
0

Since for large real s the transfer function G1 (s) is approximately equal to 1 0 ,
we see that G3 (s) grows like s for large s. By Lemma 5.2.4 we know that any well-posed
system has a transfer function which is bounded in some right half-plane. Thus the wave
equation with input (5.76) and output (5.77) is not well-posed.
The reason for the well-posedness of this system is dierent than for the choice u2
and y2 . Since if we put u3 to zero, then this implies that u1 is zero as well, and so we
know that this homogeneous equation is well-dened and has a unique solution. So if
the controls and/or observations are not formulated in the boundary eort and ow,
then we may loose well-posedness even if there is a semigroup.

5.6. Technical lemmas

In this section we present the proofs for the technical results which we needed. We begin
by considering the boundary control system
The following result is essential in our proof, and it has been proved by G. Weiss in
[30].
v(t)

u(t)
S

y(t)

Figure 5.3.: The closed loop system

Theorem 5.6.1. Let S be a well-posed system with input space U and output space Y ,
both being Hilbert spaces. Denote the transfer function by G(s). Let F be a bounded
linear operator from Y to U such that I G(s)F is invertible in some right half-plane.
If the inverse is bounded in a right half-plane, then the closed loop system as depicted
in Figure 5.3 is again well-posed. If G is regular, then so is the closed loop transfer
function.

5.7. Exercises
5.1. Prove that the function dened by (5.6) is a continuous function in t with values
in L2 (0, 1) for any x0 L2 (0, 1) and any u L2 (0, ). That is prove that
1

|x(, t) x(, t + h)|2 d 0

89

5. Well-posedness
for h 0.
5.2. In this exercise we study the p.d.e. (5.8) with boundary condition (5.9).
a) Show that for an initial condition x0 which is continuously dierentiable and
satises x0 (1) = 2x0 (0) the (classical) solution of (5.8)(5.9) is given by (5.10).
b) Show that the mapping x0 x(, t) with x(, t) given by (5.10) denes a
C0 -semigroup on L2 (0, 1).
c) Conclude that (5.10) is the mild solution of (5.8)(5.9) for any initial condition
x0 L2 (0, 1).
5.3. Prove Lemma 5.2.1.
5.4. In this exercise, we show some more results for the system dened by (5.27)(5.29).
a) Show that the system (5.27)(5.29) is a port-Hamiltonian system of the form
(3.32)(3.34).
b) Show that the solution of (5.27)(5.28) is given by
w(, t) = f (p() + t)()1 ,

(5.78)

where

p() =

()1 d

f (p()) = ()w0 (),

f (p(b) + t) = u(t),

[a, b]

(5.79)

[a, b]

(5.80)

t > 0.

(5.81)

5.5. Lemma 5.3.2: its solution is given as

w(, t) = f (n() + t)()1 ,

(5.82)

where
z

n() =

()1 d

(5.83)

f (n()) = ()w0 (),

f (t) = u(t),

t > 0.

z [a, b]

(5.84)
(5.85)

5.6. Show that the transfer function is regular with feed-though term zero if the constant mf in equation (5.17) can be chosen such that mf 0 if tf 0 for x0 = 0.

90

5.8. Notes and references

5.8. Notes and references

The third condition tells us that P1 L is diagonalizable via a continuously dierentiable
basis transformation. In Kato [14, chapter II], one can nd conditions on P1 L() such
that this is possible. For simplicity, we have assumed that () is continuously dierentiable.
From the proof of Theorem 5.2.6, we see that we obtain an equivalent matrix condition
for condition 1., i.e., item 1. of Theorem 5.2.6 holds if and only if K is invertible, see
Theorem 5.3.3. Since the matrix K is obtained after a basis transformation, and depends
on the negative and positive eigenvalues of P1 H, it is not easy to rewrite this condition
in a condition for Mij .
A semigroup can be extended to a group, if the homogeneous p.d.e. has for every
initial condition a solution for negative time. Using once more the proof of Theorem
5.3.3, we see that A in item 1. of Theorem 5.2.6 generates a group if and only if K and
Q are invertible matrices.
That the system remains well-posed after feedback was proved by Weiss [30].
This chapter is completely based on the paper by Zwart et al, [33].

91

5. Well-posedness

92

Chapter 6

Stability and Stabilizability

6.1. Introduction
In this chapter we study the stability of our systems. We study the stability of the
state, i.e., we only look at the solutions of the homogeneous dierential equation. As for
non-linear systems there are two dierent notions of stability. Namely strong stability
and exponential stability, which are dened next. Let A be the innitesimal generator
of the C0 -semigroup (T (t))t0 on the Hilbert space X. We know that x(t) := T (t)x0 is
the (unique) mild solution of the dierential equation
x(t) = Ax(t),

x(0) = x0 .

(6.1)

For the solutions we dene two concepts of stability.

Denition 6.1.1. The system (6.1) is strongly stable if for every x0 X the state
trajectory x(t) converges to zero for t going to innity.

Denition 6.1.2. The system (6.1) is exponentially stable, if there exists a M > 0,
< 0 such that
T (t) M et ,
t 0.
(6.2)

Since x(t) = T (t)x0 , it is easy to see that exponential stability implies strong stability.
The converse does not hold as is shown in the following example.
Example 6.1.3 Let X be the Hilbert space L2 (0, ) and let the C0 -semigroup be given
as
(T (t)f ) () = f (t + ).
(6.3)
As in Example 2.2.3 it is not hard to show that this is a strongly continuous semigroup.
Furthermore, for all t 0 we have T (t) = 1, see Exercise 6.1. The later implies that
this semigroup is not exponentially stable. It remains to show that it is strongly stable.

93

6. Stability and Stabilizability

Let > 0 and let f X = L2 (0, ). There exists a function f such that f f
and f () is zero for suciently large. Let denote f the point from which f () equals
zero. For t > f , we have that T (t)f = 0 and so
T (t)f T (t)f T (t)f + T (t)f

=
0

|f (t + ) f (t + )|2 d
|f (z) f (z)|2 dz .

Since this holds for any , we conclude that (T (t))t0 is strongly stable.
In this book we study systems in strong connection with their energy. This energy
serves as our norm. In the following section we show that if the energy is decaying we
have (under mild conditions) exponential stability. We urge to say that this holds for
our nice class of port-Hamiltonian system, and does not hold generally as the following
example shows.
Example 6.1.4 In this example we construct a contraction semigroup whose norm is
strictly decreasing, but the semigroup is not strongly stable.
We take the Hilbert space L2 (0, ). However, not with its standard inner product,
but we choose as inner product

f, g =

f ()g() e + 1 d.

(6.4)

As semigroup, we choose the right shift semigroup:

(T (t)f ) () =

f ( t) > t
0
[0, t)

= f ( t)[0,) ( t).
Using the formula for the norm and the formula of the semigroup, we see that
T (t)f

=
0

=
0

94

|f ( t)[0,])( t)|2 e + 1 d
|f ()|2 e(+t) + 1 d
|f ()|2 d

1
|f ()|2 e + 1 d
2 0
1
f 2.
2

(6.5)

6.2. Exponential stability of port-Hamiltonian systems

Hence (T (t))t0 cannot be strongly stable. Even more importantly, with the exception
of f = 0, there is no initial condition for which T (t)f 0 as t .
Next we show that the norm of the trajectory is always decreasing. Let t2 > t1 and
let f = 0. From (6.5) we know that
T (t2 )f

=
0

<
0

|f ()|2 e(+t2 ) + 1 d
|f ()|2 e(+t1 ) + 1 d

= T (t1 )f

where we have used that the negative exponential is strictly decreasing. Hence the norm
of any trajectory is decaying, but the system is not strongly stable.
Note that the above example shows that the second method of Lyapunov1 is not
directly applicable for p.d.e.s2 .
We end with a small technical lemma, which will be useful later on.
Lemma 6.1.5. Let (T (t))0 be a strongly continuous semigroup on the Hilbert space
X. If for some t1 > 0 we have that T (t1 ) < 1, then the C0 -semigroup is exponentially
stable.
Proof: If there exists a t1 > 0 such that T (t1 ) < 1, then we have that t1 log T (t1 ) <
1
0. Hence 0 = inf t>0 1 log T (t) < 0. By Theorem 2.5.1.e, we can nd a negative
t
such that T (t) M et . Hence we have exponential stability.
In the following section we consider our class of port-Hamiltonian system, and we
show that a simple condition is guaranteeing exponential stability.

6.2. Exponential stability of port-Hamiltonian systems

We return to our homogeneous port-Hamiltonian system of Section 2.3. That is we
consider the p.d.e.

x
(, t) = P1
[H()x(, t)] + P0 [H()x(, t)] .
t

(6.6)

with the boundary condition

WB
where

f (t)
e (t)

= 0,

1 P1 P1
(Hx) (b, t)
=
I
(Hx) (a, t)
2 I
As in Theorem 2.3.2 we assume that the following holds
1
2

f (t)
e (t)

(6.7)

(6.8)

Given a Lyapunov function V such that V < 0, implies (asymptotic) stability

One additionally needs pre-compactness of the trajectories

95

6. Stability and Stabilizability

Assumption 6.2.1:
P1 is an invertible, symmetric real n n matrix;
P0 is an anti-symmetric real n n matrix;
For all [a, b] the n n matrix H() is real, symmetric, and mI H() M I,
for some M, m > 0 independent of ;
H is continuously dierentiable on the interval [a, b];
WB be a full rank real matrix of size n 2n;
T
WB WB 0, where =

0I
I 0

The above, with the exception of the dierentiability of H, have been our standard
assumptions in many previous chapters. However, we would like to remark that our
T
main theorem 6.2.3 also holds if P0 satises P0 + P0 0. Under the conditions as listed
in Assumption 6.2.1 we know that the operator A given by
Ax := P1

d
[Hx] + P0 [Hx]
d

(6.9)

with domain
D(A) = {x L2 ((a, b); Rn ) | Hx H 1 ((a, b); Rn ), WB

f
e

= 0}

(6.10)

generates a contraction semigroup on the state space

X = L2 ((a, b); Rn )
with inner product
f, g

1
2

(6.11)

f ()T H()g()d.

(6.12)

In the following lemma, we show that the norm/energy of a state trajectory can be
bounded by the energy at one of the boundaries. The proof is based on an idea of Cox
and Zuazua in [4].
Lemma 6.2.2. Consider the generator A of the contraction semigroup (T (t))t0 given
by (6.9) and (6.10). Let x0 X be any initial condition, then for suciently large > 0
the state trajectory x(t) := T (t)x0 satises
x( )
x( )

2
X

2
X

(Hx)(b, t)

dt

(Hx)(a, t)

dt,

and

where c > 0 is a constant that only depends on and not on x0 .

96

(6.13)
(6.14)

6.2. Exponential stability of port-Hamiltonian systems

Proof: Let x0 be an element in the domain of A and let x(t) := T (t)x0 . Since x0
D(A), we know that x(t) D(A) for all t. Furthermore, x(t) is the classical solution of
(6.6).
For this trajectory, we dene the (positive) function for [a, b]
(b)

F () =

xT (, t)H()x(, t) dt,

(6.15)

(b)

where we assume that > 0 and > 2(b a). This last condition implies that we are
not integrating over a negative time interval.
Dierentiating this function with respect to gives
dF
() =
d

(b)

xT (, t)

(b)
(b)

(H()x(, t)) dt+

x(, t)

(b)
T

H()x(, t) dt+

x (, (b ))H()x(, (b ))+

xT (, (b ))H()x(, (b )).
Since P1 is non-singular and since x satises (6.6), we obtain (for simplicity we omit the
dependence on and t)
dF
() =
d

(b)
(b)
T

x
P0 Hx
t

1
xT P1

dt

+ x (, (b ))H()x(, (b ))
+

(b)

(b)
T

1
P1

x dH
1

x P1 P0 Hx
t
d

x dt

+ x (, (b ))H()x(, (b ))
(b)

1
xT P1

(b)
(b)

(b)
(b)
(b)
T

xT

x x T 1
+
P x dt
t
t 1

dH
x dt
d

1
T 1
xT HP0 P1 + P1 P0 H x dt

+ x (, (b ))H()x(, (b ))

+ xT (, (b ))H()x(, (b ))

T
where we have used that P1 = P1 , HT = H. The rst integral can be solved, and so we

97

6. Stability and Stabilizability

nd
dF
t= (b)
1
() = xT (, t)P1 x(, t) t=(b)
d
(b)
dH
xT

x dt
d
(b)

(b)
(b)
T

1
T 1
xT HP0 P1 + P1 P0 H x dt

+ x (, (b ))H()x(, (b ))

+ xT (, (b ))H()x(, (b )).
By simplifying the equation above one obtains
dF
() =
d

(b)

xT

(b)
T

1
T 1
HP0 P1 + P1 P0 H +

dH
d

x dt

1
+ x (, (b )) P1 + H() x(, (b ))
1
+ xT (, (b )) P1 + H() x(, (b )).

1
1
By choosing large enough, i.e., by choosing large, we get that P1 + H and P1 +
H are coercive (positive denite). This in turn implies that (for large enough)

dF
()
d

(b)
(b)

T 1
1
xT HP0 P1 + P1 P0 H +

Since P1 and P0 are constant matrices and, by assumption,

nd a > 0 such that for all [a, b] there holds
T 1
1
H()P0 P1 + P1 P0 H() +

dH
x dt.
d

dH
d ()

is bounded, we can

dH
H().
d

Thus we nd that
dF
()
d

(b)
(b)

xT (, t)H()x(, t) dt = F (),

(6.16)

where we used (6.15).

This inequality implies an inequality for F at dierent points. To prove this, we denote
(for simplicity) the derivative of F by F . From (6.16) we have that for all 1 , 2 [a, b]
with 1 2
2
2
F ()
d
(6.17)
d
1
1 F ()
Or, equivalently,
ln F (2 ) ln F (1 ) (2 1 ).

98

(6.18)

6.2. Exponential stability of port-Hamiltonian systems

Taking the exponential of this expression, and choosing 2 = b, gives:
F (b) F (1 ) e (b1 ) F (1 ) e (ba) for 1 [a, b].
On the other hand, since x(t2 ) X x(t1 )
property of the semigroup), we deduce that
(ba)

x(t)

(ba)

2
X

(6.19)

for any t2 t1 (by the contraction

dt x( (b a))

(ba)

2
X

dt

(ba)

= ( 2(b a)) x( (b a))

2
X.

Using the denition of F () and x(t) 2 , see (6.15) and (6.12), together with the equaX
tion above, the estimate (6.19), and the coercivity of H we obtain
2( 2(b a)) x( )

2
X

2( 2(b a)) x( (b a))

b

(ba)

(b)

xT (, t)H()x(, t) dt d

(b)
b

xT (, t)H()x(, t) dt d

(ba)
b

2
X

F () d (b a)F (b) e (ba)

= (b a) e (ba)
Mm

xT (b, t)H(b)x(b, t) dt

(ba)

(b a) e

(Hx)(b, t)

dt.

Hence for our choice of we have that

x( )

2
X

(Hx)(b, t)

dt,

(6.20)

(ba)

where c = M (ba) e
2( 2(ba))m . This proves estimate (6.13) for x0 D(A). Although, in
Theorem 6.2.3 we only need inequality (6.13) for x0 D(A), with the help of the
previous chapter we can obtain it for all x0 X.
f (t)
We replace (6.7) by the relation WB e (t) = u(t). Furthermore, we dene the

output as y(t) = (Hx)(b, t). Since the homogeneous equation, i.e., u=0, generates a
(contraction) semigroup, we have by Theorem 5.2.6 that this system is well-posed. In
particular, this means that if u = 0 small changes in the initial condition, gives small
changes in the state at time and the L2 -norm of the output. More specically, see
(5.17),
x( )

2
X

+
0

y(t) 2 dt mf x0

2
X.

Since the domain of A is dense in X, and since c was not depending on the initial
condition, we conclude that (6.13) holds for all initial conditions.

99

6. Stability and Stabilizability

The second estimate follows by replacing F () in the argument above by
(a)

F () =

xT (, t)H()x(, t) dt.

(a)

With this technical lemma, the proof of exponential stability is easy.

Theorem 6.2.3. Consider the operator A dened in (6.9) and (6.10). Furthermore,
we assume that the conditions in Assumption 6.2.1 are satised. If some some positive
constant k one of the following conditions is satised for all x0 D(A)
Ax0 , x0

+ x0 , Ax0

Ax0 , x0

+ x0 , Ax0

k (Hx0 )(b)

k (Hx0 )(a) 2 ,

(6.21)
(6.22)

then the system is exponentially stable.

Proof: Without loss of generality we assume that the rst inequality (6.21) holds. Let
x0 D(A), and let be the same as in Lemma 6.2.2. We denote T (t)x0 by x(t), and
since x0 D(A) we know that x(t) D(A) and x(t) = Ax(t). Using this dierential

equation, it is easy to see that

d x(t)
dt

2
X

d x(t), x(t)
= Ax(t), x(t) + x(t)Ax(t) .
dt

(6.23)

Using this and equation (6.21), we have that

x( )

2
X

x(0)

2
X

=
0

d x(t)
dt

2
X

(t) dt

(Hx)(b, t) 2 dt.

Combining this with (6.13), we nd that

x( )

2
X

x(0)

2
X

k
x( )
c

2
X.

c
Thus x( ) 2 c+k x(0) 2 . From this we see that the semigroup (T (t))t0 generated
X
X
by A satises T ( ) < 1, from which we obtain exponential stability, see Lemma 6.1.5.

Estimate (6.21) provides a simple way to prove the exponential stability property. We
note that Theorem 1.2.1 implies
Ax, x

+ x, Ax

= (Hx)T (b)P1 (Hx)(b) (Hx)T (a)P1 (Hx)(a).

(6.24)

This equality can be used on a case by case to show exponential stability. However,
T
when WB WB > 0, then the system is exponentially stable.

100

6.2. Exponential stability of port-Hamiltonian systems

T
Lemma 6.2.4. Consider the system (6.6)(6.7). If WB WB > 0 and the conditions of
Assumption 6.2.1 hold, then the system is exponentially stable.

Proof: By Lemma 2.4.1 we know that WB can be written as WB = S (I + V, I V ).

T
Since WB WB > 0 we have that V V T < I.
We dene WC = I + V T , I + V T . It is easy to see that with this choice, WC is a
I+V
IV
I+V T I+V T

n 2n matrix with rank n. Furthermore, since V V T < I, the matrix

is

invertible. This implies that WB is invertible.

WC
With the matrices WB and WC we dene a system. Namely, we take the p.d.e. (6.6)
with input and output
f (t)
e (t)

u(t) = WB

f (t)
e (t)

y(t) = WC

Applying Theorem 3.4.2 to this system, we see that

d
x(t)
dt

2
X

1
2

uT (t) y T (t)

u(t)
y(t)

PWB ,WC

Choosing u = 0, x D(A), and t = 0 is the above equation, we nd that

Ax, x

+ x, Ax

1
2

0 yT

WB
WC

The matrix PWB ,WC is the inverse of the matrix

WC we nd that
WB
WC

T
WB

T
WC

0
y

PWB ,WC
T
( WB

S[2I 2V V T ]S T
0

T
WC

(6.25)

). By the choice of

0
2I + 2V T V

Combining this with (6.25) we nd that

Ax, x

+ x, Ax

1
= y T [I + V T V ]1 y m1 y
4

(6.26)

for some m1 > 0. Here we have used that V V T < I, and hence V T V I < 0.
The relation between u, y and x is given by
0
y

1
=
2

WB
WC

P1 P1
I
I

(Hx)(b)
(Hx)(a)

where we have used that u = 0. Since P1 and

WB
WC

the matrix W is invertible and, in particular, W w

Taking norms on both sides yields
y

= W

(Hx)(b)
(Hx)(a)

m2

(Hx)(b)
(Hx)(a)

:= W

(Hx)(b)
(Hx)(a)

are non-singular, it follows that

2

m2 w

for some real m2 > 0.

m2 (Hx)(b) 2 .

(6.27)

101

6. Stability and Stabilizability

Combining the estimates (6.26) with (6.27) gives
Ax, x

+ x, Ax

m1 y

m1 m2 (Hx)(b) 2 .

Similarly, we nd
Ax, x

+ x, Ax

m1 m2 (Hx)(a) 2 .

Concluding we see that (6.21) holds, and so by Theorem 6.2.3) we conclude exponential
stability.
The situation as described in this lemma will not happen often. It implies that you
have as many dampers as boundary controls. In practice less dampers are necessary as
is shown in the example of the following section.

6.3. Examples
In this section we show how to apply the results of the previous section. We show
that once the input (boundary conditions) and the output are selected, a simple matrix
condition allows to conclude on the exponential stability.
Example 6.3.1 Consider the transmission line on the spatial interval [a, b]
Q
(, t)
(, t) =
t
L()
Q(, t)

(, t) =
.
t
C()

(6.28)

Here Q(, t) is the charge at position [a, b] and time t > 0, and (, t) is the ux
at position and time t. C is the (distributed) capacity and L is the (distributed)
inductance. This example we studied in Example 1.1.1 and 4.2.2. To the p.d.e. we add
the following input and output, see also Example E:4.2.2
u(t) =

y(t) =

Q(b,t)
C(b)
Q(a,t)
C(a)
(b,t)
L(b)
(a,t)
L(a)

V (b, t)
V (a, t)

I(b, t)
I(a, t)

(6.29)

(6.30)

First we want to know whether the homogeneous system is (exponentially) stable. There
for the determine the WB associated to (6.29). The boundary eort and ow are given
by

I(b) + I(a)
1 V (b) + V (a)
f

=
e
2 V (b) + V (a)
I(b) + I(a)
102

6.4. Exercises
Hence WB is given as
1
WB =
2

0 1 1 0
0 1 1 0

T
A simple calculation gives that WB WB = 0. Hence the homogeneous p.d.e. generates
a unitary group, and cannot be (strongly) stable, see Exercise ??
Now we apply an output feedback. If we apply a full output feedback, then it is not
hard to show that we have obtained an exponentially stable system, see Exercise ??
We want to consider a more interesting example, in which we only apply a feedback
on one of the boundaries. This is, we set the rst input to zero, and put a resistor at
the other end. This implies that we have the p.d.e. (6.28) with boundary conditions

V (a, t) = 0,

V (b, t) = RI(b, t),

(6.31)

with R > 0. This leads to the following (new) WB ;

1
WB =
2

R 1 1 R
0 1 1 0

(6.32)

T
0
Hence we have that WB WB = 2R 0 . Since this is not positive denite, we may not
0
apply Lemma 6.2.4. Hence we must check whether one of the condition of Theorem 6.2.3
is satised. For this we return to the original balance equation, equation (6.24).
Using (6.24), we get

Ax, x

+ x, Ax

= V (a)I(a) V (b)I(b) = RI(b)2 .

Furthermore, we have that (Hx)(b) =

(Hx)(b)

V (b)
I(b)

(6.33)

. Thus

= V (b)2 + I(b)2 = (R2 + 1)I(b)2 .

(6.34)

Combining the two previous equations, we nd that

Ax, x

+ x, Ax

R
(Hx)(b) 2 .
1 + R2

(6.35)

Hence by Theorem 6.2.3 we conclude that putting a resistor at one end of the transmission line stabilizes the system exponentially.

6.4. Exercises
6.1. Prove that the expression given in (6.3) denes a strongly continuous semigroup
on L2 (0, ).
6.2. Show that a unitary group cannot be strongly stable.

6.5. Notes and references

The results in this chapter can be found in [29].

103

6. Stability and Stabilizability

104

Chapter 7

Systems with Dissipation

7.1. Introduction
As the title indicates, in this chapter we study systems with dissipation. These systems
appear naturally in many physical situations. For instance, by internal damping in a
vibrating string, or by diusion of heat in a metal bar. The behavior of these models
will be dierent in nature than the model we have seen until now. For instance, since
energy/heat dissipates, and hence cannot be recovered, we will not have a group, i.e.,
cannot go backward in time. Although the behavior is dierent, the results as obtained
in the previous chapters can be used to prove existence and uniqueness for our class of
dissipative systems. We begin by recapitulating two examples from Chapter 1.
Example 7.1.1 (Damped wave equation) Consider the one-dimensional wave equation of Example 1.1.2. One cause of damping is known as structural damping. Structural damping arises from internal friction in a material converting vibrational energy
into heat. In this case the vibrating string is modeled by
1
ks 2 w
w
2w
(, t) =
(, t) +
(, t) ,
T ()
t2
()

() 2 t

[a, b], t 0,

(7.1)

where ks is a positive constant and the other variables have the same meaning as in
Example 1.1.2.
In the previous example, we can still recognize a system which we have studied before,
namely the (undamped) wave equation. In the model of the heat conduction the relation
with our class of port-Hamiltonian systems seems completely lost.
Example 7.1.2 (Heat conduction) The model of heat conduction is given by
1
T
(, t) =
t
cV

()

T (, t)

(7.2)

where T (, t) denotes the temperature at position [a, b] and time t, cV is the heat
capacity, and () denotes the heat conduction coecient.

105

7. Systems with Dissipation

As we did for the undamped system, we see these models as examples of a general class
of models. For this new model class, we investigate existence of solutions, boundary control, etc. We begin by showing that this new model class is the class of port-Hamiltonian
system with an extra closure relation. Based on the underlying port-Hamiltonian system,
proving existence of solutions, etc. will be easy.

7.2. General class of system with dissipation.

The general equation describing our class of port-Hamiltonian system was given by
Hx
x
(, t) = P1
(, t) + P0 (Hx) (, t)
t

(7.3)

Now we add a dissipation term, and we obtain the following p.d.e.

x

(, t) = (J GR SGR ) (Hx) (, t),

t

x(, 0) = x0 (), [a, b],

(7.4)

x
+ GT x,
0

(7.5)

where
J x = P1

x
+ P0 x,

GR f = G1

f
+ G0 f,

GR x = GT
1

G is known as the formal adjoint of GR . As before, we assume

R
Assumption 7.2.1:
P1 is a symmetric real n n matrix;
P0 is an anti-symmetric real n n matrix;
For all [a, b] the n n matrix H() is real, symmetric, and mI H() M I,
for some M, m > 0 independent of .
On the new term we assume
G1 and G0 are real n r matrices;
For all [a, b] the r r matrix S() is real, symmetric, and m1 I S() M1 I,
for some M1 , m1 > 0 independent of .

Note that we have removed the invertibility assumption on P1 . This assumption will be
replaced by another assumption, see Assumption 7.2.7.
First we check whether the two examples from the introduction are in the class dened
by (7.4) and (7.5).
Example 7.2.2 For the damped wave equation of Example 7.1.1 we have that the state
is given by x =

w
t
w

H() =

106

. The matrices H, P1 and P0 are the ones found in (1.20), i.e.,

1
()

0
T ()

P1 =

0 1
1 0

P0 = 0.

7.2. General class of system with dissipation.

Using this and the formula for the state we nd that (here omit and t)
x1
x2

0 1
1 0

0
0 T

x1
x2

0 1
1 0

0
0 T

x1
x2

+ ks

x1
x2

ks

2 1 x

0 1
1 0

ks

0
0 T

0
0 T

0
2
2

0
0

0
0 T

x1
x2

x1
x2

Hence
G1 =

1
0

G0 = 0

and S = ks .

Concluding, we see that the damped wave equation can be written in the form (7.4)(7.5).
Furthermore, it is easy to see that the conditions in Assumption 7.2.1 are satised.
Example 7.2.3 The equation for heat condition can also be written in the format (7.4)
(7.5). We choose as state x(, t) the temperature T (, t) and furthermore, we choose
P1 = P0 = G0 = 0,

G1 = 1,

H() =

1
and S() = ().
cV

Then we easily see that (7.4) becomes (7.2). The conditions of Assumption 7.2.1 are
trivially satised.
Similar to Theorem 1.2.1, there holds a balance equation for the system (7.4)(7.5)
for the energy
1 b
E(t) =
x(, t)T H()x(, t)d.
(7.6)
2 a
Lemma 7.2.4. Under the assumptions 7.2.1 the following balance equation holds for
classical solution to the p.d.e. (7.4) and (7.5)
b
1
dE
(t) = (Hx)T (, t)P1 (Hx) (, t)
dt
2
a
1

(Hx)T (, t)G1 (SGR Hx) (, t) + (SGR Hx)T (, t)GT (Hx) (, t)

1
2
b
a

(7.7)
b
a

(GR Hx)T (, t)S() (GR Hx) (, t)d,

where E(t) is given by (7.6).

107

7. Systems with Dissipation

Proof: The proof is very similar to that of Theorem 1.2.1. From that proof we easily
see that
2

dE
(t) = (Hx)T (, t)P1 (Hx) (, t)
dt
b
a

b
a

(7.8)

[(GR SGR ) (Hx) (, t)]T (Hx) (, t) + x(, t)T H() [(GR SGR ) (Hx) (, t)] d.

We concentrate on the second term, and we introduce some notation to simplify the

formula. We write z = Hx, q = GR z. Furthermore, we omit the t. Using this combined

with the fact that H() is symmetric, we nd that the second term of (7.8) becomes
b

[GR Sq()]T z() + z()T [GR Sq()] d.

(7.9)

Using integration by parts, we have that

b

z()T [GR Sq()] d =

z()T G1

a
b

=
a
b

=
a
b

=
a
b

=
a

(Sq) () + G0 q() d

z T
() [G1 (Sq) () + G0 (Sq) ()] d + z()T G1 (Sq) ()

GT
1

z
() + GT z()
0

b
a

(Sq) ()d + z()T G1 (Sq) ()

(GR z)T () (Sq) ()d + z()T G1 (Sq) ()

b
a

b
a

(GR z)T ()S() (GR z) ()d + z()T G1 (SGR z) ()

b
.
a

Using this and its transpose in (7.9) we nd that (7.8) becomes

2

dE
(t) = (Hx)T (, t)P1 (Hx) (, t)
dt
b

2
a

b
a

(7.10)

(GR Hx)T ()S() (GR Hx) ()d

(Hx)T ()G1 (SGR Hx) () + (SGR Hx)T ()GT (Hx) ().

1

b
a

From (7.7) we clearly see that if the boundary conditions are such that the term
(Hx)T (, t)P1 (Hx) (, t)

(Hx)T (, t)G1 (SGR Hx) (, t) + (SGR Hx)T (, t)GT (Hx) (, t)

1

b
a

0,

then the energy is decaying. However, this is under the assumption of the existence of
a solution.

108

7.2. General class of system with dissipation.

As said in the introduction, we want to use the results form the previous chapters
for proving existence and uniqueness of solutions. However, the nature of the system
with dissipation seems completely dierent to that of the system we studied till now.
The following example indicates how we may use an extended port-Hamiltonian system
together with a closure relation for obtaining a system with dissipation.
Example 7.2.5 (General idea) Consider the following model. This could be the
model of the transmission line, equation (1.1), in which we have taken all physical
parameters equal to one,
x2
x1
(, t) =
(, t)
t

x2
x1
(, t) =
(, t).
t

(7.11)
(7.12)

Instead of looking at it as a dierential equation, we regard this as an relation between

the variables e and f , given by
f1
fp

e1
ep

:= Je e

(7.13)

in which we substituted f = x and e = x. The power balance for the p.d.e. came directly

from the following relation, see Section 1.3

f T ()e() + eT ()f ()d = 0.

(7.14)

f1 ()e1 () + fp ()ep ()d = 0.

(7.15)

Or equivalently,

If we add the closure relation

ep = Sfp

(7.16)

with S = S() a bounded and strictly positive function, to the equation (7.13), then we
see that for a pair (f, e) satisfying both equations
f1 ()e1 ()d =

f1 ()e1 () + fp ()ep ()d

=0

fp ()ep ()d

fp ()S()fp ()d 0,

(7.17)

where we used (7.15).

Next we dened a new system using (7.13) and (7.16). We take f1 =
Using equations (7.13) and (7.16), we nd that
ep
Sfp

x
= f1 =
=
=
t

e1

x
t

and e1 = x.

(7.18)

109

7. Systems with Dissipation

Hence we have obtained the diusion equation. Furthermore, from (7.17) we nd that
d
dt

x
(, t)x(, t)d = 2
t

x(, t)2 d = 2

f1 e1 d 0.

x(, t)2 d is dissipating.

Hence the energy

e1

f1
Je

ep

fp

S
Figure 7.1.: Interconnection structure.
In the example we see that we can obtain a system with dissipation by adding a closure
relation to a larger port-Hamiltonian system. Our system (7.4)(7.5) can be seen in the
same way.

Lemma 7.2.6. The operator J GR SGR dened (7.5) can be seen as the mapping Je
together with the closure relation ep = Sfp , where Je is dened by

Je

e1
ep

P1
GT
1

G1
0

e1
ep
f1
fp

e1
Proof: Dene the image of Je ( ep ) as

P0
GT
0

G0
0

e
ep

(7.19)

, see also Figure 7.1. Then

ep

e
+ G1
+ P0 e1 + G0 ep = J e1 + G1
+ G0 ep

= J e1 + GR ep = J e1 + GR Sfp

f1 = P1

e1
GT e1
0

= J e1 GR SGR e1 .
= J e1 + GR S GT
1

This proves the assertion.

From equation (7.19) we see that it is natural to introduce a new P1 and P0 as
P1,ext =

P1
GT
1

G1
0

P0,ext =

P0
GT
0

G0
0

(7.20)

By the conditions on P1 and P0 , see Assumption 7.2.1, we have that P1,ext is symmetric,
and P0,ext is anti-symmetric. To the conditions listed in Assumption 7.2.1 we add the
following.

110

7.2. General class of system with dissipation.

Assumption 7.2.7: The matrix P1,ext as dened in (7.20) is invertible.

Under this assumption, we see that our operator (7.19) ts perfectly in the theory
as developed in Section 2.3. The following theorem is a direct consequence of Theorem
2.3.2. By the Hilbert space Xext we denote the space L2 ((a, b); Rn+r ) with inner product
x
xp

z
zp

=
a

Xext

x()T H()z()d +

Hence Xext = X L2 ((a, b); Rr ). Furthermore, we dene

1
=
2

f,Hx,xp
e,Hx,xp

P1,ext P1,ext
In+r
In+r

xp ()T zp ()d.

(7.21)

(Hx)(b)
xp (b)
(Hx)(a)
xp (a)

(7.22)

The matrix WB is a full rank, real matrix of size (n + r) 2(n + r). With this notation,
we can formulate the following theorem, which is an adaptation of Theorem 2.3.2 to our
extended setting.
Theorem 7.2.8. Under the conditions of Assumptions 7.2.1 and 7.2.7 we have that the
operator Aext dened as
Aext

x
xp

= P1,ext

(Hx)
xp

+ P0,ext

(Hx)
xp

(7.23)

with domain given by

x
D(Aext ) = ( xp ) Xext |

Hx
xp

H 1 ((a, b); Rn+r ) with WB

f,Hx,xp
e,Hx,xp

=0

(7.24)

T
generates a contraction semigroup on Xext if and only if WB WB 0.

Based on this theorem and the fact S dissipates energy, we can prove that the operator
associated to p.d.e. (7.4) generates a contraction semigroup provided that one uses the
correct boundary conditions. The proof of this result is an application of the general
result Theorem 7.3.3 combined with the previous theorem.

Theorem 7.2.9. Denote by J , GR and GR the operators as dened in (7.7). Furthermore, let the Assumptions 7.2.1 and 7.2.7 be satised and let WB be a (n + r) 2(n + r)
T
matrix of full rank such that WB WB 0. Then the operator AS dened as

AS x = (J GR SGR ) (Hx)

(7.25)

with domain
D(AS ) = {x X | Hx H 1 ((a, b); Rn ) with WB

f,Hx,SGR (Hx)
e,Hx,SGR (Hx)

= 0}

(7.26)

generates a contraction semigroup on X.

111

7. Systems with Dissipation

Proof: If we can show that the operator AS with its domain can be written as (7.39)
(7.40), and if the other conditions of Theorem 7.3.3 holds, then the proof becomes a
straightforward application of this general theorem.
We take Aext to be the operator dened by (7.23) with domain given by (7.24). Using

the notation J , GR and GR we see that this operator can be written as

Aext =

J
GR

GR 0

H 0
0 I

(7.27)

This is easy formulated into the set-up of Theorem 7.3.3 be dening

A2 = GR H,

D(A2 ) = {x L2 ((a, b); Rn ) | Hx H 1 ((a, b); Rn )}

(7.28)

and
A1 =

J H GR

D(A1 ) = D(Aext ).

(7.29)

Hence A1 has the domain given by (7.24). Note the domain of A2 imposes no extra
restriction the domain of A1 .
Next we dene S to be the multiplication operator
(Sf )() = S()f ().

(7.30)

Using the assumption on S it is easy to see that S L(L2 ((a, b); Rr )) and it satises
(7.37).
h
By Theorem 7.3.3 we have that A1 S(A2 h) with domain D(AS ) = {h H1 |
h
S(A2 h)

D(Aext )} generates a contraction semigroup on X. From (7.26) and (7.27)

we see that this AS is also given by (7.25) with domain (7.26). This concludes the proof.
We begin by applying the theorem to the example of the heat conduction, see Example
7.1.2.
Example 7.2.10 In this example we want to investigate for which boundary conditions
the p.d.e. (7.2) describing the heat conductivity generates a contraction semigroup. We
begin by identifying the extended state. Using Example 7.5 and equation (7.20) we see
that
0 1
0 0
P1,ext =
,
P0,ext =
.
1 0
0 0
This implies that
f,g,xp
e,g,xp

0
1 1
=
2 1
0

1 0 1
g(b)
0 1 0 xp (b)

0 1
0 g(a)
1 0
1
xp (a)

xp (b) xp (a)

1
= g(b) g(a) . (7.31)

2 g(b) + g(a)
xp (b) + xp (a)

For the matrix WB describing the boundary condition, we observe the following. It
T
should be full rank, and WB WB 0, and we describe the boundary conditions as

112

7.2. General class of system with dissipation.

being in its kernel. By Lemma 2.4.1 we see that without loss of generality, we may
assume that
WB = (I + V I V )
f,g,xp
e,g,xp

with V V T I. Using (7.31) we see that WB

(I + V )

xp (b) xp (a)
g(b) g(a)

= 0 is equivalent to
g(b) + g(a)
xp (b) + xp (a)

= (V I)

(7.32)

From the theory of Chapter 2 we know that with the above boundary conditions the extended operator generates a contraction semigroup, provided V V T I. Using Theorem
7.2.9, we see that under the same conditions on V the operator
AS :=

1 d
cV d

()

d
d

(7.33)

with domain
D(AS ) = { x L2 ((a, b); R) | x H 1 ((a, b); R),
(I + V )

(b) dx
cv d (b)
1
cv x(b)

(a) dx
cv d (a)
1
cv x(a)

dx
H 1 ((a, b); R) and
d

= (V I)

(b) dx
cv d (b) +
1
cv x(b) +

(7.34)
(a) dx
cv d (a)
1
cv x(a)

generates a contraction semigroup on L2 ((a, b); R). This implies that the homogeneous
p.d.e. (7.2) with boundary conditions
(I + V )

(b) dx (b) (a) dx (a)

d
d
x(b) x(a)

= (V I)

(b) dx (b) + (a) dx (a)

d
d
x(b) + x(a)

(7.35)

has a mild solution for every initial condition in L2 ((a, b); R). Choosing V = 0, we nd
as boundary conditions
dx
(b) (b) = 0
x(b) = 0,
d
whereas V = I gives
(b)

dx
dx
(b) = (a) (a)
d
d

x(b) = x(a).

We end by saying that Theorem 7.2.9 cannot be used for the damped wave equation
of Example 7.1.1. The reason this lies in the fact that the extended P1 matrix is noninvertible. Namely, using Example 7.2.2 we nd that

0 1 1
P1,ext = 1 0 0
1 0 0
which is clearly no invertible. We remark that Theorem 7.2.8 and hence Theorem 2.3.2
can be extended such that non-invertible P1 s are possible, see [28].

113

7. Systems with Dissipation

7.3. General result

In this section we prove a general result, which is very useful in proving existence of the
(homogeneous) dierential equation. For the proof of this theorem we need the following
two lemmas
Lemma 7.3.1. Let A be a linear operator from D(A) X to X, where X is a Hilbert
space, then A generates a contraction semigroup if and only if the following two conditions are satised:
1. For all x0 D(A)
2Re ( Ax0 , x0

X)

= Ax0 , x0

+ x0 , Ax0

0,

(7.36)

2. The range of I A equals X for some > 0.

Lemma 7.3.2. Let A be the generator of a contraction semigroup on the Hilbert space
X, and let P L(X) satisfy Re ( P x, x ) 0 for all x X, then A + P generates a
contraction semigroup.
Theorem 7.3.3. Let H1 and H2 be two Hilbert spaces. Furthermore, Let A1 be a linear
operator from D(A1 ) H1 H2 to H1 and A2 is a linear operator from D(A2 ) H1 to
H2 , and let S L(H2 ) such that it is invertible and satises
Re ( Sx2 , x2 ) m2 x2 2 ,

x2 H2

(7.37)

for some m2 > 0 independent of x2 .

If the operator
Aext :=

A1
A2 0

(7.38)

with the domain D(Aext ) = {(h1 , h2 ) H1 H2 | h1 D(A2 ) and (h1 , h2 ) D(A1 )}

generates a contraction semigroup on H1 H2 , then
AS h = A1
with domain
D(AS ) = {h H1 |

h
S (A2 h)
h
S (A2 h)

(7.39)

D(Aext )}

(7.40)

generates a contraction semigroup on H1 .

Proof: By Lemma 7.3.1 we have to check two conditions for AS ; that is
Re ( AS x1 , x1 ) 0

for all x1 D(AS )

(7.41)

for some > 0.

(7.42)

and
ran(I AS ) = H1

114

7.3. General result

We start by showing (7.41). Let x1 D(AS )
AS x1 , x1

x1
S(A2 x1 )

A1

Aext

x1
S(A2 x1 )

x1
0

Aext

x1
S(A2 x1 )

x1
S(A2 x1 )

x1
S(A2 x1 )

Aext
=

Aext

, x1

x1
S(A2 x1 )

,
,

0
S(A2 x1 )
x1
S(A2 x1 )

A2 x1 , S(A2 x1 ) .

Using the fact that Aext generates a contraction semigroup, we nd

Re ( AS x1 , x1 ) = Re

Aext

x1
S(A2 x1 )

x1
S(A2 x1 )

Re ( A2 x1 , S(A2 x1 ) )

0 Re ( A2 x1 , S(A2 x1 ) ) 0,
where in the last step we used that S satises (7.37).
Next we prove the range condition (7.42) on AS . That is, for a > 0 we have to show
that for any given f X we can nd an x D(AS ) such that
f = (I AS )x.
From (7.37) we nd that
Re

S 1 z2 , z2
m2
,
S 2

Choose such that 0 < <

H1 H2

m2 S 1 z2

m2
z2 2 .
S 2

(7.43)

and dene the following mapping from H1 H2 to

P =

0
0
1 + I
0 S

By our assumption we have that Re ( P x, x ) 0 for all x H1 H2 . Hence by Lemma

7.3.2, we conclude that Aext + P is generates a contraction semigroup. In particular, the
range of I Aext P is the whole space. This implies that for all f H1 H2 there
0
exists an

h1
h2

D(Aext ) such that

f
0

= (I Aext P )

h1
h2

(7.44)

This is equivalent to
f = h1 A1

h1
h2

115

7. Systems with Dissipation

and
0 = h2 A2 h1 + (S 1 )h2 .
This last relation, implies that h2 = S(A2 h1 ). Hence h1 D(AS ), and f = (I AS )h1 .
Concluding, we see that the range of I AS is full, and so it generates a contraction
semigroup on H1 .

7.4. Exercises
7.5. Notes and references
This chapter is based on Chapter 6 of [28]. Theorem 7.2.9 can be extended to allow for
non-linear dissipation terms.

116

Chapter A

Mathematical Background

A.1. Complex analysis

In this section, we present important facts from complex function theory that are used in this
book. As main references, we have used Levison and Redheer [17] and Rudin [24].
By a domain we shall mean a nonempty, open, and connected subset of C; in some books the
name region is used.
Denition A.1.1. Let be a domain in C, and let f be a function dened on with values
df
in C. The function f is holomorphic on if ds (s0 ) exists for every s0 in .
The function is said to be entire if it is holomorphic on C.
The function g is meromorphic on if g can be expressed as g = f1 , where f1 and f2 are
f2
holomorphic on .

We remark that some texts use the the term analytic instead of holomorphic. Examples of
holomorphic functions are all polynomials and exponential powers; the latter are entire functions.
Rational functions are meromorphic on C and holomorphic on every domain not containing the
zeros of the denominator. It is easily veried that these classes of holomorphic functions have the
property that the derivative is again a holomorphic function. In fact, this is a general property
of holomorphic functions.
Proposition A.1.2. A function f that is holomorphic on the domain possesses the following
properties:
a. The derivative

df
ds

is holomorphic on the same domain;

b. If f = 0 in , then f 1 is holomorphic on ;
c. f is uniformly bounded on every compact subset of C contained in .

Proof a. See theorem 5.2 in Levison and Redheer [17] or the corollary of theorem 10.16 in
Rudin [24].
b. This follows easily from the fact that
df
df 1
= f 2 .
ds
ds
c. This follows directly from the fact that f is continuous on .

117

A. Mathematical Background
The last property shows that holomorphic functions have similar properties to real-valued
functions. However, in contrast to functions on R, it cannot be bounded on the whole complex
plane, unless it is a constant.
Theorem A.1.3. Liouvilles Theorem. If f is an entire function that is bounded on the
whole complex plane, then it is a constant.
Proof See theorem 5.5 of chapter 3 in Levison and Redheer [17] or theorem 10.23 of Rudin
[24].
The zeros of a holomorphic function have very nice properties, as can be seen in the next
theorem.
Theorem A.1.4. The zeros of a function f that is holomorphic on the domain have the
following properties:
a. If s0 is a zero of f , then f (s) is either identically zero on or the order of the zero is
nite, that is, there exists an m such that f (s) = (s s0 )m g(s), where g is holomorphic
on and g(s0 ) = 0;
b. If f is not identically zero on , then the zeros of f are isolated; that is, for every zero s0
of f there exists a > 0 such that f (s) = 0 for all s satisfying 0 < |s s0 | < ;
c. If the zeros of f have a limit point in , then f is identically zero;
d. In every compact subset V of C with V , there are only nitely many zeros, provided
that f is not identically zero.
Proof This result can be found in theorem 10.18 of Rudin [24] and in theorems 7.27.4 of Levison
and Redheer [17].
A corollary of Theorem A.1.4.c is that two functions f1 , f2 , that are holomorphic on the
domains 1 and 2 , respectively, and are equal on a set containing a limit point in 1 2 , are
in fact equal on 1 2 . Furthermore, there exists a unique function f that is holomorphic on
1 2 such that f = f1 on 1 and f = f2 on 2 . This f is called the holomorphic continuation.
Denition A.1.5. A curve in the complex plane is called a rectiable curve if there exists
an interval [a, b] R and a continuously dierentiable mapping from [a, b] to C such that the
image of equals , that is, = ([a, b]). The rectiable curve is called simple if (x) = (y)
for all x and y in (a, b) such that x = y. It is called closed if (a) = (b). By a contour we
shall mean a nite collection of rectiable curves j , j = 1, . . . , n, such that the nal point of j
is the initial point of j+1 for 1 j n 1. The notions of simple and closed are the same for
these curves.

Theorem A.1.6. Rouchs Theorem. Let f1 and f2 be functions that are holomorphic on
e
the domain , and suppose that contains a simple, closed contour . If |f1 (s)| > |f2 (s)| for
s , then f1 and f1 + f2 have the same number of zeros inside . (A zero of order p counts for
p zeros.)
Proof See theorem 6.2 in Levison and Redheer [17] or theorem 10.43 in Rudin [24].

118

A.1. Complex analysis

Denition A.1.7. For a function f that is continuous on the domain we dene its integral
along the rectiable curve by
b

f (s)ds :=

f ((x))

d
(x)dx.
dx

(A.1)

Its integral over a contour is dened by

n

f (s)ds,

(A.2)

where j , 1 j n, are the curves that form the contour .

f (s)ds =
j=1
j

Before we can state one of the most important theorems of complex analysis, we need the
concept of the orientation of a rectiable, simple, closed contour. Let the contour be composed of
dj
the rectiable curves j = j ([aj , bj ]), and choose a point x0 from (aj , bj ) such that dx (x0 ) = 0.
dj
If the vector obtained by rotating the tangent vector dx (x0 ) in a counterclockwise sense through
an angle of points inside the interior bounded by the contour , then the rectiable, closed,
2
simple contour is said to be positively oriented. For a circle it is easily seen that it is positively
oriented if one transverses the circle in a counterclockwise sense going from a to b.
Theorem A.1.8. Cauchys Theorem. Consider the simply connected domain that contains
the positively oriented, closed, simple contour . If f is holomorphic on , then
f (s)ds = 0,

and for any point s0 inside

1
2

f (s)
ds = f (s0 ).
s s0

Proof See Levison and Redheer [17, pp. 180 and 183] or theorem 10.35 in Rudin [24].
Denition A.1.9. Let g be a function that is meromorphic on the domain . A point s0 in
is dened to be a pole of g if lim |g(s)| = . The order of the pole is dened to be the smallest
ss0

positive integer m such that lim |(s s0 )m g(s)| < .

ss0

f1
It is easily seen that if g can be expressed as g = f2 , where f1 and f2 are holomorphic on ,
then s0 is a pole of g only if s0 is a zero of f2 . Since the zeros have nite order (see Theorem
A.1.4), so do the poles.
If g is a meromorphic function on the domain with no poles on , then it is continuous on
and hence (A.2) is well dened.

Theorem A.1.10. Cauchys Residue Theorem. Let g be a function that is meromorphic

on the simply connected domain with s0 as its only pole inside the positively oriented, simple,
closed contour . Assume further that there are no poles on the contour . Then
1
2

g(s)ds =

dm1
1
(s s0 )m g(s)
(m 1)! dsm1

(A.3)

s=s0

where m is the order of the pole s0 .

119

A. Mathematical Background
Proof See theorem 2.1 in Levison and Redheer [17] or theorem 10.42 in Rudin [24].
The value on the right-hand side of equation (A.3) is called the residue of g at s0 . If the
meromorphic function f contains nitely many poles inside the contour , then the integral in
equation (A.3) equals the sum over all the residues.
In the next theorem, we see that it is possible to express a meromorphic function with a pole
at s0 as an innite series of positive and negative powers of s s0 .
Theorem A.1.11. Let f be a holomorphic function on the punctured disc {s C | 0 < |ss0 | <
R} and let C be the circle {s C | |s s0 | = r} for any r satisfying 0 < r < R. If we dene
ak :=

1
2
C

f (s)
ds
(s s0 )k+1

for k Z, where C is transversed in a counterclockwise sense, then the Laurent series given by

f (s) =
k=

ak (s s0 )k

converges uniformly to f (s) in any closed annulus contained in the punctured disc {s C | 0 <
|s s0 | < R}.
Proof See Levison and Redheer [17, theorem 9.2].
We remark that if the function is holomorphic on the disc {s C | |s s0 | < R}, then
aj = 0 for negative values of j. Hence for every holomorphic function there exists a sequence of
polynomials that approximate it on an open disc. In the next theorem, we shall see how good
this approximation is on the closed disc.
Theorem A.1.12. We dene the disc D(z0 , R) := {z C | |z z0 | R}. If f is a holomorphic
function on the interior of D(z0 , R) and continuous on the boundary, then for every > 0 there
exists a polynomial P such that
sup
zD(z0 ,R)

|f (z) P (z)| < .

(A.4)

Proof See theorem 20.5 in Rudin [24].

We remark that if a sequence of polynomials converges to a function in the norm in equation
(A.4), then this limit function is continuous on the boundary.
df
For the special case that the meromorphic function in Theorem A.1.10 is given by f 1 ds , we
have the following result.
Theorem A.1.13. Principle of the Argument. Let be a simply connected domain and
let be a positively oriented, simple, closed contour contained in . Let g be a function that is
meromorphic on with no zeros or poles on , and let N () and P () denote the number of
zeros and the number of poles, respectively, inside . The following equalities hold
dg
ds (s)

1
2

g(s)

ds =

1
arg(g(s))| = N () P ().
2

(A.5)

Furthermore, N () P () equals the number of times that {g(s) | s } winds around the
origin as s transverses once in a counterclockwise sense.

120

A.1. Complex analysis

Proof See theorem 6.1 in Levison and Redheer [17] or theorem 10.43 in Rudin [24].
We would like to apply this theorem to the imaginary axis, but this is not a closed curve. To
overcome this, we introduce an extra assumption on the functions.
Theorem A.1.14. Nyquist Theorem. Let g be a function that is meromorphic on an open
set containing C+ and suppose that g has no poles or zeros on the imaginary axis. Furthermore,
0
we assume that g has a nonzero limit at in C+ ; that is, there exists a g() C, g() = 0
0
such that

lim

sup

{sC+ ||s|>}
0

|g(s) g()| = 0.

(A.6)

Then g has at most nitely many poles and zeros in C+ and

0
1
2

dg
ds ()

g()

1
lim [arg(g(j)) arg(g(j))]
2

= N0 P0 ,

(A.7)

where N0 and P0 are the number of zeros and poles, respectively, in C+ . Furthermore, N0 P0
0
equals the number of times that {g() | R} winds around the origin as decreases from
+ to .
Proof This follows from Theorem A.1.13 by a limiting argument.
This theorem can be extended to allow for isolated poles or zeros on the imaginary axis in the
following manner.
If g has a pole or a zero at 0 , then we integrate around this point via the half-circle in C :
0
C0 = {s C | s = 0 e ; < < , > 0}, and the principle of the argument also
2
2
applies for this indented imaginary axis. Notice that the crucial requirement in Theorem A.1.14
has been the limit behavior of g as |s| in C+ .
0
This last version of the principle of the argument (A.7) motivates the following concept of
the Nyquist index of a meromorphic, scalar, complex-valued function. As we have already
noted, meromorphic functions have isolated poles and zeros (see Denitions A.1.1 and A.1.9 and
Theorem A.1.4.b).
Denition A.1.15. Let g be a function that is meromorphic on C+ for some > 0 and suppose

that g has a nonzero limit at in C+ (see (A.6)). This implies that the graph of g(s) traces
0
out a closed curve in the complex plane, as s follows the indented imaginary axis. We dene the
number of times the plot of g(s) encircles the origin in a counterclockwise sense as s decreases
from to over the indented imaginary axis to be its Nyquist index, which we denote by
ind(g). Thus, by Theorem A.1.14 we have that
ind(g) =

1
lim [arg(g(j)) arg(g(j))] = N0 P0 .
2

(A.8)

If g has no poles or zeros on the imaginary axis, then the Nyquist index is just the number of
times the plot of g() encircles the origin in a counterclockwise sense as decreases from to
.

121

A. Mathematical Background
In complex analysis books, the index for a curve is normally dene as a winding number. Note
that our Nyquist index is the winding number of the curve g(s) with s on the indented imaginary
axis.
From the properties of the argument, it follows that the Nyquist index has a similar property:
ind(g1 g2 ) = ind(g1 ) + ind(g2 ).

(A.9)

The Nyquist index is a homotopic invariant, which basically means that deforming the closed
curve g(R) does not change the index, provided that the curve remains closed and does not pass
through the origin. We recall the denition of homotopic maps.
Denition A.1.16. Let X be a topological space and let 1 = 1 ([0, 1]), 2 = 2 ([0, 1]) be two
closed curves in X. 1 and 2 are X-homotopic if there exists a continuous map : [0, 1][0, 1]
X such that
(y, 0) = 1 (y),
(y, 1) = 2 (y),
(0, t) = (1, t)
(A.10)
for all y, t [0, 1].

Theorem A.1.17. If 1 and 2 are C \ {0}-homotopic closed contours in the domain C \ {0},
then the number of times that 1 and 2 wind around 0 is the same.
Proof This follows from theorem 10.40 in Rudin [24].
We apply this theorem to show that the indices of two functions that can be continuously
transformed from the rst into the second have the same Nyquist index.
Lemma A.1.18. Let g1 and g2 be meromorphic functions on an open set containing C+ , with
0
nonzero limits g1 () and g2 () at innity in C+ . If there exists a continuous function h(s, t) :
0
(, ) [0, 1] C such that h(, 0) = g1 (), h(, 1) = g2 () and h(, t) and h(, t)
are nonzero for all t [0, 1] and R, then the Nyquist indices of g1 and g2 are the same.
Proof First we suppose that neither g1 nor g2 has poles or zeros on the imaginary axis. For
t [0, 1] and y (0, 1) we dene (y, t) := h( tan(y ), t), 1 (y) := g1 ( tan(y )) and
2
2
2 (y) := g2 ( tan(y )). Furthermore, we dene the end point of (, t) by (0, t) = (1, t) =
2
h(, t) and the end points of 1 , 2 by 1 (0) = 1 (1) = g1 () and 2 (0) = 2 (1) = g2 ().
By Denition A.1.16 we easily see that the closed curves 1 ([0, 1]) and 2 ([0, 1]) are C \ {0}homotopic, and so by Theorem A.1.17 the number of encirclements of 0 are the same. Since
these curves are the same as g1 () and g2 (), respectively, we have by Denition A.1.15 that
their Nyquist indices are the same.
The proof for the case that g1 or g2 has poles and zeros on the imaginary axis is similar,
replacing the imaginary axis with the indented version.

A.2. Normed linear spaces

The results in this section are well known in functional analysis and may be found in almost any
book on this subject. The basic source is Kreyszig [16]; secondary sources are Kato [14], Naylor
and Sell [19], Taylor [25], and Yosida [31].

122

A.2. Normed linear spaces

A.2.1. General theory

The concept of normed linear spaces is fundamental to functional analysis and is most easily
thought of as a generalization of the n-dimensional Euclidean vector space Rn with the euclidean
length function : Rn R+ = [0, ) given by
n

=
i=1

|xi |2 .

In fact, it is just a linear vector space with a length function (norm) dened on it. First, we
dene a linear vector space; other terms are vector space or linear space.
Denition A.2.1. A linear vector space W over a scalar eld F is a nonempty set W with
a mapping: (x1 , x2 ) x1 + x2 from W W to W , which we call addition, and a mapping:
(, x) x from F W to W , which we call scalar multiplication. These mappings satisfy the
following conditions for all x, y, z in W and all , F:
a. x + y = y + x (the commutative property);
b. (x + y) + z = x + (y + z) (the associative property);
c. There exists a unique element 0 in W such that x + 0 = x
(the existence of the zero element);
d. For each x W , there exists a unique element x W such that x + x = 0
(the existence of an inverse);
e. (x) = ()x;
f. ( + )x = x + x;
g. (x + y) = x + y;
h. 1x = x, where 1 is the unit element of the scalar eld F .

In this book, F will be either the real number eld R or the complex number eld C; W over
R is called a real vector space, and W over C is called a complex vector space.
Denition A.2.2. If W is a linear vector space over the eld F , then a subset S of W is a
linear subspace if x, y S x + y S for all scalars , F (i.e., S is closed under addition
and scalar multiplication and so is itself a linear vector space over F ).

Denition A.2.3. A linear combination of vectors x1 , . . . , xn of a linear vector space W is an

expression of the form 1 x1 + 2 x2 + . . . n xn , where the coecients 1 , . . . , n are scalars.
Denition A.2.4. For any nonempty subset M of the linear vector space W , the set of all
linear combinations of vectors of M is called span of M and is denoted by
span{M }.

Obviously, this is a linear subspace Y of W , and one can easily show that it is the smallest
(with respect to inclusion) linear subspace that contains M . We say that Y is spanned by M .
Denition A.2.5. If x1 , . . . , xn are elements of W , a linear vector space over F , and there exist
scalars 1 , . . . , n , not all zero, such that the linear combination 1 x1 + . . . + 2 xn = 0, then we
say that x1 , . . . , xn is a linearly dependent set. If no such set of scalars exist, then we say that
x1 , . . . , xn are linearly independent.

123

A. Mathematical Background
Denition A.2.6. If the linear vector space W is the span of a nite set of linearly independent
vectors x1 , . . . , xn , then we say that W has dimension n. If there exists no nite set M of vectors,
such that W =span{M }, W is said to be innite-dimensional.

Denition A.2.7. A norm is a nonnegative set function on a linear vector space,

R+ = [0, ), such that:
a.

x = 0 if and only if x = 0;

b.

x + y x + y for all x, y W (the triangular inequality);

c.

x = || x for all x W and F.

:W

Denition A.2.8. A normed linear space is a linear vector space X with a norm X on it
and it is denoted by (X, X ). If the meaning is clear from the context, we usually write simply
X and use for the norm.

Example A.2.9 Let p 1 be a xed real number. By denition, each element in the space p
is a sequence x = (j ) = (1 , 2 , . . .) of numbers in C such that

j=1

|j |p < .

This is a linear vector space over C with componentwise addition and scalar multiplication.
It is also a normed linear space with the norm

x =

j=1

1/p

|i |p

The triangular inequality for the space p is commonly know as the Minkowski inequality for
sums.
Example A.2.10 The space consists of all sequences x = (1 , 2 , . . .), where i C and
sup |i | < . This is a linear vector space over C with componentwise addition and scalar
i1

multiplication. Furthermore, it is a normed linear space with the norm

x = sup |i |.
i1

Example A.2.11 Let p 1 be a xed real number and let a < b . Consider the set
b

of measurable functions x(t) with

a

|x(t)|p dt nite, and with the norm

x =

1/p

|x(t)|p dt

This is a linear vector space with addition and scalar multiplication dened by:
(x + y)(t) = x(t) + y(t);

124

A.2. Normed linear spaces

(x)(t) = x(t).
However, it is not a normed linear space, since x = 0 only implies that x(t) = 0 almost
everywhere. To make it into a normed linear space we have to consider (equivalence) classes of
functions, [x], where [x] is the class of all functions that equal x almost everywhere. Clearly,
these equivalence classes form a linear space and [x] := x1 for any x1 [x] denes a norm;
we call this normed linear space Lp (a, b). Following usual practice, we write x1 instead of [x],
where x1 is any element of the equivalence class [x].
The triangular inequality for Lp (a, b) is called the Minkowski inequality for functions.
Example A.2.12 Let a < b and consider all measurable functions x from (a, b) to
C with the property that esssupt(a,b)|x(t)| < . As in Example A.2.11, we form equivalence
classes [x] that contain functions that equal x almost everywhere on (a, b). With the norm
[x]

:= ess supt(a,b) |x1 (t)|

for any x1 [x],

this space is a normed linear space, which we denote by L (a, b). As in Example A.2.11, we
usually write x1 instead of [x], where x1 is any element of [x].
Denition A.2.13. Let (X, X ) and (Y, Y ) be two normed linear spaces. Then X and
Y are topologically isomorphic if there exists a linear, bijective map T : X Y and positive
constants a, b such that
a x

Tx

b x

for all x X.

The norms X and Y are then called equivalent norms.

The normed linear spaces are isometrically isomorphic if there exists a linear, bijective map
T : X Y such that
T x Y = x X.

Denition A.2.14. A sequence {xn } in a normed linear space (X,

lim

xn x

xi is said to converge to x, if the sequence

i=0

converges to x if

= 0.

The series

X)

i=0

xi converges to x as n .

Denition A.2.15. A set V in a normed linear space X is closed if every convergent sequence
in V has its limit point in V . A set V is open if its complement is closed. Alternatively, a set
V is open if for any point x V , there exists an > 0 such that the sphere with centre x and
radius , B(x, ) := {y X | y x < } is contained entirely in V .
If we add to a set V all the limit points of sequences in V , we obtain the smallest closed set
that contains V . This closed set is called the closure of V , which we write as V .
A set V in a normed linear space (X, X ) is bounded if sup x X < . A set V in a normed
xV

linear space is compact if every sequence in V contains a convergent subsequence with its limit
point in V ; V is relatively compact if its closure is compact.
Denition A.2.16. A subset V of a normed linear space is dense in X if its closure is equal to
X.

125

A. Mathematical Background
This important property means that every element x of X may be approximated as closely as
we like by some element v of V , i.e., for any x in X and > 0 there exists a v V such that
v x < .
All normed linear spaces have dense subsets, but they need not be countable. Normed linear spaces that do have countable dense subsets have special properties that are important in
applications.
Denition A.2.17. A normed linear space (X,
that is countable.

X)

is separable if it contains a dense subset

The concept of Cauchy sequence in R is very important, since even without evaluating the
limit one can determine whether a sequence is convergent or not. We shall start by generalizing
the concept of Cauchy sequences to general normed linear spaces.
Denition A.2.18. A sequence {xn } of elements in a normed linear space (X,
sequence if
xn xm X 0,
as n, m .

X)

is a Cauchy

As stated above, every Cauchy sequence in R is convergent. Unfortunately, this does not hold
for general normed linear spaces, as can be seen from the next example.
Example A.2.19 Let X = C[0, 1], the space of continuous functions on [0, 1] and as a norm we
1

take x = ( |x(t)|2 dt)1/2 . Now consider the sequence of functions {xn } X given by
0

xn (t) =

nt

2
1

n
4

1
2

1
for 0 t 1 n
2
1
for 1 n t 1 +
2
2
1
1
for 2 + n t 1.

1
n

{xn } is Cauchy, since for n > m we have that

1

xm xn

=
0

|xm (t) xn (t)|2 dt

1
1
2n

mt m 1 2

+ ) dt +
2
4
2

1
1
2m

1
1
2+n

mt m nt n 2

+ ) dt+
2
4
2
4

1
1
2n

1
1
2+m

mt m 1 2

) dt
2
4
2

1
1
2+n

=
since n > m.
Thus xm xn

1 m
1
1
1
1
2 + ]
[

,
6 n2
n m
6m 6n

0 as m, n . Clearly, the pointwise limit of xn is

x(t) =

0 for 0 t <

1 for

1
2

1
2

< t 1.

However, this function is not in C[0, 1], because of the discontinuity at t = 1 .

2

126

A.2. Normed linear spaces

This situation is clearly unsatisfactory and we prefer to work with spaces where Cauchy sequences always have limits in the same space. A normed linear space X is complete if every
Cauchy sequence has a limit in X.
Denition A.2.20. A Banach space is a complete, normed linear space.

The spaces p , , Lp (a, b), and L (a, b) introduced in Examples A.2.9 to A.2.12 are all
Banach spaces.
Example A.2.19 was an example of a normed linear space that was not complete. Under a
dierent norm it is complete.
Example A.2.21 Consider the space C[0, 1] of continuous functions on [0, 1] and dene the sup
norm
x := sup |x(t)|.
t[0,1]

Clearly, with C[0, 1] denes a normed linear space. It is complete, since it is known that
uniformly convergent sequences of continuous functions converge to a continuous function. Hence
C[0, 1] is a Banach space under this norm.
Another way of producing a complete normed linear space is given in the next theorem.

Theorem A.2.22. Let X be a normed linear space. Then there exists a Banach space X and

a linear, injective map T : X X such that T(X) is dense in X and T x X = x X for all

x X. The space X is called the completion of X.

Proof See Kreyszig [16, theorem 2.3-2] or Yosida [31, section I.10] or Taylor [25, theorem 2.41A].
It is not hard to show that the completion of C[0, 1] with the norm as in Example A.2.19 is
L2 (0, 1).

A.2.2. Hilbert spaces

A Banach space generalizes the notion of Rn as a linear space with a length function, but in
order to generalize the useful geometric property of orthogonality we need some extra structure.
Denition A.2.23. An inner product on a linear vector space Z dened over the complex or
real eld F is a map
, : Z Z F

such that for all x, y Z and , F it holds that

a. x + y, z = x, z + y, z ;

b. x, y = y, x ;
c. x, x 0 and x, x = 0 if and only if x = 0.
Properties a and b imply that x, z + y = x, z + x, y ; we say that x, z is semilinear
in z. A linear space Z with an inner product , is called an inner product space.
Using the inner product we can make an inner product space into a normed linear space
(Z, Z ) by dening the induced norm by
z

:=

z, z .

In general, Z will not be a Banach space, since it need not be complete. Complete inner
product spaces have a special name.

127

A. Mathematical Background
Denition A.2.24. A Hilbert space is an inner product space that is complete as a normed
linear space under the induced norm.

Before we look at some examples of Hilbert spaces, we list some properties of inner products
and their induced norms:
a. x, y = 0 for all x Z implies y = 0;
b. | x, y | x
c.

x+y

y , (Cauchy-Schwarz inequality);

+ xy

=2 x

+ 2 y 2, (parallelogram law);

d. If the norm in a normed linear space satises the parallelogram law, then the following
denes an inner product
x, y =

1
4

x+y

xy

+ x + y

x y

and the norm is induced by this inner product.

Example A.2.25 The spaces 2 and L2 (a, b) dened in Examples A.2.9 and A.2.11, respectively,
are Hilbert spaces under the inner products
b

x, y

xn yn

:=

and

x, y

L2

:=

n=1

x(t)y(t)dt,
a

respectively. As in Example A.2.11, by x we really mean the equivalence class [x]. We remark
that the much used Cauchy-Schwarz inequality on L2 (a, b) becomes
b

b
2

x(t)y(t)dt|

b
2

|x(t)| dt

|y(t)|2 dt.

(A.1)

Using the Cauchy-Schwarz inequality one can show that functions in L2 (0, ) with their
derivative in L2 (0, ) have zero limit at innity.
Example A.2.26 Let f be an element of L2 (0, ), and assume that f is dierentiable with its
derivative in L2 (0, ). Then for all t > s we have that
t
2

|f (t)| |f (s)|

=
s

d
|f ( )|2 d =
dt
t

2
s

f( )f ( )d

f ( )f( )d +
s

|f ( )|2 d

|f( )|2 d .

Since f, f are elements of L2 (0, ) we see that |f (t)| converges for t . Using the fact that
f is square integrable, we see that its limit can only be zero.
We now illustrate how it is possible to dene several inner products on the same linear vector
space.

128

A.2. Normed linear spaces

Example A.2.27 Consider L2 (a, b) dened above with < a < b < and dene the
subspace
Z := {u L2 (a, b) | u is absolutely continuous on (a, b)
with du L2 (a, b) and u(a) = 0}.
dt
We remark that an element in L2 (a, b) is said to be absolutely continuous if there is an absolutely
continuous function in the equivalence class (see Example A.2.11). One can easily show that there
can at most be one absolutely continuous function in every equivalence class.
Z can be regarded as a subspace of L2 (a, b) and it is in fact a dense subspace. On the other
hand, we can introduce a dierent norm that is well dened for all u, v Z
u, v

du dv
,
dt dt

L2 (a,b) .

With the above inner product we obtain the new Hilbert space Z2 .
The above example brings us naturally to the following class of Hilbert spaces (see Yosida [31,
sections I.9 and I.10] or Naylor and Sell [19, section 5.13]).
Denition A.2.28. For < a < b < we dene the following subspace of L2 (a, b)
m
S2 (a, b) :=

m1

u
{u L2 (a, b) | u, . . . , d m1 are absolutely
dt

continuous on (a, b) with

dm u
dtm

L2 (a, b)}.

This is a Hilbert space with respect to the inner product

m

z1 , z2

m
S2 (a,b)

=
n=0

dn z1 dn z2
,
dtn dtn

L2 .

(A.2)

These Hilbert spaces are called Sobolev spaces1 .

m
One can show that S2 (a, b) is the completion of C m [a, b] or C [a, b] with respect to the norm
induced by (A.2) (see Yosida [31, sections I.9 and I.10] or Naylor and Sell [19, section 5.13]). It
m
is not dicult to show that S2 (a, b) is topologically isomorphic to
m1

u
{u L2 (a, b) | u, . . . , d m1 are absolutely continuous on (a, b)
dt

with

dm u
dtm

L2 (a, b)}

under the inner product

z1 , z2 = z1 , z2

L2

dm z1 dm z2
,
dtm dtm

L2 .

(A.3)

The inner product structure allows a simple generalization of the concept of orthogonality.
Denition A.2.29. We say that two vectors x and y in a Hilbert space Z are orthogonal if
x, y = 0,
in which case we write x y.
1

Another notation for S m is H m . However, in this book we use H m for the Hardy spaces.

129

A. Mathematical Background
If x y, then the parallelogram law reduces to a generalized statement of Pythagoras theorem,
namely,
x + y 2 = x 2 + y 2.
Denition A.2.30. If V is a subspace of a Hilbert space Z, then the orthogonal complement
V is dened by
V = {x Z | x, y = 0 for all y V }.

It can be shown that V is a closed linear subspace of Z and that Z can be uniquely decomposed as the direct sum
Z = V V ,
(A.4)

where V is the closure of V . This means that any z Z has the unique representation
z = zV + zV ,

where zV V , zV V , and z 2 = zV 2 + zV 2 .
Furthermore, we see that a subspace V is dense in the Hilbert space Z if and only if V = {0}.
Denition A.2.31. An orthonormal set in a Hilbert space Z is a nonempty subset {n , n 1}
of Z such that
1 if n = m
n , m = nm :=
0 if n = m.

In other words, {n , n 1} are mutually orthogonal unit vectors. Of course, any mutually
orthogonal set {xn , n 1} may be normalized by dening new vectors
xn
.
xn

n =

Denition A.2.32. Let {en , n 1} be a subset of Z. We say that it is maximal if

spann1 {en } = Z.

In Rn any element can be expressed as a linear combination of any set of n mutually orthonormal elements; such an orthonormal set is called a basis. For innite-dimensional Hilbert spaces
we have a similar property.
Denition A.2.33. We say that an orthonormal sequence of a separable Hilbert space Z is an
orthonormal basis if it is maximal. Then for any x Z, we have the Fourier expansion

x=

x, n n .
n=1

The terms x, n are called the Fourier coecients of x with respect to n . Furthermore, we
have the important Parseval equality. Any two vectors x, y in Z satisfy

x, n y, n .

x, y =
n=1

In particular, for x = y we have

=
n=1

130

| x, n |2 .

A.2. Normed linear spaces

Example A.2.34 Consider the Hilbert space L2 (0, 1) introduced in Examples A.2.11 and A.2.25.
This space has several dierent orthonormal bases. The best known one is the following:

{1, 2 sin(2nt), 2 cos(2nt), n 1}.

(A.5)
This is an orthonormal basis, so any x L2 (0, 1) may be represented in the form

x(t) = a0 + 2

an cos(2nt) + 2

n=1

bn sin(2nt),
n=1

where
1

a0

x(t)dt,

=
0
1

an

=
0
1

bn

=
0

x(t) 2 cos(2nt)dt for n 1,

x(t) 2 sin(2nt)dt for n 1.

This is the classical Fourier expansion, and an and bn are the Fourier coecients.
Other orthonormal bases are given by the sequences

{1, 2 cos(nt), n 1},

see Example A.4.22, and

{ 2 sin(nt), n 1},

(A.6)
(A.7)

see Example A.4.21. A fourth orthonormal basis is given by the Legendre polynomials
Pn (t) =

8n dn 2
(t t)n , n 1.
2 n! dtn

We remark that the expansions in Example A.2.34 are not valid pointwise, but only in the
sense of the L2 (0, 1) norm. For example, equality in the Fourier expansion means that
N

x, i i 0 as N .

i=1

Example A.2.35 Let {n , n 1} be an orthonormal basis of the Hilbert space Z, and let
{n , n 1} be a positive sequence with n 1. Now we dene the following linear subspace of
Z.

Z := {z Z | z =

zn n , with
n=1

n=1

n |zn |2 < }.

(A.8)

It is clear that Z is a dense, linear subspace of Z. On this linear vector space we dene the
following inner product

z1 , z2

n z1 , n z2 , n .

:=

(A.9)

n=1

Under this inner product, the space Z is a Hilbert space.

131

A. Mathematical Background
A recurring problem in innite dimensions is the question of approximation. For example, we
N

may ask how good an approximation

x, n n is to x and how one should improve upon

n=1

this approximation by introducing extra terms. It turns out that there is a simple answer to
this question if {n , n 1} is an orthonormal basis in the Hilbert space Z. It is based on the
following generalization of dropping a perpendicular in the three-dimensional Euclidean space.
Theorem A.2.36. Let Z be a Hilbert space and V a closed subspace of Z. Then, given x Z,
there exists a unique v0 V such that
x v0 = min x v .
vV

Furthermore, a necessary and sucient condition for v0 V to be the minimizing vector is that
(x v0 ) V .
Proof See Kreyszig [16, theorem 3.3-1 and lemma 3.3-2] or Naylor and Sell [19, theorem 5.14.4].

Notice that in the above theorem the vector x v0 is to be seen as that obtained by dropping
a perpendicular onto V . We now apply this theorem to the approximation problem. Let
1 , 2 , . . . , N be an orthonormal sequence of vectors that span a nite-dimensional subspace V .
For any given x Z we seek the vector x in V such that x x is minimized. By Theorem

A.2.36, we see that

x x, n = 0, n = 1, 2, . . . , N.

Supposing that x =

n n , the above equality implies that

n=1

x, n = n .
So the best estimate of any vector x Z using N orthonormal vectors n , n = 1, . . . , N is x

given by
N

x=

x, n n .
n=1

To improve this estimate, all that is necessary is to add an extra term x,N +1 N +1 . We remark
that this would not be the case if the sequence {n } were not orthonormal; then it would be
necessary to recalculate all of the coecients every time a better approximation were required.

A.3. Operators on normed linear spaces

The theory of operators on normed linear spaces is treated in any introductory book on functional
analysis and most of the denitions, lemmas, theorems, and examples in this section are standard;
useful references are Kato [14], Kreyszig [16], Rudin [23], and Naylor and Sell [19].

A.3.1. General theory

In this section we shall be concerned with transformations T from one normed linear space X to
another Y . Usually, X and Y will be either Banach or Hilbert spaces and T will be linear. Later
in this section we treat the special case where Y is the scalar eld F ; there the transformations
are called functionals. We start with the following xed-point theorem.

132

A.3. Operators on normed linear spaces

Theorem A.3.1. Contraction Mapping Theorem. Let X be a Banach space, T a mapping
from X to X, m N, and < 1. Suppose that T satises T m (x1 ) T m (x2 ) x1 x2 for
all x1 , x2 X. Then there exists an unique x X such that T (x ) = x ; x is the xed point
of T .
Furthermore, for any x0 X the sequence {xn , n 1} dened by xn := T n (x0 ) converges to

x as n .
Proof See Kreyszig [16, theorem 5.4-3] or Naylor and Sell [19, theorem 3.15.2 and corollary
3.15.3].
In above theorem the mapping T does not need to be linear. However, in the rest of this
section we only consider linear transformations.
Denition A.3.2. A linear operator, or simply an operator, T from a linear space X to a linear
space Y over the same eld F is a map T : D(T ) X Y , such that D(T ) is a subspace of X,
and for all x1 , x2 D(T ) and scalars , it holds that
T (x1 + x2 ) = T x1 + T x2 ,
T (x1 ) = T x1 .

It follows immediately from this denition that if i F and xi D(T ) for i = 1, . . . , n, then
n

i T xi .

i xi ) =

T(

i=1

i=1

The set D(T ) is called the domain of T . In fact, changing the domain changes the operator; for
example, the operator T1 : D(T1 ) = {x L2 (0, 1) | x continuous} L2 (0, 1), T1 x = 2x diers
from the operator T2 : L2 (0, 1) L2 (0, 1), T2 x = 2x.
Example A.3.3 It is easy to see that the following mappings are all linear operators:
the shift operator dened by
: 2 2 ,
where
((x))n = xn+1 ;
the integral operator
Tg : L2 (0, 1) L2 (0, 1)
dened by
1

Tg f =

f (t)g(t)dt
0

for a g L2 (0, 1);

the dierentiation operator

T : D(T ) = C 1 (0, 1) L2 (0, 1) L2 (0, 1)
dened by
Tf =

df
.
dx

133

A. Mathematical Background
Denition A.3.4. The set of all possible images of the operator T : D(T ) Y is a subspace of
Y , in general. It is called the range of T and we denote this by ran T . If the range of an operator
is nite-dimensional, then we say that the operator has nite rank.

Operators for which the domain and the range are in one-to-one correspondences are called
invertible.
Denition A.3.5. A operator T : D(T ) X Y between two linear spaces X and Y is
invertible if there exists a map S : D(S) := ran T Y X such that
ST x =
T Sy =

x, x D(T ),
y, y ran T.

S is called the algebraic inverse of T and we write T 1 = S.

Lemma A.3.6. Linear operators T from X to Y , where X and Y are linear vector spaces, have
the following properties:
a. T is invertible if and only if T is injective, that is, T x = 0 implies x = 0;
b. If T is an operator and it is invertible, then its algebraic inverse is also linear.
Proof See Kreyszig [16, theorem 2.6-10] or Kato [14, section III.2].
The set of all elements in the domain of T such that T x = 0 is called the kernel of T and
is denoted by ker T . If T is a linear operator, then ker T is a linear subspace. From the above
lemma we see that the linear operator T has an inverse if ker T = {0}.
The continuity of a map from one normed linear space to another is another very important
property, since it says that a small change in the original vectors gives rise to a corresponding
small change in their images.
Denition A.3.7. A map F : D(F ) X Y between two normed linear spaces (X, X )
and (Y, Y ) is said to be continuous at x0 X if, given > 0, there exists a > 0 such that
F (x) F (x0 ) Y < , whenever x x0 X < . F is continuous on D(F ) if it is continuous at
every point in D(F ).

Denition A.3.8. Let T be a linear operator from D(T ) X Y , where X and Y are normed
linear spaces. T is a bounded linear operator or T is bounded if there exists a real number c such
that for all x D(T )
T x Y c x X.

The above formula shows that a bounded linear operator maps bounded sets in D(T ) into
bounded sets in Y , and it leads naturally to the following denition of a norm.
Denition A.3.9. Let T be a bounded linear operator from D(T ) X to Y . We dene its
norm, T , by
Tx Y
.
T = sup
x X
xD(T ),
x=0

134

A.3. Operators on normed linear spaces

If D(T ) = {0}, we dene T = 0, since in this uninteresting case from denition A.3.2 we
have that T 0 = 0.
An equivalent denition of T is
T = sup

Tx

Y.

xD(T )
x X =1

This norm satises the conditions for a norm (see denition A.2.7). An automatic consequence
of denition A.3.9 is that
T x Y T x X;
(A.1)
this result will be used frequently.
Continuity and boundedenitioness are equivalent concepts for linear operators.
Theorem A.3.10. If T : D(T ) X Y is a linear operator, where X and Y are normed
linear spaces, then:
a. T is continuous if and only if T is bounded;
b. If T is continuous at a single point, it is continuous on D(T ).
Proof See Kato [14, section III.2], Kreyszig [16, theorem 2.7-9], Naylor and Sell [19, theorem
5.6.4 and lemma 5.6.5], or Rudin [23, theorem 1.32].
Bounded linear operators that map into a Banach space always have a unique extension to
the closure of their domain.
Theorem A.3.11. Let T : D(T ) X Y be a bounded linear operator, where X is a normed

linear space and Y is a Banach space. Then T has a unique bounded extension T : D(T ) Y .
= T .
Furthermore, T
Proof See Kato [14, theorem 1.16] or Kreyszig [16, theorem 2.7-11].
Of special interest are bounded linear operators whose domain is a normed linear space.
Denition A.3.12. If X and Y are normed linear spaces, we dene the normed linear space
L(X, Y ) to be the space of bounded linear operators from X to Y with D(T ) = X and with
norm given by denition A.3.9.
If it is necessary to distinguish between various norms, we shall write the norm as L(X,Y ) .
For the special case that X = Y we denote L(X, X) by L(X). First we consider L(X, Y ), where
X and Y are nite-dimensional spaces.
Example A.3.13 Recall that matrices with k rows and m columns are linear mapping from
Cm to Ck . If we take the norm on Ck and Cm to be the Euclidian norm, then it is easy to see
that this mapping is also bounded. We shall calculate the exact norm. Let T be a k m-matrix.
Since the matrix T T is symmetric and nonnegative, we have that
m
2
i x, i i ,

T T x =

(A.2)

i=1

135

A. Mathematical Background
2
where {i , 1 i m} is an orthonormal basis of Cm and i are the eigenvalues of T T . i are
the singular values of T . Without loss of generality, we assume that 1 2 . . . m 0.
Since {i , 1 i m} is an orthonormal basis, we have that
m

=
i=1

| x, i |2 .

(A.3)

Now from (A.2), we deduce

m

Tx

2
i x, i i , x

T x, T x = T T x, x =
i=1

But T 1

2
1

using equation (A.3).

2
= T T 1 , 1 = 1 , and so T = 1 .

In the next lemma, we summarize some properties of the space of linear bounded operators,
L(X, Y ).
Lemma A.3.14. Let L(X, Y ) denote the space of bounded linear operators from X to Y . Then
the following properties hold:
a. If Y is a Banach space, then so is L(X, Y );
b. If X, Y , and Z are normed linear spaces, T1 L(X, Y ) and T2 L(Y, Z), then T3 , dened
by T3 x = T2 (T1 x), is an element of L(X, Z) and T3 T2 T1 ;
c. For the special case that X = Y , L(X) is an algebra; that is, T1 , T1 + T2 and T1 T2 are
in L(X) for every T1 , T2 in L(X); furthermore, T1 T2 T1 T2 .
Proof a. Proofs are given in the following texts: Kato [14, section III.3.1], where the notation
B(X, Y ) is used instead of L(X, Y ); Kreyszig [16, theorem 2:10-2]; Naylor and Sell [19, theorem
5.8.6], where Blt[X, Y ] is used instead of L(X, Y ); Rudin [23, theorem 4.1], where the notation
B(X, Y ) is used instead of L(X, Y ); Taylor [25, theorem 4.1-A], where [X, Y ] is used for L(X, Y ).
b. See Kreyszig [16, section 2.7, equation (7)], Yosida [31, proposition I.6.2], or Naylor and Sell
[19, theorem 5.8.4], where the last reference uses Blt[X, Y ] instead of L(X, Y ).

c. See Kreyszig [16, section 2.10] or Taylor [25, theorem 4.1-B], where [X] is used instead of
L(X).

Example A.3.15 Consider the Banach space Z with norm Z and let W be a linear subspace
of Z. Suppose that another norm, W , is also dened on W and that W is a Banach space
under this norm.
Consider the linear operator from W to Z dened by
iw = w,
where on the left-hand side w is seen as an element of W and on the right-hand side as an element
of Z. This mapping is called a continuous embedding if the operator i is an element of L(W, Z).
In this case, we have that
w Z c w W
(A.4)

136

A.3. Operators on normed linear spaces

for some positive constant c. If W is a dense subspace of Z (with respect to the norm Z ), we
call i a dense injection. In the case that i is a continuous dense injection, we use the notation
W Z.

(A.5)

Let us now take W to be the Hilbert space Z with the norm induced by ,
It is easy to show that W is contained in Z with continuous, dense injection
Z Z.

given by (A.9).
(A.6)

It is possible to introduce several dierent notions of convergence in the space of bounded

linear operators. The natural one based on the norm in L(X, Y ) is called uniform convergence,
but this is a very strong property. Consequently, we nd the following weaker concept very
useful.
Denition A.3.16. Let {Tn , n 1} be a sequence of bounded linear operators in L(X, Y ),
where X and Y are normed linear spaces. If
Tn x T x

0 as n for all x X,

then we say that Tn converges strongly to T .

Frequently, the bounded linear operator will depend on a parameter t, where t is usually from
some interval in R. We can dene strong continuity and uniform continuity with respect to t in
an analogous manner.
Denition A.3.17. If T (t) is in L(X, Y ) for every t [a, b], where X and Y are normed linear
spaces, then
a. T (t) is uniformly continuous at t0 , if
T (t) T (t0 )

L(X,Y )

0 as t t0 ;

b. T (t) is strongly continuous at t0 , if

T (t)x T (t0 )x

0 for all x X as t t0 .

Using this notion of continuity, we can dene the following linear space.
Denition A.3.18. Let X be a normed linear space, and suppose that a < b . Let
f be a function from [a, b] to X that satises
f (s) f (s0 )

0, as s s0

for all s0 [a, b]. This function is called continuous and we denote by C([a, b]; X) the space of
continuous functions from [a, b] to X. It is easy to show that C([a, b]; X) is a linear space.
Combining denitions A.3.17 and A.3.18 we see that T (t) L(X) is strongly continuous if
and only if T (t)x C([a, b]; X) for every x X.
There are two very important theorems on linear operators that are used frequently in applications.

137

A. Mathematical Background
Theorem A.3.19. The Uniform Boundedenitioness Theorem (Banach Steinhaus Theorem). Let {Tn } be a family of bounded linear operators in L(X, Y ), where X is a Banach space
and Y a normed linear space. If the family {Tn x} is bounded for each x (that is,
Tn x

Mx ,

where Mx depends on x, but is independent of n), then { Tn } is uniformly bounded in n.

Proof See Kato [14, theorem III.1.26], Kreyszig [16, theorem 4.7-3], Rudin [23, theorem 2.5],
Taylor [25, theorem 4.4-E], or Yosida [31, corollary II.1.1].
Theorem A.3.20. The Open Mapping Theorem. Let T L(X, Y ), where both X and Y
are Banach spaces and T maps X onto Y . Then T maps every open set of X onto an open set
of Y .
Proof See Kreyszig [16, theorem 4.12-3], Rudin [23, corollary 2.12(a)], or Yosida [31, theorem
in section II.5].
A special subclass of bounded linear operators with useful properties is the following.
Denition A.3.21. Let X and Y be normed linear spaces. An operator T L(X, Y ) is said
to be a compact operator if T maps bounded sets of X onto relatively compact sets of Y . An
equivalent denition is that T is linear and for any bounded sequence {xk } in X, {T xk } has a
convergent subsequence in Y .
Compact operators have properties rather similar to those enjoyed by operators on nitedimensional spaces.
Lemma A.3.22. Let X and Y be normed linear spaces and let T : X Y be a linear operator.
Then the following assertions hold:
a. If T is bounded and dim(T (X)) < , then the operator T is compact;
b. If dim(X) < , then the operator T is compact;
c. The range of T is separable if T is compact;
d. If S, U are elements of L(X1 , X) and L(Y, Y1 ), respectively, and T L(X, Y ) is compact,
then so is U T S;
e. If {Tn } is a sequence of compact operators from X to the Banach space Y , that converge
uniformly to T , then T is a compact operator;
f. The identity operator, I, on the Banach space X is compact if and only if dim(X) < ;
g. If T is a compact operator in L(X, Y ) whose range is a closed subspace of Y , then the
range of T is nite-dimensional.
Proof a. See Kreyszig [16, theorem 8.1-4(a)], Naylor and Sell [19, theorem 5.24.3], or Rudin
[23, theorem 4.18(a)].
b. See Kreyszig [16, theorem 8.1-4(b)].
c. See Kato [14, theorem III.4.10], Kreyszig [16, theorem 8.2-3], or Taylor [25, theorem 5.5-A].
d. See Kato [14, theorem III.4.8], Naylor and Sell [19, theorem 5.24.7], Rudin [23, theorem
4.18(f)], or Yosida [31, part (ii) of the theorem in section X.2].

138

A.3. Operators on normed linear spaces

e. See Kato [14, theorem III.4.7], Kreyszig [16, theorem 8.1-5], Naylor and Sell [19, theorem
5.24.8], Rudin [23, theorem 4.18(c)], or Yosida [31, part (iii) of the theorem in section X.2].
f. See Kreyszig [16, lemma 8.1-2(b)].
g. See Rudin [23, theorem 4.18(b)].
Parts a and e of this lemma are extremely useful in proving the compactness of operators, as
seen in the next example.
Example A.3.23 Let X = 2 and consider T : 2 2 dened by
T x = (x1 ,

x2 x3
, , . . .).
2 3

Clearly, T is linear. Now dene Tn by

Tn x = (x1 ,

x2
xn
,...,
, 0, 0 . . .).
2
n

For every n, this operator is clearly linear and bounded and it has nite-dimensional range. So
from Lemma A.3.22.a. we see that Tn is a compact operator. Now we prove that Tn converges
uniformly to T

T x Tn x

1
1
|xi |2
i2
(n + 1)2
i=n+1

i=n+1

|xi |2

1
x 2.
(n + 1)2

So we have that

1
,
n+1
and Tn converges uniformly to T . Lemma A.3.22.e shows that T is compact.
Tn T

An important class of compact operators on the space L2 (a, b) are the integral operators.
Theorem A.3.24. Let k(t, s) be an element of L2 ([a, b] [a, b]). Then the operator K from
L2 (a, b) to L2 (a, b) dened by
b

(Ku)(t) =

k(t, s)u(s)ds
a

is a compact operator.
Proof See Naylor and Sell [19, example 6 of section 5.24], Taylor [25, example 2 of 5.5], or
Yosida [31, example 2 of section X.2].
We now consider linear operators from a normed linear space X to F , the scalar eld of X.
Denition A.3.25. A linear functional f is a linear operator from D(f ) X, a normed linear
space, to F , the scalar eld of X. Thus
f : D(f ) X F ,
where F = R if X is real and F is C if X is complex.

Denition A.3.26. A bounded linear functional, f , is a bounded linear operator from D(f )
X, a normed linear space, to F , the scalar eld of X.

139

A. Mathematical Background
Example A.3.27 Consider the following special case of the Hilbert space from Example A.2.25:
Z = {z L2 (0, 1) | z is absolutely continuous on (0, 1) with dz L2 (0, 1) and z(0) = 0} with
dt
the inner product
dz1 dz2
,
.
z1 , z2 Z =
dt dt L2 (0,1)
Dene the following functional on Z
1
f (z) = z( ).
2
Since z is absolutely continuous, this functional is well dened. We prove that it is also bounded.
1
2

|f (z)| =

1
1

|z( 2 )| = |z( 2 ) z(0)| = | z(s)ds|

1

|z(s)|ds

0
1/2

|z(s)|2 ds

= z

1
2

|z(s)|ds

Z,

where in the last inequality we have used the Cauchy-Schwarz inequality, (A.1). So f is a bounded
linear functional.
From Theorem A.3.11, we know that any bounded linear functional can be extended to the
closure of its domain without increasing its norm. The following important theorem says that
any bounded linear functional can be extended to the whole space without increasing its norm.
A consequence of this theorem is the existence of nontrivial bounded linear functionals on any
normed linear space.
Theorem A.3.28. The Hahn-Banach Theorem. Every bounded linear functional f : D(f )
F dened on a linear subspace D(f ) of a normed linear space X can be extended to a bounded
linear functional F on all X with preservation of norm.
Proof See Kato [14, theorem III.1.21], Kreyszig [16, theorem 4.3-2], Rudin [23, theorem 3.6] or
Taylor [25, theorem 4.3-A], or Yosida [31, theorem IV.5.1].
To see that this guarantees the existence of nontrivial continuous linear functionals, consider
the subspace D(f ) =span{x0 }, where x0 is an arbitrary nonzero element of X. A linear functional
f dened on D(f ) is given by
f (y) = x0

for y = x0 .

We have
|f (y)| = y ,
and so f = 1. Thus the Hahn-Banach Theorem A.3.28 says there exists an F dened on X
with F (x0 ) = x0 and norm one.
Following our previous notation we can denote all bounded linear functionals by L(X, F ), but
it is customary to use the following notation.
Denition A.3.29. The (topological) dual space of a normed linear space X is the space of all
bounded linear functionals on X with domain all of X. This space will be denoted by X .

140

A.3. Operators on normed linear spaces

Lemma A.3.30. X is a Banach space with norm
f

= sup |f (x)|.

xX
x X =1

Furthermore, we have the following duality between

x

sup
f X
f X =1

and

|f (x)|.

Proof See Kato [14, section III.1.4], Kreyszig [16, theorem 2.10-4 and corollary 4.3-4], Rudin
[23, theorem 4.3], Taylor [25, theorem 4.3-B], or theorem 1 in section IV.7 of Yosida [31].
Example A.3.31 In this example, we shall show that the dual of p is q , where
f be any element of (p ) ; then since f is linear and bounded we have

1
q

1
= 1 p . Let

f (x) =

xk k ,
k=1

where k = f (ek ), ek = (kj ); i.e., all components are zero except that in position k, which
p
equals one. Let q be p1 and consider the following sequence in p
n

(x )(k) =

|k |q
k

if k n and k = 0

if k > n or k = 0.

So

f (xn ) =
k=1

|k |q .

Now we show that {k } is a sequence in q

f (xn ) = |f (xn )| f
n

f
k=1

xn

|k |q
|k |
q

f
k=1

q1 p

(|k |

= f

|k |

1/p

n
q

|k | = f (x ) f

Dividing by the last factor and using 1

q

k=1

1
p

|k |

k=1

|k |

1
= q , we obtain

1/q

k=1

Hence

k=1

1/p

1/p

p
p 1/p

1
1 p

n
q

=
k=1

|k |

f .

141

A. Mathematical Background
Since n is arbitrary, we have
1/q

f .

|k |

=
k=1

Thus (k ) q .
Conversely, for any y = (yn ) q we get a bounded linear functional on p , if we dene

g(x) =

xk yk .
k=1

Then g is linear, and the boundedenitioness follows from the Hlder inequality
o

1/p

k=1

xk yk |

k=1

1/q

|xk |

k=1

|yk |

(A.7)

So, nally, (p ) = q .
The above results can be extended to the Lebesgue spaces to obtain
(Lp (a, b)) = Lq (a, b),
where

1
p

1
q

= 1 and 1 < p < .

Example A.3.32 Consider the Hilbert space Z as dened in A.2.35. We want to calculate the
dual space of Z .
Dene {n , n 1} by n =

1
n ,

and consider the sequences {zn , n 1} with

With respect to these sequences we dene the following linear space:

Z = {z | z =

n=1

|zn |2 n < .

zn n with
n=1

n=1

|zn |2 n < }.

(A.8)

If we dene on this (formal) space the inner product

x, y

xn yn n ,

:=
n=1

then Z is a Hilbert space. We shall show that Z can be identied with Z . Let f be any

element of Z . Since f is linear and bounded, we have

f (z) =

zn fn ,
n=1

where fn = f (n ). Let n be

1
n

and consider the following sequence of elements in Z .

(z N )(n) =

142

f n n
0

if n N

if n > N.

A.3. Operators on normed linear spaces

So
N

f (z N ) =
n=1

|fn |2 n .

Now we shall show that y =

fn n is an element of Z .
n=1
1/2

N
N

|f (z )| f

f (z ) =

= f
n=1

1/2

2
n |fn |2 n

f
n=1

|fn |2 n

We have shown that

N

1/2

N
2

n=1

|fn | n = f (z ) = f

n=1

|fn | n

and dividing by the last factor, we obtain

1/2

N
n=1

|fn |2 n

f .

Since N is arbitrary, we conclude that

1/2

n=1

f .

|fn | n

So y :=

fn n is an element of Z .
n=1

Conversely, for any y =

n=1

dening

yn n Z we obtain a bounded linear functional on Z , by

g(z) =

zn yn .

(A.9)

n=1

This g is linear, and the boundedenitioness follows from the Cauchy-Schwarz inequality

n=1

zn yn | = |

n=1

n zn

1/2

n y n |

n=1

n |zn |

1/2

2
n=1

n |yn |

So Z can be identied with the space Z .

In the previous example it is easily seen that Z Z ; this holds more generally.
Lemma A.3.33. Let X and Y be Hilbert spaces, such that X Y . Then Y X .
Proof See proposition 3 in Aubin [1, chapter 3, section 5].

143

A. Mathematical Background
Since the dual of a normed linear space X is a Banach space, we may consider the bounded
linear functionals on X , which we shall denote by X . Moreover, each element x in X gives rise
to a bounded linear functional fx in X , by

fx (f ) = f (x), f X .

It can be shown that the map x fx is an isometric isomorphism of X into X , and it is called

the natural embedding of X in X . Sometimes it happens that X equals X ; these spaces have
a special name.

Denition A.3.34. A space is reexive if its second dual X is isometrically isomorphic to X

under the natural embedding.
Examples of reexive spaces are p and Lp (a, b) for 1 < p < and Hilbert spaces; see Theorem
A.3.52.
The introduction of the space X leads to a new concept of convergence of a sequence.
Denition A.3.35. A sequence {xn } in the normed linear space X converges weakly to x if
f (xn ) f (x) as n for all f X .

Lemma A.3.36. If {xn } is a weakly convergent sequence in a normed linear space with weak
limit x, then {xn } is uniformly bounded in norm and
x lim inf xn < .
Proof See Kato [14, section III.1, equation (1.26)] or Yosida [31, section V.1, theorem 1].
The next example will show that weak convergence is indeed weaker than strong convergence.
Example A.3.37 Consider X = p , p > 1 and the sequence {xn }, where xn = (nk ). Then for
f X = q and f = (f1 , f2 , . . .) we see that
f (xn ) = fn ,
and since f q , we have that fn 0 as n . Therefore xn converges weakly to 0. However,
xn 0 q = 1, so we see that xn does not converge strongly.
From Lemma A.3.36 we see that every weakly converging sequence is bounded. The following
theorem shows that the converse is (almost) true for reexive Banach spaces.
Theorem A.3.38. A Banach space X is reexive if and only if every bounded sequence in X
contains a weakly convergent subsequence.
Proof See Yosida [31, Eberlein-Shmulyan theorem].
A consequence of this result is the following theorem.
Theorem A.3.39. Let X1 be a separable Banach space and let X2 be a reexive Banach space.
Assume further that {Tn } L(X1 , X2 ) is a sequence of uniformly bounded operators. Then
there exists a T L(X1 , X2 ) and a subsequence (n) N such that
lim f T(n) x f (T x)

for every x X1 and f X2 .

144

A.3. Operators on normed linear spaces

Proof Let {en , n N} be a basis for X1 . Without loss of generality, we assume that for
every n, en = 1. Since Tn are uniformly bounded, the sequence {Tn e1 } is also uniformly
bounded. Hence by Theorem A.3.38 there exists a subsequence (1, n) such that T(1,n) e1
converges weakly to some y1 X2 . Next we consider the sequence T(1,n) e2 . This is again
bounded; hence there exists a subsequence (2, ) (1, ) such that T(2,n) e2 converges weakly
to some y2 X2 . Repeating this argument, we obtain subsequences (i, ) and elements yi such
that (i + 1, ) (i, ) and T(i,n) ei converges weakly to yi .
If we dene (n) := (n, n), that is, the nth element of the nth subsequence, then (n) (i, )
for n > i. Hence T(n) ei converges weakly to yi . Dening the linear operator T by T ei = yi
gives
N

T(n)
i=1

i ei T

i ei ,

(A.10)

i=1

where the convergence is in the weak sense. Combining (A.10) with Lemma A.3.36 gives
T
i=1

i ei lim inf T(n)

n

i=1

i ei M

i ei ,
i=1

since {Tn } is uniformly bounded. Hence T is an element of L(X1 , X2 ) and T M .

Choose an f X2 and an x X1 . For this x there exist an N and 1 , . . . , N such that

N

i=1

i ei

.
3M f

Thus we obtain that

|f T(n) x f (T x)|

|f T(n) x f

i ei

T(n)
i=1

i ei

T(n)

|+

i=1

i ei

i=1

|f

i ei

T
i=1

f (T x)|
N

/3 + f

i ei

T(n)
i=1

i ei

+ /3.

i=1

From (A.10) it follows that the last expression is smaller than for n suciently large. is
arbitrary, and so we have proved the theorem.
On the dual spaces there exists a natural operator dual to a given operator.
Denition A.3.40. Let Q be an operator in L(X, Y ), where X and Y are Banach spaces. The
operator Q from Y to X , dened by
(Q y ) (x) = y (Qx),

(A.11)

is the dual operator of Q.

145

A. Mathematical Background
Lemma A.3.41. Let Q L(X, Y ), where X and Y are Banach spaces. The dual operator Q
of Q has the following properties:
a. Q L(Y , X ) with Q = Q ;
b. (Q) = Q .

Proof a. See Aubin [1, chapter 3, section 3, proposition 1], Kato [14, section III.3.3], Kreyszig
[16, theorem 4.5-2], Rudin [23, theorem 4.10], Taylor [25, 4.5], or Yosida [31, theorem 2 in
section VII.1].
b. See 4.5 in Taylor [25].
Until now we have concentrated mainly on bounded linear operators. However, in applications
one often comes across unbounded (not bounded) linear operators. Before we can introduce an
important class of these operators, we need the concept of the graph of a linear operator.
Denition A.3.42. Let X and Y be normed linear spaces and T : D(T ) X Y a linear
operator. The graph G(T ) is the set
G(T ) = {(x, T x) | x D(T )}
in the product space X Y .

Denition A.3.43. A linear operator T is said to be closed if its graph G(T ) is a closed linear
subspace of X Y . Alternatively, T is closed if whenever
xn D(T ), n N and lim xn = x, lim T xn = y,
n

it follows that x D(T ) and T x = y.

From this denition, we see that the domain of denition is important for an operator to be
closed. We shall illustrate this by the following example.
Example A.3.44 Let X be an innite-dimensional normed linear space, and let V be a linear
subspace of X that is not closed. If we consider the operator I on V , dened by
Ix = x

for x V,

then I is trivially bounded, but it is not closed. If we take any x in V and not in V , there exists
a sequence {xn } in V converging to x. So we have a sequence in V that converges and so does
{Ixn }. However, x is not in D(I) = V so I : V X X is not closed.
This example is rather special, since one can easily show that any bounded linear operator on
a closed domain is closed. However, there are many unbounded linear operators that are closed,
as in the following example.
Example A.3.45 Let Z be the Hilbert space L2 (0, 1) and consider the following operator on
L2 (0, 1)
d
T =
dx
with
D(T ) = {z(x) Z | z is absolutely continuous with
dz
z(0) = 0 and dx L2 (0, 1)}.

146

A.3. Operators on normed linear spaces

We show that T with this domain is closed.
Let {zn } D(T ) be a sequence such that zn z and
dz
and dx = y. Dene f by

dzn
dx

y; we must show that z D(T )

f () =

y(x)dx.
0

f is an element of D(T ) and

f z

L2 (0,1)

df
dx

= y. We show that f = z by considering

Since zn z and

dzn
dx

f zn + zn z f zn + zn z
1/2
1
zn z +

zn z +

zn z +

0
1

y(x)dx zn ()|2 d

1/2
dzn
y(x)
(x)dx|2 d
dx

1[0,]

2
L2

dzn
y
dx

1
dzn
zn z + y
.
3
dx

1/2

L2 d

y, this last expression can be made arbitrarily small, and so z = f .

In many examples, it is rather dicult to prove that an operator is closed. The next theorem
states that if the operator is the algebraic inverse of a bounded linear operator, then it is closed.
With this theorem we can more easily prove the result in Example A.3.45 (see Example A.3.47).
Theorem A.3.46. Assume that X and Y are Banach spaces and let T be a linear operator
with domain D(T ) X and range Y . If, in addition, T is invertible with T 1 L(Y, X), then
T is a closed linear operator.
Proof This follows from theorem 4.2-C of Taylor [25] with f = T 1 .
Example A.3.47 Let Z be the Hilbert space L2 (0, 1) and consider the operator of Example
A.3.45 again, i.e.,
d
T =
dx
with
D(T ) = {z(x) Z | z is absolutely continuous with
dz
z(0) = 0 and dx L2 (0, 1)}.
We show that T with this domain is closed.
Dene the following operator on Z:

z(s)ds.

(Sz)(x) =
0

It is easy to see that S L(Z) and that ST = ID(T ) and T S = IZ . So S = T 1 and from
Theorem A.3.46 we conclude that T is a closed operator.

147

A. Mathematical Background
Example A.3.48 Let Z be the Hilbert space L2 (0, 1) and consider the following operator on
Z:
d2
T = 2
dx
with domain
dz
D(T ) = {z L2 (0, 1) | z, dx are absolutely continuous
dz
d2
dz
with dx (0) = dx (1) = 0 and dxz L2 (0, 1)}.
2

Using Theorem A.3.46, we show that T with this domain is closed. Since T 1 = 0, we have that
T is not injective and thus is not invertible. Instead, we shall consider the operator T + I.
Dene the following operator on Z:
1

g(x, )h()d +

(Sh)(x) =

g(, x)h()d,
x

where
g(, x) = cot(1) cos(x) cos() + sin() cos(x).
This operator is clearly in L(Z), and by Theorem A.3.24 it is even compact. If we set f (x) =
(Sh)(x), then f is absolutely continuous and
df
(x)
dx

[ cot(1) cos() sin(x) + cos() cos(x)] h()d +

=
0
1

From this we see that

once more, we obtain

df
dx (0)

[ cot(1) cos() sin(x) sin() sin(x)] h()d.

df
dx (1)

= 0 and

df
dx

is absolutely continuous. Dierentiating

df
dx

d2 f
(x) = h(x) f (x).
dx2
Thus S is the bounded inverse of T + I. Thus, by Theorem A.3.46 T + I is closed, and hence T
is also closed.
Theorem A.3.46 gives an easy condition to check the closedenitioness of an operator. The
following theorem gives a similar result for the boundedenitioness of a linear operator.
Theorem A.3.49. Closed Graph Theorem. A closed linear operator dened on all of a
Banach space X into a Banach space Y is bounded.
Proof See Kato [14, theorem III.5.20], Kreyszig [16, theorem 4.13-2], Rudin [23, theorem 2.15],
Taylor [25, theorem 4.2-I], or Yosida [31, theorem II.6.1].

Corollary A.3.50. If T is a closed linear operator from a Banach space X to a Banach space Y
and T has an algebraic inverse T 1 , then T 1 is an element of L(Y, X) if and only if D(T 1 ) =
ran T = Y .

Proof See theorem 4.7-A in Taylor [25].

Many of the denitions that we gave for bounded linear operators have extensions to closed
operators. One of these notions is that of the dual operator.

148

A.3. Operators on normed linear spaces

Denition A.3.51. Let A be a closed, densely dened operator from D(A) X to Y , where
X and Y are Banach spaces. A is constructed in the following way. D(A ) consists of all g Y
such that there exists an f X with the property
g(Ax) = f (x)

for all x D(A).

The dual operator A g is dened by

A g = f

for g D(A ).

A.3.2. Operators on Hilbert spaces

In the last subsection, we introduced linear operators on a normed linear space. A Hilbert space
is a special normed linear space and so all the denitions made in that subsection are valid
for Hilbert spaces. However, since we have additional structure on Hilbert spaces (the inner
product), we can deduce extra properties of operators that exploit this structure.
One of the most important properties of a Hilbert space is that there is a particularly simple
representation for its dual space.
Theorem A.3.52. Riesz Representation Theorem. If Z is a Hilbert space, then every
element in Z induces a bounded linear functional f dened by
f (x) = x, z

Z.

On the other hand, for every bounded linear functional f on Z, there exists a unique vector
z0 Z, such that
f (x) = x, z0 Z for all x Z,
and furthermore, f = z0 .
Proof See Kato [14, p. 252 and 253], Kreyszig [16, theorem 3.8-1], Naylor and Sell [19, theorem
5.21.1], Taylor [25, theorem 4.81-C], or Yosida [31, section III.6].
Using this theorem, one can easily give a representation of nite-rank bounded operators. In
the next example, we do this for an operator of rank one.
Example A.3.53 Let Z be a Hilbert space and T L(Z) be an operator with one-dimensional
range. This means that there exists a v Z such that T z span{v} for all z Z. Hence,
T z = f (z)v for some mapping f . Since T is a linear and bounded operator, it follows directly
that f is bounded linear functional. Thus by the Riesz Representation Theorem A.3.52 there
exists a z0 Z such that f (z) = z, z0 , and so T z = z, z0 v.
The Riesz representation theorem gives an isometry between Z and Z . Usually, we identify
Z with its dual Z .
Example A.3.54 In Example A.3.27, we showed that
1
f : Z C; f (z) = z( )
2
denes a bounded linear functional on the Hilbert space Z := {z L2 (0, 1) | z is absolutely
dz
continuous on (0, 1) with dx L2 (0, 1) and z(0) = 0}. The Riesz representation theorem gives

149

A. Mathematical Background
that there exists an element y of Z such that z, y
y, we consider

= f (z) for every z in Z. To determine this

1
2

1
z( ) =
2

z(x)dx

and choose y(x) = 1[0, 1 ] (x), for then

2
1
2

z(x)y(x)dx =

1
z(x)dx = z( ).

So, if we dene

then y is an element of Z and z, y

1
0x 2
1
2 x 1,

y(x) =

1
2

1
= z( 2 ).

It is important to realize that a Hilbert space may be identied with several spaces, as can be
seen in the next example.
Example A.3.55 Consider the Hilbert space Z as dened in Example A.2.35

Z := {z = (zn ) |
with inner product

n=1

n |zn |2 < },

z, w

n zn wn .

:=
n=1

In Example A.3.32, we showed that the dual space can be identied with the Hilbert space

Z = {z = (zn ) |
with inner product

n=1

n |zn |2 < },

z, w

n zn wn ,

=
n=1

where n = n . However, from the Riesz Representation Theorem A.3.52 we see that Z can
also be identied with itself. For every element of Z , we calculate the element of Z such that
they dene the same linear functional. For (yn ) in Z the corresponding functional is dened by

g(z) =

zn yn

see (A.9).

n=1

An easy calculation shows that

g(z) =

n zn
n=1

150

1
yn = z, w
n

Z ,

A.3. Operators on normed linear spaces

1
where w = (wn ) = ( n yn ). This is an element of Z , since

n=1

n |wn |2 =

n=1

n |

1
1
|yn |2 =
n |yn |2 < .
yn |2 =
n
n
n=1
n=1

Similarly, for every element of Z we can construct an element of Z such that their corresponding
linear functionals are the same.
In the previous example, we saw that there is some freedom in identifying the dual of a Hilbert
space. However, in the situation that there are two Hilbert spaces W and Z such that W Z,
then we have from Lemma A.3.33 that Z W . If we could identify W with W and Z with
Z , then W would equal Z, but this is not true in general. For Hilbert spaces identied with
their dual we use the term pivot space. So, if in the previous discussion Z is the pivot space,
then
W Z = Z W .
(A.12)
This identication implies that if w W is also an element of Z, then
w (w) = w, w

Z.

(A.13)

It is usual to represent the action of the bounded linear functional w W on w W as a

duality pairing
w (w) := w, w W,W .
(A.14)
For more details about this we refer the reader to Aubin [1, chapter 3].
Example A.3.56 Consider the Hilbert space Z = 2 and Z dened in Example A.2.35. Since
n 1, we have
Z Z
and if we choose Z as the pivot space, we obtain

Z Z Z .

Consider the operator T : Z Z dened by

(T z)n =

1
zn .
n

Clearly, T is linear and bounded, and its dual T : Z Z. Since we have identied Z with
its dual, by the Riesz Representation Theorem A.3.52, there exists a bounded bijective operator

J : Z Z such that z (z) = z, Jz for any z Z and any z Z . Taking an arbitrary

z Z, we have that

z, T z

(T z )(z)

=
=

z (T z)
T z, Jz

by denition A.3.40

=
where wn =

n (Jz )n .

1
n zn (Jz )n = z, w
n
n=1

Z,

So we have shown that

(T z )n =

n
(Jz )n .
n

151

A. Mathematical Background
Another consequence of the Riesz Representation Theorem A.3.52 is the existence of the
adjoint operator.
Denition A.3.57. Let T L(Z1 , Z2 ), where Z1 and Z2 are Hilbert spaces. Then there exists
a unique operator T L(Z2 , Z1 ) that satises
T z1 , z2

Z2

= z1 , T z2

for all z1 Z1 , z2 Z2 .

Z1

This operator is called the adjoint operator of T .

Example A.3.58 Let Z be a complex Hilbert space and dene T z = z, zT for some zT Z.
It is easily seen that T L(Z, C). To calculate the adjoint of T , let z Z and C be arbitrary,
and consider T z, C = z, zT Z = z, zT Z . Thus T = zT .
b

Example A.3.59 Let Z = L2 (a, b) and dene K : Z Z by Kz() =

k(, s)z(s)ds, where

a

k L2 ([a, b] [a, b]). Then from Theorem A.3.24, K L(Z), and for z, w Z the following
holds:
b

Kz, w

k(t, s)z(s)dsw(t)dt

=
a

a
b

z(s)
a

k(t, s)w(t)dtds
a

=
a

Hence K w() =

b
a

k(t, s)w(t)dtds.

z(s)
a

k(t, )w(t)dt.

Since for Hilbert spaces we may identify the dual space with the space itself, there is a relationship between the adjoint and the dual operator. We shall show that such a relationship
exists for bounded linear operators (see also [16, section 4.5]). Let Z1 and Z2 be Hilbert spaces
and suppose that T L(Z1 , Z2 ). From denitions A.3.40 and A.3.57, we have

T : Z2 Z1

with (T z2 )(z1 ) = z2 (T z1 ),

and
T : Z2 Z1

with z1 , T z2

Z1

= T z1 , z2

Z2 .

From the Riesz Representation Theorem A.3.52, we have that Z1 is isometrically isomorphic to

Z1 . Thus there exists a bounded, bijective operator J1 from Z1 to Z1 such that

z1 (z) = z, J1 z1 .

A similar relationship holds for the Hilbert space Z2 . All these operators between the spaces
are given in Figure A.1. We remark that, for complex Hilbert spaces, J1 and J2 are not linear
operators, since
z, J1 (z1 + z2 )

Thus J1 (z1 + z2 ) = J1 z1 + J1 z2 .

152

=
=

(z1 + z2 )(z) = z1 (z) + z2 (z)

z, J1 z1 + z, J1 z2 .

A.3. Operators on normed linear spaces

Z1

T T

J1 6

Z2
J2 6

 T

Z1

Z2

Figure A.1.: The relationship between T and T

Using the denitions of T , T , J1 , and J2 , we see that for every z1 Z1 and z2 Z2 there
holds
z1 , T z2

Z1

=
=

1
T z1 , z2 Z2 = (J2 (z2 ))(T z1 )
1
1

(T (J2 (z2 )))(z1 ) = z1 , J1 T J2 z2

Z1 .

So T has the following representation:

1
T = J1 T J2 .

(A.15)

Usually, we identify Hilbert spaces with their dual, and in this case we use the adjoint and
the notation T , as in Example A.3.59. Only in situations where we choose not to identify the
Hilbert spaces do we use the dual notation T , as in Example A.3.56.
Many of the results presented in the next lemma can be proved using this relationship. Compare property a of Lemma A.3.41 with c of Lemma A.3.60.
Lemma A.3.60. Let T1 , T2 L(Z1 , Z2 ) and S L(Z2 , Z3 ), where Z1 , Z2 , and Z3 are Hilbert
spaces. The adjoint has the following properties:
a. I = I;

b. (T1 ) = T1 ;

c.

T1 = T1 ;

d. (T1 + T2 ) = T1 + T2 ;

e. (ST1 ) = T1 S ;

f.

T1 T1 = T1 2 .

Proof See Kreyszig [16, theorem 3.9-4], Naylor and Sell [19, theorem 5.22.2 and corollary 5.22.3],
Rudin [23, section 12.9], Taylor [25, section 4.9], or Yosida [31, section VII.2].
Theorem A.3.61. If T L(Z1 , Z2 ), where Z1 and Z2 are Hilbert spaces, then we have the
following equalities:
a. ran (T ) = ker(T );
b. ran (T ) = ker(T ) ;

153

A. Mathematical Background
c. ran (T ) = ker T ;
d. ran (T ) = ker(T ) .
Proof See proposition 1 and corollary 2 in Aubin [1, chapter 3, section 4], Naylor and Sell [19,
theorem 5.22.6], or Rudin [23, theorem 12.10].
Denition A.3.62. A bounded linear operator T on a Hilbert space Z is:
a. normal if T T = T T ;
b. unitary if T T = T T = I.
In the theory of this book we also need the adjoint of an unbounded linear operator.
Denition A.3.63. Let A be a linear operator on a Hilbert space Z. Assume that the domain
of A, D(A), is dense in Z. Then the adjoint operator A : D(A ) Z Z of A is dened as
follows. The domain D(A ) of A consists of all y Z such that there exists a y Z satisfying
Ax, y = x, y

for all x D(A).

For each such y D(A ) the adjoint operator A is then dened in terms of y by
A y = y .

In addition, it can be shown that if A is a closed, densely dened operator, then D(A ) is
dense in Z and A is closed. Furthermore, one can show the same relationship exists between
A and A as in (A.15). In the following example we shall calculate the adjoint of an unbounded
operator heuristically.
Example A.3.64 Let Z = L2 (0, 1) and consider the operator A given by
(Az)(x) =

dz
(x),
dx

dz
where D(A) = {z L2 (0, 1) | z is absolutely continuous with dx L2 (0, 1), z(0) = 0}.
In Example A.3.45 we have shown that A is a closed linear operator. Now we calculate its
adjoint heuristically as follows:
1

Az, y

dz
(x)y(x)dx = z(x)y(x)
dx

0
1

= z(1)y(1)

z(x)

1
0

z(x)

dy
(x)dx
dx

dy
(x)dx.
dx

dy
This can be written in the form z, y if and only if y(1) = 0 and dx Z. So the logical choice
dy
dy
is D(A ) = {y Z | y is absolutely continuous with dx Z and y(1) = 0} and A y = dx .

In order to justify the above we need the following results.

Lemma A.3.65. Let A be an arbitrary, densely dened operator and let T be a bounded linear
operator dened on the whole of the Hilbert space Z. The following holds:

154

A.3. Operators on normed linear spaces

a. (A) = A ; D((A) ) = D(A ) if = 0 and Z if = 0;
b. (A + T ) = A + T , with domain D((A + T ) ) = D(A );
c. If A has a bounded inverse, i.e., there exists an A1 L(Z) such that AA1 = IZ ; A1 A =
ID(A) , then A also has a bounded inverse and (A )1 = (A1 ) .
Proof a. This is shown in Rudin [23, theorem 13.2].
b. Suppose that there exist y, y such that (A + T )x, y = x, y for all x D(A + T ). This
implies that
Ax, y = x, y T y
for all x D(A + T ) = D(A).

Hence y D(A ) and A y = y T y, and so we conclude that D((A + T ) ) D(A ) and

(A + T ) = A + T on D((A + T ) ). The inclusion D(A ) D((A + T ) ) follows similarly.
c. See Kato [14, theorem III.5.30] or Kreyszig [16, theorem 10.2-2].

Example A.3.66 Let Z be the Hilbert space L2 (0, 1) and consider the operator of Example
A.3.64 again, i.e.,
d
A=
dx
with
D(A) = {z Z | z is absolutely continuous
dz
with z(0) = 0 and dx L2 (0, 1)}.

From Example A.3.47, we have that the algebraic inverse is bounded and given by
x
1

(A

z(s)ds.

z)(x) =
0

We calculate A via (A1 ) ; so we consider

x

1
1

z1 , z2

z1 (s)dsz2 (x)dx

=
0

0
1

z1 (s)z2 (x)dxds

=
s

0
1

where [(A1 ) z2 ](s) =

z2 (x)dxds = z1 , (A1 ) z2 ,

z1 (s)

by Fubinis Theorem A.5.22

z2 (x)dx. From this it is easy to see that

s

A z =

dz
dx

with domain
D(A) = {z Z | z is absolutely continuous with z(1) = 0 and

dz
dx

L2 (0, 1)}.

Thus we see that this is the same as in Example A.3.64. The dierence here is that we have
proven it rigorously.

155

A. Mathematical Background
Example A.3.67 Let Z be the Hilbert space L2 (0, 1) and consider the operator of Example
A.3.48 on L2 (0, 1)
d2
A= 2
dx
with domain
dz
D(A) = {z L2 (0, 1) | z, dx are absolutely continuous
dz
d2
dz
with dx (0) = dx (1) = 0 and dxz L2 (0, 1)}.
2
From Example A.3.48, we have that A + I has a bounded, algebraic inverse given by
1

x
1

((I + A)

g(x, )h()d +

h)(x) =

g(, x)h()d,
x

where
g(, x) = cot(1) cos(x) cos() + sin() cos(x).
If we calculate the adjoint of (I + A)1 , then we have that
(I + A)1 h, z
x

g(x, )h()dz(x)dx +

=
0

0
1

g(, x)h()dz(x)dx
x

g(x, )h()z(x)dxd +

=
0

h()

=
0

g(, x)h()z(x)dxd
0

h()

g(x, )z(x)dxd +

g(, x)z(x)dxd
0

h, (I + A)1 z .

So we see that ((I + A)1 ) = (I + A)1 . Thus from Lemma A.3.65.c it follows that (I + A) =
I + A, and from Lemma A.3.65.b we conclude that A = A.
Example A.3.67 belongs to a special class of operators.
Denition A.3.68. We say that a densely dened, linear operator A is symmetric if for all
x, y D(A)
Ax, y = x, Ay .
A symmetric operator is self-adjoint if D(A ) = D(A).

All bounded, symmetric operators are self-adjoint. It can be shown that the adjoint of an
operator is always closed, so, in particular, every self-adjoint operator is closed. Furthermore,
we have from Lemma A.3.65 that an invertible operator is self-adjoint if and only if its inverse is
(see Example A.3.67). For a self-adjoint operator, we always have that Az, z = z, Az . Thus
from property b of denition A.2.23 we conclude that Az, z must be real for all z D(A). The
converse is also true.
Lemma A.3.69. Let T be an element of L(Z), with Z a complex Hilbert space. T is self-adjoint
if and only if T z, z is real for all z Z.

156

A.3. Operators on normed linear spaces

Proof See Kreyszig [16, theorem 3.10-3] or Naylor and Sell [19, theorem 5.23.6].
Lemma A.3.70. Let T be a self-adjoint operator in L(Z), where Z is a Hilbert space. We have
the following relation between the norm and the inner product:
T = sup | T z, z |.
z =1

Proof See Kreyszig [16, theorem 9.2-2], Naylor and Sell [19, theorem 5.23.8], or theorem 3 in
Yosida [31, section VII.3].
So for every self-adjoint operator A the range of Az, z is real. Operators for which this range
is nonnegative have a special name.
Denition A.3.71. A self-adjoint operator A on the Hilbert space Z is nonnegative if
Az, z 0

for all z D(A);

A is positive if
for all nonzero z D(A);

Az, z > 0

and A is coercive if there exists an > 0 such that

Az, z z

for all z D(A).

We shall use the notation A 0 for nonnegativity of the self-adjoint operator A, and A > 0 for
positivity. Furthermore, If T, S are self-adjoint operators in L(Z), then we shall write T S for
T S 0.
With this new notation, it is easy to see that A is coercive if and only if A I, for some
> 0. Some of the special properties of self-adjoint, nonnegative operators are collected in the
following theorem and lemmas.
Theorem A.3.72. Let Z be a complex Hilbert space, and let Tn be a sequence of bounded,
nonnegative, self-adjoint operators on Z such that Tn+1 Tn and I Tn , for some positive
R. Then the sequence {Tn } is strongly convergent; that is, there exists a T L(Z) such
that Tn z T z for every z Z. Furthermore, T is nonnegative, self-adjoint, and I T Tn
for all n.
Proof See Kreyszig [16, theorem 9.3-3].
Lemma A.3.73. If A is self-adjoint and nonnegative, then A has a unique nonnegative square
1
1
1
1
1
1
root A 2 , so that D(A 2 ) D(A), A 2 z D(A 2 ) for all z D(A), and A 2 A 2 z = Az for
1
z D(A). Furthermore, if A is positive, then A 2 is positive too.
Proof See Kato [14, theorem V-3.35] for a general proof. Kreyszig [16, theorem 9.4-2] and Rudin
[23, theorem 12.33] only prove the bounded case.
Lemma A.3.74. Let T be a nonnegative, self-adjoint operator in L(Z), where Z is a Hilbert
space. It has the following properties:
a.

T2 = T

1
2

b. | T z1 , z2 |2 T z1 , z1 T z2 , z2 for all z1 , z2 Z;

157

A. Mathematical Background
c.

Tz

z, T z for all z Z.

Note that b is a generalization of the Cauchy-Schwarz inequality.

1

Proof a. This follows from Lemma A.3.60.f with T1 = T 2 .

b. For z1 , z2 Z we have
1

| T z1 , z2 |2

T 2 z1 2 T 2 z2 2
T z1 , z1 T z2 , z2 .

c. It is easy to see that T z = sup

Tz

= | T 2 z1 , T 2 z2 |2

y =1

by the Cauchy-Schwarz inequality

T z, y , and thus using part b we obtain

= sup | T z, y |2 sup T y, y T z, z = T
y =1

T z, z ,

y =1

where we used Lemma A.3.70 and the fact that T is nonnegative.

One of the most important classes of nonnegative operators is the orthogonal projections.
Denition A.3.75. An operator P L(Z) is a projection if P 2 := P P = P , and a projection
operator is called orthogonal if P = P .
In Appendix A.2, we have seen that given a closed linear subspace V we can decompose the
Hilbert space Z into Z = V V . Let z be any element of Z, then there exist zV V and
zV V such that z = zV + zV .
Dene the operator P : Z Z by P z = zV . Then, since z 2 = zV 2 + zV 2 , we have
that P L(Z), ker P = V , and P is an orthogonal projection. We call P the orthogonal
projection on V . On the other hand, if P is an orthogonal projection, then Z = ran P ker P
with ran P ker P . So an orthogonal projection is naturally associated with an orthogonal
decomposition of the Hilbert space.
We close this section with an important lemma concerning minimalization problems that
complements Theorem A.2.36.
Lemma A.3.76. Orthogonal Projection Lemma. Let Z be a Hilbert space and V a closed
subspace of Z. Then, given z0 Z, there exists a unique v0 in V such that
z0 v0 = min z0 v .
vV

Furthermore, the element v0 is given by v0 = PV z, where PV is the orthogonal projection on V .

We see that z0 v0 = PV z0 .
Proof See Kato [14, page 252], Kreyszig [16, theorem 3.3-1 and lemma 3.3-2], or Naylor and
Sell [19, theorem 5.14.4].
Corollary A.3.77. Let Z be a Hilbert space, V a closed subspace of Z, z0 Z, and dene the
ane set
Vz0 := {z Z | z = z0 + v for some v V }.
There exists a unique element zV in Vz0 such that

zV = min z .
zVz0

This element is given by zV := PV z, where z is an arbitrary element of Vz0 .

158

A.4. Spectral theory

Proof This follows from Lemma A.3.76, since
min z = min z0 + v = min z0 v .

zVz0

vV

vV

A.4. Spectral theory

A.4.1. General spectral theory
In this section, we consider abstract equations of the form
(I A)x = y,

(A.1)

where A is a closed linear operator on a complex Banach space X with D(A) X, x, y X,

and C. As an example of this formulation, which will be considered in more detail later in
this appendix, we consider the boundary value problem
d2 z
(x) + z(x) = v(x)
dx2
z(0) = 0 = z(1),

on L2 (0, 1),

where v is a given function in L2 (0, 1).

The solutions of these problems are reduced to asking under what conditions (I A) has
a bounded inverse on the particular Banach space X. When X is nite-dimensional, it is well
known that this depends on whether is an eigenvalue of A. For the innite-dimensional case, we
need to generalize the concept of eigenvalues. We shall generalize this to the class of closed linear
operators, and a study of these will give useful information about the existence and uniqueness
of solutions to (A.1). This abstract approach to studying linear equations on a Banach space
is what is known as spectral theory. This theory can be found in almost any book on operator
theory. In the nite-dimensional case, not every matrix on a real space has eigenvalues and
eigenvectors. To overcome this situation, one has to consider the matrix on a complex space. In
this section, we shall therefore only consider complex normed spaces.
As our motivation is the study of linear equations of the form
y = (I A)x
on a complex normed linear space X, where A : D(A) X X is a closed linear operator,
we are interested in those C for which (A.1) has a unique solution for all y X, which,
following denition A.3.5, we may write as
x = (I A)1 y,

(A.2)

where (I A)1 is the algebraic inverse of I A. Here we also require that this inverse is
bounded.
Denition A.4.1. Let A be a closed linear operator on a (complex) normed linear space X. We
say that is in the resolvent set (A) of A, if (I A)1 exists and is a bounded linear operator
on a dense domain of X.

159

A. Mathematical Background
Now, from the fact that (I A) is closed and invertible, (I A)1 must be a closed
operator. Since it is also a bounded linear operator, its domain must be a closed subspace. So
its domain is both closed and dense, which by the Closed Graph Theorem A.3.49 means that
(I A)1 L(Z). So (A) if and only if (I A)1 L(Z). We shall call (I A)1 the
resolvent operator of A. Other names that are used are: bounded algebraic inverse or bounded
inverse.
Example A.4.2 Let Z be a Hilbert space. Consider the positive, self-adjoint operator A on Z
that is coercive, i.e., Az, z z 2 for all z D(A) and a given > 0. From this it is clear
that A is injective, and so from Lemma A.3.6.a we obtain that the algebraic inverse exists. For
z ran A we have
1
1
AA1 z, A1 z
z A1 z ,
A1 z 2

where we have used the Cauchy-Schwarz inequality. This implies that

A1 z

1
z ,

and A1 is bounded on its range. If ran A is dense in Z, then 0 is in the resolvent set of A and
A1 L(Z).
Let x be in the orthogonal complement to the range of A, i.e., for all z D(A) the following
holds:
Az, x = 0.
By denition A.3.63, this implies that x D(A ) and A x = 0. Since A is self-adjoint, we
conclude that Ax = A x = 0. The positivity of A shows that this can only happen if x = 0, and
so ran A is dense in Z.
Example A.4.3 Let Z be a Hilbert space and consider the positive, self-adjoint operator A on
Z that is coercive. In the previous example, we saw that A is boundedly invertible. This inverse
is positive, since for every nonzero z Z
A1 z, z = y, Ay > 0,

where y = A1 z.

Since A and A1 are positive operators, they have a positive square root (see Lemma A.3.73).
1

We show that A 2

1
2

= A1
1
2

.
1

Dene the operator Q = A A . Since ran A1 = D(A) and D(A 2 ) D(A) (see Lemma
A.3.73), we have that Q is a well dened linear operator. If Q is closed, then by the Closed
Graph Theorem A.3.49 Q L(Z). Let zn z and Qzn y. Then xn := A1 zn A1 z and
1
1
A 2 xn y. From the fact that A 2 is closed (see the remark after denition A.3.68), we conclude
1
1
that A 2 A1 z = y. So Q is closed and hence bounded. It is easy to see that ran Q D(A 2 ) and
1

A 2 Q = IZ .
1

(A.3)

Dene x = QA 2 z, for z D(A 2 ). Then

1

A 2 x = A 2 QA 2 z = A 2 z
1

by (A.3).
1

1 . Thus A 2 is invertible,
The operator A 2 is positive, and so z = x. In other words, QA 2 = I
D(A 2 )
and
1
1
1
= Q = A 2 A1 .
(A.4)
A2

160

A.4. Spectral theory

1

To see that A 2

is positive, consider the following for z Z

1

A2

Multiplying both sides of (A.4) by A 2

A2
Thus
1

A2

A2
1

z, z = y, A 2 y > 0,

where y = A 2

= A1

z.

gives
1

A2

is a positive square root of A1 .

1
2

= A1 .
Since the positive square root is unique,
1

. We shall denote this operator by A 2 .

Denition A.4.4. Let A be a closed linear operator on a (complex) normed linear space X.
The spectrum of A is dened to be
(A) = C\(A).
The point spectrum is
p (A) = { C | (I A) is not injective}.
The continuous spectrum is
c (A) = { C | (I A) is injective, ran (I A) = X, but
(I A)1 is unbounded}
= { C | (I A) is injective, ran (I A) = X, but
ran (I A) = X}.

The residual spectrum is

r (A) = { C | (I A) is injective, but ran (I A) is not dense in X}.
So (A) = p (A) c (A) r (A).
A point p (A) is an eigenvalue, and x = 0 such that (I A)x = 0, an eigenvector.

For eigenvalues, we have natural generalizations of the nite-dimensional concepts.

Denition A.4.5. Let 0 be an eigenvalue of the closed linear operator A on the Banach space
X. Suppose further that this eigenvalue is isolated ; that is, there exists an open neighbourhood
O of 0 such that (A) O = {0 }. We say that 0 has order 0 if for every x X
lim ( 0 )0 (I A)1 x

exists, but there exists an x0 such that the following limit does not
lim ( 0 )0 1 (I A)1 x0 .

If for every N there exists an x X such that the limit

lim ( 0 ) (I A)1 x

does not exist, then the order of 0 is innity.

For the isolated eigenvalue 0 of nite order 0 , its (algebraic) multiplicity is dened as
dim(ker(0 I A)0 ). The elements of ker(0 A)0 are called the generalized eigenvectors
corresponding to 0 .

161

A. Mathematical Background
We remark that if the kernel of (0 IA) is nite-dimensional, then so is the kernel of (0 IA)
for any 1.
In nite-dimensional spaces, we always have that c (A) and r (A) are empty, but this is not
the case if X is innite-dimensional, as can be seen from the following example.
Example A.4.6 Let X = 1 and let T : X X be given by
T x = (x1 ,

xn
x2
,...,
, . . .).
2
n

Consider
(I T )x = y.
Now

1
)xn , . . .).
n
1
So = n , n = 1, 2, . . ., are the eigenvalues of the operator T with associated eigenvectors
1
en := (0, . . . , 0, 1, 0 . . .). Let = n , n = 1, 2, . . ., and = 0; then
(I T )x = (( 1)x1 , . . . , (

1
1
x = (I T )1 y = (( 1)1 y1 , ( )1 y2 , . . . , ( )1 yn , . . .).
2
n
This denes a bounded linear operator. Thus (T ) C \ {0, 1, 1 , 1 , . . .}.
2 3
So we only have to investigate in which part of the spectrum the point 0 lies. If = 0, then
1
x = (y1 , 2y2 , . . . , nyn , . . .), and for y = ( n2 ) we have that x

=
i=1

1
n

is not nite, and so

0 (T ). We know that T is one-one, so we must determine the range of T to decide whether

0 c (T ) or r (T ). Now T (nen ) = en , and so en ran T . Since span{en } is X, we have
ran T = X and so 0 c (T ).
Summarizing, we have shown that
c (T ) = {0}; r (T ) = ,
1
1
p (T ) = {1, , . . . , , . . .}.
2
n
One of the most important relations for the resolvent operator is the resolvent equation
(I A)1 = (I A)1 + ( )(I A)1 (I A)1 ,

(A.5)

for , (A).
Repeatedly applying the resolvent equation gives
n

(I A)1

=
k=0

( )k (I A)k1 +

( )n+1 (I A)n1 (I A)1 .

If ( )(I A)1 < 1, then the term ( )n+1 (I A)n1 converges uniformly to zero,
and we obtain the following result.
Lemma A.4.7. Suppose that , (A), and ( )(I A)1 < 1. Then

(I A)1 =

162

( )k (I A)k1 .

k=0

(A.6)

A.4. Spectral theory

Proof See Kreyszig [16, theorem 7.3-3] or Taylor [25, theorem 5.1-C].
A direct consequence of this lemma is the following result.
Lemma A.4.8. For a closed linear operator, A, on a Banach space X the following hold:
a. If (A) and C are such that ( )(I A)1
(I A)1 is given by (A.6);

< 1, then (A), and

b. The resolvent set of A is open;

c. The resolvent operator is a holomorphic function on (A), and the derivative is given by
d
1
= (I A)2 ; see denition A.5.32.
d (I A)
Proof a. See Kreyszig [16, theorem 7.3-3], Naylor and Sell [19, theorem 6.7.3], or Taylor [25,
theorem 5.1-A].
b. See Kato [14, theorem III.6.7], Kreyszig [16, theorem 7.3-2], Naylor and Sell [19, theorem
6.7.3], Taylor [25, theorem 5.1-B], or theorem 1 in Yosida [31, section VIII.2].
c. See Kato [14, theorem III.6.7], Kreyszig [16, theorem 7.5-2], or Example A.5.33, or Taylor [25,
theorem 5.1-C].
The order of the isolated eigenvalue as dened in denition A.4.5 is the same as the order of
the pole that (I A)1 has as a holomorphic function.
So for closed linear operators we have information about the resolvent set and the resolvent
operator. For bounded linear operators we can prove even stronger results. A useful tool is the
following theorem.
Theorem A.4.9. Let X be a Banach space and T L(X) such that T < 1. Then (I T )1
exists and is in L(X) with
(I T )1 = I + T + T 2 + . . . + T n + . . . ,
where the convergence is in the uniform topology on L(X) and
(I T )1 (1 T )1 .
Proof See Kreyszig [16, theorem 7.3-1], Naylor and Sell [19, theorem 6.7.2], Rudin [23, theorem
10.7], or Taylor [25, theorem 5.2-A].
1
If we rewrite (I T ) = (I T ), then we have the following corollary of Theorem A.4.9.

Corollary A.4.10. Let T L(X), where X is a Banach space. If || > T , then (T ).

Moreover, we have

(I T )1 =
and

n1 T n
n=0

(I T )1 (|| T )1.

A consequence of this corollary and Lemma A.4.8.c is the following lemma.

Lemma A.4.11. If T L(X), where X is a Banach space, then (T ) is nonempty.

163

A. Mathematical Background
Proof See Kato [14, section III.6.2], Kreyszig [16, theorem 7.5-4], Rudin [23, theorem 10.13], or
Taylor [25, theorem 5.2-B].
We summarize results concerning the spectrum of a bounded linear operator in the following
theorem.
Theorem A.4.12. If X is a complex Banach space and T L(X), then the spectrum of T is
a nonempty compact subset of the complex plane lying in the closed ball { C | || T }.
Proof See Kato [14, section III.6.2], Kreyszig [16, theorem 7.3-4], Naylor and Sell [19, theorem
6.7.4], or Rudin [23, theorem 10.13].
Example A.4.13 Let Z denote the Hilbert space 2 (Z) and let A be the right shift operator
given by
(Az)k = zk1
for k Z,
where z = (. . . , z1 , z0 , z1 , . . .).
It is easy to see that A = 1, and so by Theorem A.4.12 we may conclude that (A) {s
C | |s| 1}. Furthermore, we have that A1 exists and, it is given by the left shift operator
(A1 z)k = zk+1

for k Z.

From this it follows that A1 = 1, and so by Theorem A.4.12 we deduce that (A1 ) {s
C | |s| 1}.
For 0 = s C the following equality holds
1
(sI A) = sA(A1 I).
s
For 0 < |s| < 1 the inverse of the right-hand side exists in L(Z), and so the inverse of the
left-hand side is in L(Z) too. Hence we see that (A) {s C | |s| = 1}. We show that, in
fact, equality holds.
Suppose that is an eigenvalue with || = 1. Then there exists a z Z satisfying Az = z.
From the denition of A, we obtain
zk1 = zk

for k Z.

The unique solution of this equation is given by zk = k z0 . However, since || = 1, we have

that this is in Z = 2 (Z) if and only if z0 = 0, and this implies that z = 0. Thus a with
modulus one cannot be an eigenvalue. Next we shall show that every on the unit circle is in
the spectrum of A. Let en denote the element of Z, which is dened as follows
en = (en ) = n,k =
k

1
0

for k = n
for k = n.

Consider the equation (A I)z = e0 , or equivalently,

zk1 zk = e0
k

for k Z.

This has the unique solution zk = 0 for k < 0 and zk = 1k for k 0. Again, since || = 1,
this is not an element of Z, and thus we have shown that ran (I A) = Z. The remark after
denition A.4.1 shows that (A). Combining these results and using Theorem A.4.12, we
conclude that (A) = {s C | |s| = 1}. For completeness we shall show that (A) = c (A). Let

164

A.4. Spectral theory

z Z be any element in the orthogonal complement of ran (A I). This implies that for every
n Z we have that
z, (A I)en
z, en+1 en

for all n Z
for all n Z

= 0
= 0

z n+1 z n

for all n Z,

= 0

where z = (zn ). This equation has the unique solution zn = n z0 . However, since || = 1, this
is in Z if and only if z0 = 0. Hence z = 0 is the only element in the orthogonal complement of
ran (A I), and thus ran (A I) = Z. From denition A.4.4 it follows that c (A) = {s C |
|s| = 1} = (A).
In the following lemma, we characterize the exact radius of the smallest ball containing the
spectrum of T . The spectral radius r (T ) of the operator T L(X) on a complex Banach space
X is dened by
r (T ) := sup ||.
(T )

We have the following result.

Lemma A.4.14. For T L(X) on the complex Banach space X, we have
r (T ) = lim

Tn .

Proof See Kato [14, section III.6.2, equation (6.13)], Kreyszig [16, theorem 7.5-5], Rudin [23,
theorem 10.13], Taylor [25, theorem 5.2-E], or Yosida [31, theorems 3 and 4 in section VIII.2].

With this lemma we can easily prove the following result.

Lemma A.4.15. Let T, S be bounded operators on the Banach space X. The following relation
holds:
r (T S) = r (ST ).
Proof If S or T is the zero operator, then the result is trivial. Suppose that they are nonzero.
We have that
r (T S) =

lim

[T S]n

lim

lim

n
n
n

[ST ]n1
lim

S
[ST ]n1

by Lemma A.3.14
lim

= 1 r (ST ) 1.
Hence we have shown that r (T S) r (ST ). Similarly, one can show that the reverse inequality
holds, and so we have proved the assertion.
Lemma A.4.14 gives information about the size of the spectrum of an operator. For self-adjoint
operators, we have more information.
Lemma A.4.16. If A is a self-adjoint operator on the Hilbert space Z, then (A) R. Furthermore, if A L(Z), then we have the following additional properties:

165

A. Mathematical Background
a. (A) [m, M ], where m = inf Az, z and M = sup Az, z ;
z =1

z =1

b. m, M (A);
c.

A = max{|m|, |M |};

d. r (A) = A .
Proof a. See Kreyszig [16, theorem 9.2-1] or Taylor [25, theorem 6.2-B].
b. See Kreyszig [16, theorem 9.2-3] or Taylor [25, theorem 6.2-B].
c. See Kreyszig [16, theorem 9.2-2], Taylor [25, theorem 6.11-C], or Yosida [31, theorem 3 in
section VII.3].
d. This follows from parts a, b, and c.
Lemma A.4.17. Consider the densely dened, closed, linear operator A on the Banach space
X. The following relation holds between the spectrum of A and of its adjoint
(A ) = (A),

(A.7)

where the bar denotes the complex conjugate.

Proof See Kato [14, theorem III.6.22].

A.4.2. Spectral theory for compact normal operators

From the previous subsection it is clear that the spectral properties of innite-dimensional operators are much more complicated than those for nite-dimensional operators. However, compact
operators have a simple spectrum, and the following theorem shows that we can expect the
theory for solutions of linear equations with compact operators to be similar to that for (nitedimensional) operators on Cn .
Theorem A.4.18. If T is a compact operator on a Banach space X, then = 0 is in either the
point spectrum or the resolvent set of T . The point spectrum of T is, at most, countably innite
with = 0 the only possible point of accumulation. Furthermore, the order of every nonzero
eigenvalue is nite, and so is its multiplicity.
Proof See Kato [14, theorem III.6.26], Kreyszig [16, theorems 8.3-1 and 8.3-3], Naylor and Sell
[19, corollary 6.10.5 and theorem 6.10.1], Rudin [23, theorem 4.25], Taylor [25, theorems 5.5-C
and 5.5-G], or Yosida [31, theorem 2, section X.5].
In Section A.3 we have seen that if the Banach space is innite-dimensional and the compact
operator T is injective, then 0 is an element of the spectrum of T ; see Lemma A.3.22.
The following general result is very useful in the applications.
Lemma A.4.19. Let A be a closed linear operator with 0 (A) and A1 compact. The
spectrum of A consists of only isolated eigenvalues with nite multiplicity.
Proof See Kato [14, theorem III.6.29].
Compact, normal operators on a Hilbert space do not have generalized eigenvectors but they
have a spectral decomposition analogous to normal matrices.

166

A.4. Spectral theory

Theorem A.4.20. If T L(Z) is a compact, normal operator on a Hilbert space Z, then
there exists an orthonormal basis of eigenvectors {i , i 1} corresponding to the eigenvalues
{i , i 1} such that

i z, i i

Tz =
i=1

for all z Z.

Proof See Kato [14, theorem V.2.10] or Naylor and Sell [19, theorem 6.11.2].
A consequence of this theorem is that every compact, normal operator induces an orthonormal
basis for the Hilbert space. We shall illustrate this by the next classical examples.
Example A.4.21 Let Z = L2 (0, 1) and let T be given by
1

g(x, )v( )d,

(T v)(x) =

(A.8)

where

(1 )x
(1 x)

g(x, ) =

for 0 x 1,
for 0 x 1.

Since g(x, ) = g(, x), we see from Example A.3.59 that T is self-adjoint. Furthermore, T is a
compact operator, by Theorem A.3.24. So we may calculate the eigenvalues and eigenvectors.
We can rewrite (A.8) as
x

(T v)(x) =
0

(1 x) v( )d +

(1 )xv( )d.

So (T v) is absolutely continuous and (T v)(0) = 0 = (T v)(1). Let C be such that T v = v.

Then
x

v( )d +
0

xv( )d

x v( )d = v(x) for x [0, 1].

(A.9)

Since the left-hand side is absolutely continuous, we may dierentiate (A.9) to obtain
1

xv(x) xv(x) +

v( )d

v( )d = v(x) for x [0, 1].

(A.10)

The left-hand side is again absolutely continuous and so we may dierentiate (A.10) to obtain
v(x) = (x).
v

(A.11)

So = 0 is not an eigenvalue. The general solution of (A.11) is given by

1

v(x) = a sin( 2 x) + b cos( 2 x).

(A.12)

1
1
Using the fact that v(0) = (T v)(0) = 0 and v(1) = (T v)(1) = 0 gives = n21 2 and v(x) =

1
a sin(nx). So the eigenvalues are { n2 2 , n 1} and the eigenvectors are {sin(nx), n 1}. By

Theorem A.4.20, we now have that { 2 sin(nx), n 1} is an orthonormal basis for L2 (0, 1).

167

A. Mathematical Background
Example A.4.22 Let Z = L2 (0, 1)and let S be the operator dened in Example A.3.48
1

g(x, )h()d +

(Sh)(x) =

g(, x)h()d,
x

where
g(, x) = cot(1) cos(x) cos() + sin() cos(x).
This operator is clearly in L(Z); by Theorem A.3.24 it is even compact. From Example A.3.67
we have that S is self-adjoint, and so it is certainly normal. From Example A.3.48, we have that
S is the bounded inverse of I + A, where
A=
with domain

D(A)

d2
,
dx2

dz
= {z L2 (0, 1) | z, dx are absolutely continuous
dz
d2
dz
with dx (0) = dx (1) = 0 and dxz L2 (0, 1)}.
2

We shall calculate the eigenvalues and eigenvectors of S. If Sz = z, then by applying the inverse
we obtain that z = (I + A)z. So we have to solve
d2 z
1
=
z.
dx2

(A.13)

Using the boundary conditions (z D(A)), this has a nonzero solution if 1 = n2 2 for

1
some n 0, and then z is given by cos(n). So the eigenvalues are given by { 1n2 2 , n 0} and

the eigenvectors are {cos(n), n 0}. From Theorem A.4.20, we obtain that {1, 2 cos(n), n
1} is an orthonormal basis of Z.
From Theorem A.4.20, we see that every compact, normal operator has a nice representation.
In the next theorem we shall show that every compact operator has a similar representation.
Theorem A.4.23. If T L(Z1 , Z2 ) is a compact operator, where Z1 and Z2 are Hilbert spaces,
then it has the following representation:

i z1 , i i ,

T z1 =

(A.14)

i=1

where {i }, {i } are the eigenvectors of T T and T T , respectively, and i 0 are the square
roots of the eigenvalues. {i } form an orthonormal basis for Z1 and {i } form an orthonormal
basis for Z2 . (i , i ) are the Schmidt pairs of T , i the singular values and (A.14) is its Schmidt
decomposition.
Furthermore, the norm of T equals its largest singular value.
Proof a. T T is clearly a self-adjoint, nonnegative operator and by Lemma A.3.22 it is compact.
2
By Theorem A.4.20 T T has the following representation in terms of its eigenvalues i , (i 0)
and its eigenvectors i , which form an orthonormal basis for Z1

2
i z, i i =

T T z =
i=1

168

2
i z, i i ,
iJ

A.4. Spectral theory

where J is the index set that contains all indices for which i > 0. For i J we dene
i =

1
T i ,
i

and we easily obtain

T i =

1
T T i = i i
i

for i J.

Notice that
2
T T i = i T i = i i

for i J,

(A.15)

2
which shows that i is the eigenvector of T T corresponding to i . They form an orthonormal
set, since
1
i
1
T i , T j =
T T i , j =
i , j = i,j .
i , j =
i j
i j
j

We shall show that we can extend the set {i ; i J} to an orthonormal basis for Z2 by adding
an orthonormal basis for the kernel of T . For this we need the following observation. If i is
an eigenvector of T T corresponding to the eigenvalue zero, then
2

T i

= T i , T i = T T i , i = 0.

(A.16)

Let z2 be orthogonal to every i with i J. For i J we have that

i , T z2 = T i , z2 = i i , z2 = 0.
For i not an element of J it follows directly from (A.16) that i , T z2 = 0. Since {i } forms an
orthonormal basis, we have that T z2 = 0. So we have shown that the orthogonal complement
of {i ; i J} equals the kernel of T . This implies that we can decompose the Hilbert space Z2
into the direct sum of the closure of the span of {i ; i J} and the kernel of T (A.4). Choosing
an orthonormal basis for the kernel of T produces an orthonormal basis for Z2 . We shall denote
this basis by {i ; i N}. For i J we have seen that i is an eigenvector for T T . However, for
i not an element of J we have that T T i = 0, and so i is also an eigenvector for all i N.
b. We now show that T has the Schmidt decomposition
Tz =

i z, i i .
iJ

From Theorem A.4.20 {i } is an orthonormal basis in Z1 , and so

for all z Z1 .

z, i i

z=
i=1

Since T is bounded, the following holds:

Tz

z, i T i =

=
i=1

i z, i i .

z, i T i

by (A.16)

iJ

by (A.15)

iJ

169

A. Mathematical Background
c. Let us number the singular values so that 1 2 . Since {i } is a orthonormal set in
Z2 , from (A.14), it follows that

Tz

=
i=1

2
2
i | z, i |2 1 z 2 .

But T 1 = 1 1 = 1 and so T = 1 .
So the class of compact operators has very special properties. Another class of operators with
useful properties is the following.
Denition A.4.24. Let A be a linear operator on a Hilbert space Z. We say that A has
compact, normal resolvent if there exists a 0 (A) for which (0 I A)1 is compact and
normal.

With the resolvent equation one can easily prove that denition A.4.24 is independent of the
particular 0 , i.e., if (0 I A)1 is compact and normal, then (I A)1 is compact and normal
for all (A).
Theorem A.4.25. Let A be a linear operator on the Hilbert space Z with domain D(A) and
let 0 (A) with A1 compact and normal. From Theorem A.4.20, it follows that for z Z we
have the representation

1 z, i i ,
i

z=
i=1

1
i

where
and i are the eigenvalues and the eigenvectors of A1 , respectively, and {i , i 1}
is an orthonormal basis. Moreover, for z D(A), A has the decomposition

i z, i i ,

Az =
i=1

with D(A) = {z Z |

i=1

|i |2 | z, i |2 < }, and A is a closed linear operator.

Proof Dene

i z, i i ,

A1 z =
i=1

with domain D(A1 ) = {z Z |

operator, and that A1 equals A.

i=1

|i |2 | z, i |2 < }. We shall show that A1 is a closed

Since {i , i 1} is an orthonormal basis, we have that z Z if and only if

i=1

| z, i |2 < .

Next we prove that A1 is a closed linear operator.

Assume that zn z and A1 zn y as n . Since zn z, we have that zn , i z, i
for all i N. Furthermore, since A1 zn y we have that i zn , i y, i . So i z, i =
y, i , and since y Z, this implies that z D(A1 ) and A1 z = y. Thus A1 is closed. For
z D(A1 ) it is easy to show that A1 A1 z = z, and since ran (A1 ) = D(A), this implies that
D(A1 ) D(A). Now we prove that A1 A1 z = z for all z in Z. If z is in Z, then xn :=
converges to z as n , and A1 zn =

170

i=1

z, i i

i=1

1 z, i i converges to A1 z as n (A1 is
i

A.5. Integration and dierentiation theory

continuous). So A1 zn converges and so does zn = A1 A1 zn . Thus by the closedenitioness of
A1 we have that A1 z D(A1 ) and A1 A1 z = z for all z Z. Since ran (A1 ) = D(A), we
have D(A) D(A1 ). So D(A) = D(A1 ) and A1 z = AA1 A1 z = Az for z D(A1 ) = D(A).
We conclude with an illustration of this result.
Example A.4.26 Let Z = L2 (0, 1) and let A be given by
Az =

d2 z
dx2

for z D(A),
2

d
dz
where D(A) = {z Z | z, dx absolutely continuous and dxz L2 (0, 1) with z(0) = 0 = z(1)}. It
2
is easy to verify that the inverse of A may be expressed by
1
1

(A

z)(x) =

g(x, )z( )d,

0

where
g(x, ) =

(1 )x
(1 x)

for 0 x 1
for 0 x 1.

So A1 equals the operator from Example A.4.21. In that example, we showed that A1 was
self-adjoint and compact with eigenvalues { n21 2 , n 1} and eigenvectors {sin(nx), n 1}.

Now, applying Theorem A.4.25 we see that A is closed and has the representation

Az =

n2 2 z, 2 sin(n) 2 sin(n),

with

n=1

D(A) = {z L2 (0, 1) |

n=1

n4 4 | z, 2 sin(n) |2 < }.

A.5. Integration and dierentiation theory

A.5.1. Integration theory
In this section, we wish to extend the ideas of Lebesgue integration of complex-valued functions
to vector-valued and operator-valued functions, which take their values in a separable Hilbert
space Z or in the Banach space L(Z1 , Z2 ), where Z1 , Z2 are separable Hilbert spaces. As main
references, we have used Diestel and Uhl [6], Dunford and Schwartz [8], and Hille and Phillips
[12].
Throughout this section, we use the notation for a closed subset of R, and (, B, dt) for the
measure space with measurable subsets B and the Lebesgue measure dt. It is possible to develop
a Lebesgue integration theory based on various measurability concepts.
Denition A.5.1. Let X be a Banach space. A function f : X is called simple if there
n
exist x1 , x2 , . . . , xn X and E1 , E2 , . . . , En B such that f = i=1 xi 1Ei , where 1Ei (t) = 1 if
t Ei and 0 otherwise.
Let Z1 , Z2 be two separable Hilbert spaces, and let F : L(Z1 , Z2 ) and f : Z1 .

171

A. Mathematical Background
a. F is uniformly (Lebesgue) measurable if there exists a sequence of simple functions Fn :
L(Z1 , Z2 ) such that
lim

F Fn

L(Z1 ,Z2 )

= 0 almost everywhere.

b. f is strongly (Lebesgue) measurable if there exists a sequence of simple functions fn :

Z1 such that
lim f fn Z1 = 0 almost everywhere.
n

F is strongly measurable if F z1 is strongly measurable for every z1 Z1 .

c. f is weakly (Lebesgue) measurable if f, z1 is measurable for every z1 Z1 .
F is weakly measurable if F z1 is weakly measurable for every z1 Z1 .

It is easy to see that uniform measurability implies strong measurability, which implies weak
measurability. For the case that Z is a separable Hilbert space, the concepts weak and strong
measurability coalesce.
Lemma A.5.2. For the case that Z is a separable Hilbert space the concepts of weak and strong
measurability in denition A.5.1 coincide.
Proof See Hille and Phillips [12, theorem 3.5.3] or Yosida [31, theorem in Section V.4].
We often consider the inner product of two weakly measurable functions.
Lemma A.5.3. Let Z be a separable Hilbert space, and let f1 , f2 : Z be two weakly
measurable functions. The complex-valued function f1 (t), f2 (t) dened by the inner product
of these functions is a measurable function.
Proof This follows directly from Lemma A.5.2 and denition A.5.1.
The notion of the Lebesgue integral follows naturally from the measurability concepts given
in denition A.5.1.
Denition A.5.4. Suppose that (, B, dt) is the Lebesgue measure space and that E B.

a. Let X be a Banach space and let f : X be a simple function given by f = n xi 1Ei ,

i=1
where the Ei are disjoint. We dene f to be Lebesgue integrable over E if f is Lebesgue
n
integrable over E, that is,
i=1 xi (Ei E) < , where () denotes the Lebesgue
measure of the set and we follow the usual convention that 0 = 0. The Lebesgue
n
integral of f over E is given by i=1 xi (Ei E) and will be denoted by E f (t)dt.

b. Let Z1 and Z2 be two separable Hilbert spaces. The uniformly measurable function F :
L(Z1 , Z2 ) is Lebesgue integrable over E if there exists a sequence of simple integrable
functions Fn converging almost everywhere to F and such that
lim

F (t) Fn (t)

L(Z1 ,Z2 ) dt

We dene the Lebesgue integral by

Fn (t)dt.

F (t)dt = lim
E

172

= 0.

A.5. Integration and dierentiation theory

c. Let Z be a separable Hilbert space. The strongly measurable function f : Z is
Lebesgue integrable over E if there exists a sequence of simple integrable functions fn
converging almost everywhere to f and such that
lim

f (t) fn (t)

Z dt

= 0.

We dene the Lebesgue integral by

fn (t)dt.

f (t)dt = lim
E

These integrals in the above denition are also called Bochner integrals in the literature. For
functions from R to a separable Hilbert space Z, there is a simple criterion to test whether a
function is Lebesgue integrable.
Lemma A.5.5. Let f (t) : Z, where Z is a separable Hilbert space Z. E f (t)dt is well
dened as a Lebesgue integral for E B if and only if the function z, f (t) is measurable for
every z Z and E f (t) dt < .
Proof See Hille and Phillips [12, theorem 3.7.4], noting that weak and strong measurability are
the same for separable Hilbert spaces (Lemma A.5.2).
In the case of operator-valued functions F (t) : L(Z1 , Z2 ), where Z1 and Z2 are separable
Hilbert spaces, we need to distinguish between the Lebesgue integral E F (t)dt for the case that
F (t) is uniformly (Lebesgue) measurable and the Lebesgue integral E F (t)zdt for the case that
F (t) is only strongly (Lebesgue) measurable.
Example A.5.6 Let T (t) be a C0 -semigroup on a separable Hilbert space Z. Since T (t) is
strongly continuous, it is strongly measurable. In fact, Hille and Phillips [12, theorem 10.2.1]
show that the C0 -semigroup is uniformly measurable if and only if it is uniformly continuous. Now
the only uniformly continuous semigroups are those whose innitesimal generator is a bounded
operator, Hille and Phillips [12, theorem 9.4.2], and so T (t) will only be strongly measurable
1
1
in general. Thus 0 T (t)zdt is a well dened Lebesgue integral for any z Z, but 0 T (t)dt is
not.

Example A.5.7 Next consider 0 T ( s)F (s)ds, where T (t) is a C0 -semigroup on a separable
Hilbert space Z, F () L(U, Z), U is a Hilbert space, F is weakly measurable, and F
L1 (0, ). Since T (t) is also a C0 -semigroup, T (t)z is continuous and so strongly measurable.
Furthermore, by denition, we have that F (s)u is weakly measurable. Hence Lemma A.5.3 shows
that z, T ( s)F (s)u = T ( s)z, F (s)u is measurable in s for all z Z, u U . So from

Lemma A.5.5 we have that for each u U 0 T ( s)F (s)uds is a well dened Lebesgue integral.

However, 0 T ( s)F (s)ds need not be a well dened Lebesgue integral, since the integrand
will not to be uniformly measurable in general.
This example motivates the need for a weaker concept of integration based on weak measurability. We now introduce the Pettis integral, which satises this requirement.
Lemma A.5.8. Let Z1 and Z2 be separable Hilbert spaces, and let F (t) : L(Z1 , Z2 ).
Assume furthermore, that for every z1 Z1 and z2 Z2 the function z2 , F (t)z1 is an element
of L1 (). Then for each E B, there exists a unique zF,E (z1 ) Z2 satisfying
z2 , F (t)z1 dt.

z2 , zF,E (z1 ) =
E

173

A. Mathematical Background
Proof Set
z2 , F (t)z1 dt.

G(z2 ) =
E

It is clear that G is well dened for every z2 Z2 , and that it is linear on Z2 . It remains to show
that G is bounded. To do this, we dene the following operator from Z2 to L1 (E) : Q(z2 ) =
n
n
z2 , F (t)z1 and we show that it is closed. This follows since if z2 z2 and z2 , F (t)z1 h(t)
n
in L1 (E), we have z2 , F (t)z1 z2 , F (t)z1 everywhere on E, and so z2 , F (t)z1 = h(t). Thus
Q is a closed linear operator with domain Z2 ; so with the Closed Graph Theorem A.3.49, we
conclude that Q is a bounded linear operator. Thus
| G(z2 ) |

| z2 , F (t)z1 | dt Q

z2 ,

and G is bounded. Applying the Riesz Representation Theorem A.3.52, we obtain the existence
of a zF,E (z1 ) such that
z2 , F (t)z1 dt.

z2 , zF,E (z1 ) = G(z2 ) =

(A.1)

In the next lemma, we shall show that zF,E (z1 ) denes a bounded linear operator from Z1 to
Z2 .
Lemma A.5.9. The mapping zF,E (z1 ) in (A.1) is a linear function of z1 , and zF,E denes a
bounded linear operator from Z1 to Z2 .
Proof The linearity of zF,E (z1 ) in z1 follows easily from the uniqueness of zF,E (z1 ). The boundedenitioness will follow from the closedenitioness of the operator
z1 zF,E (z1 ).
n
n
If z1 z1 in Z1 and zF,E (z1 ) z E , then for all z2 Z2 we have
n
E
z2 , zF,E (z1 ) z2 , z2 ,

and
n
z2 , zF,E (z1 )

=
=

n
z2 , F (t)z1
E
n
z1 , zF ,E (z2 )

dt =

z1 , zF ,E (z2 )

n
z1 , F (t)z2 dt

as n

z2 , F (t)z1 dt = z2 , zF,E (z1 ) ,

=
E

where we have used Lemma A.5.8. Thus z E = zF,E (z1 ).

In fact, we have established the following result.
Theorem A.5.10. Let Z1 and Z2 be separable Hilbert spaces, and let F (t) : L(Z1 , Z2 ).
Assume further that for all z1 Z1 and z2 Z2 , the function z2 , F (t)z1 is an element of L1 ().
Then for each E B, there exists a bounded linear operator, denoted by zF,E , satisfying
z2 , F (t)z1 dt.

z2 , zF,E z1 =
E

174

A.5. Integration and dierentiation theory

This last result leads naturally to the denition of the Pettis integral.
Denition A.5.11. Let Z1 and Z2 be separable Hilbert spaces and let F : L(Z1 , Z2 ). If
for all z1 Z1 and z2 Z2 the function z2 , F (t)z1 L1 (), then we say that F () is Pettis
integrable. Furthermore, for all E B, we call E F (t)dt dened by
z2 , F (t)z1 dt.

F (t)dtz1 := z2 , zF,E z1 =

z2 ,

(A.2)

the Pettis integral of F (t) over E and

F (t)z1 dt the Pettis integral of F (t)z1 over E.

One can easily prove the usual properties such as linearity of the integral
F2 (t)dt.

F1 (t)dt +

(F1 (t) + F2 (t)) dt =

(A.3)

From the denition of the Pettis integral, we always have that

| z2 , F (t)z1 |dt < .

(A.4)

In particular, if E F (t) dt < , then the condition (A.4) is satised. Furthermore, it is easy
to see that if F is an integrable simple function, then the Pettis integral equals the Lebesgue
integral. From the denition of the Lebesgue integral, it follows easily that if the Lebesgue
integral of a function exists, then the Pettis integral also exists, and they are equal.
In the next example we shall show that a function may be Pettis integrable, but not Lebesgue
integrable.
Example A.5.12 Let Z be 2 from Example A.2.9 and dene en to be the nth-basis vector.
Dene the function
1
for n 1 t < n.
f (t) = en
n
It is easy to show that

f (t) dt =
0

1
= .
n
n=1

So by Lemma A.5.5 we see that f is not Lebesgue integrable. On the other hand, we have that
for any z 2 ,

| z, f (t) |dt =

1
| z, en |
n
n=1

1
n2
n=1

n=1

| z, en |2 = c z .

So z, f (t) L1 (0, ), and we conclude that the Pettis integral exists.

In the following example, we re-examine Examples A.5.6 and A.5.7, which we considered as
Lebesgue integrals.
Example A.5.13 We recall from Example A.5.6 that the C0 -semigroup T (t) on the separable
1
Hilbert space Z is in general only strongly measurable and so while 0 T (t)zdt exists as a Lebesgue
1
integral 0 T (t)dt does not. We show that it does exist as a Pettis integral. Since T (t) is strongly
continuous, we have that z1 , T (t)z2 is measurable for every z1 , z2 Z. From Theorem 2.5.1

175

A. Mathematical Background
1

we have that 0 T (t) dt < . Thus by denition A.5.11 the Pettis integral 0 T (t)dt is well
dened. If the innitesimal generator A of T (t) is invertible, then using Theorem 2.5.2 we can
even calculate this Pettis integral to obtain
1

T (t)dt = A1 T (1) A1 .

Example A.5.14 From Example A.5.7 we recall that 0 T ( s)F (s)ds was not a well dened
Lebesgue integral. There we already showed that z, T ( s)F (s)u is Lebesgue measurable for
all z Z, u U . Furthermore, we see that

T ( s)F (s) ds M e

So by denition A.5.11 the integrals

as Pettis integrals.

F (s) ds < .

T ( s)F (s)ds and

T ( s)F (s)uds are well dened

Most of the integrals we use in this text satisfy the conditions in Lemma A.5.5, and so we may
speak about the integral, as in the following example.

Example A.5.15 Consider 0 T ( s)Bu(s)ds, where T (t) is a C0 -semigroup on a separable

Hilbert space Z, B L(U, Z), U is a separable Hilbert space and u L1 ([0, ]; U ) (see denition
A.5.16). Then, as in Example A.5.14, z, T ( s)Bu(s) is measurable in s for all z Z and

T ( s)Bu(s) ds M e B 0 u(s) ds < . So by Lemma A.5.5, the integral is well

0
dened as a Pettis or as a Lebesgue integral.
To avoid confusion between the Pettis and Lebesgue integrals we introduce the following
notation.
Denition A.5.16. Let Z1 , Z2 , and Z be separable Hilbert spaces, and let be a closed subset
of R. We dene the following spaces:
P (; L(Z1 , Z2 ))
Pp (; L(Z1 , Z2 ))

:= {F : L(Z1 , Z2 ) | z2 , F ()z1 is measurable for every z1 Z1 and z2 Z2 }.

:= {F P (; L(Z1 , Z2 )) | F
F (t)

P (; L(Z1 , Z2 ))

:=

1/p

p
L(Z1 ,Z2 )

< }; 1 p < .

:= {F P (; L(Z1 , Z2 )) | F
ess sup F (t)

L(Z1 ,Z2 )

:=

< }.

L(; Z) := {f : Z | z, f () is measurable for all z Z}.

1/p

Lp (; Z)
L (; Z)

176

:= {f L(; Z) | f
1 p < .
:= {f L(; Z) | f

:=

f (t)

p
Z dt

:= ess sup f (t)

< };
Z

< }.

A.5. Integration and dierentiation theory

The reason for using the L notation is that these integrals are also dened in the Lebesgue
sense. For example, if T (t) is a strongly continuous semigroup, then T (t)z Lp ([0, ]; Z) for all
z Z, but we only have that T (t) Pp ([0, ]; L(Z)) instead of the Lebesgue space Lp ([0, ]; L(Z))
(see Example A.5.13).
We remark that if Z1 and Z2 are nite-dimensional, then L(Z1 , Z2 ) is also nite-dimensional, and so L (; L(Z1 , Z2 )) is well dened as a Lebesgue space (see Lemma A.5.5) and equals
P (; L(Z1 , Z2 )).
Lemma A.5.17. If we do not distinguish between two functions that dier on a set of measure
zero, then the spaces Pp (; L(Z1 , Z2 )), P (; L(Z1 , Z2 )), Lp (; Z), and L (; Z) are Banach
spaces.
Furthermore, L2 (; Z) is a Hilbert space with inner product
h, f =

h(t), f (t)

Z dt.

(A.5)

Proof See Thomas [26] or [27].

The completeness property of Lp is also shown in theorem III.6.6 of Dunford and Schwartz
[8].
In Section 3.5 of Balakrishnan [2] it is shown that L2 (, Z) is a Hilbert space.
It is interesting to remark that Pp is not a Banach space under the norm
1/p

:=

sup
z1 =1, z2 =1

| F (t)z1 , z2 |dt

see Thomas [27].

From Lemmas A.5.8 and A.5.9 it should be clear that the integrals share the usual properties
of their nite-dimensional Lebesgue counterparts.
Theorem A.5.18. If f P1 (; L(Z1 , Z2 )), where Z1 and Z2 are separable Hilbert spaces, then
the following hold:
a.
b.

f (t)dt

lim

(E)0 E

f (t) dt;

f (t)dt = 0, where (E) denotes the Lebesgue measure of E B.

Proof a. This follows easily from denitions A.5.16 and A.3.9 and Lemma A.3.30, since from
these last results, it follows that
f (t)dt

sup

| z2 ,

z1 Z1 ,z2 Z2

sup
z1 Z1 ,z2 Z2

sup
z1 Z1 ,z2 Z2

sup
z1 Z1 ,z2 Z2

f (t)dtz1 |
z1 z2

z2 , f (t)z1
dt
z1 z2

| z2 , f (t)z1 |
dt
z1 z2
| z2 , f (t)z1 |
dt =
z1 z2

f (t) dt.

b. This follows directly from part a and the standard Lebesgue theory.

177

A. Mathematical Background
Lemma A.5.19. Let Z be a separable Hilbert space and let a and b be real numbers such that
< a < b < . Lp ([a, b]; Z) has the following dense subspaces:
a. The space of all continuous functions on [a, b], C([a, b]; Z);
b. The space of all piecewise constant functions that are functions of the form f (x) =
n

i=0

zi 1[ai ,bi ] (x) with a = a0 < b0 a1 . . . bn = b and zi Z.

Proof See page 86 of Hille and Phillips [12].

Lemma A.5.20. Let Z be a separable Hilbert space and let 1 p < . Lp ((, ); Z) has
the following dense subspaces:
a. The functions in Lp ((, ); Z) that are zero outside some nite interval;
b. Lp ((, ); Z) Lq ((, ); Z) for every q 1.
Proof a. Let f be an arbitrary function in Lp ((, ); Z). For suciently large N we have
that fN (t) := f (t)1[N,N ](t) is arbitrarily close to f in the Lp -norm, since
f fN

1/p

f (t) dt

f (t) dt +
N

converges to zero as N approaches .

b. Let > 0 and let N be chosen such that f fN p < , where f and fN be the same as
in part a. Now we have that fN Lp ([N, N ]; Z), so by Lemma A.5.19 there exists a function
gN C([N, N ]; Z) such that fN gN Lp ([N,N ],Z) < . Since gN is continuous, it is easy to
see that it is an element of Lq ((N, N ); Z). Now we dene
g(t) =

gN (t)
0

for t [N, N ]
for t [N, N ]

and so g Lp ((, ); Z) Lq ((, ); Z). Furthermore,

gf

g fN

+ fN f

< 2.

was arbitrary, so we have proved the result.

We remark that part a of this lemma is false for L .
Theorem A.5.21. Lebesgue-Dominated Convergence Theorem. Let Z be a separable
Hilbert space and let fn be a sequence in L1 (; Z). Suppose that fn converges almost everywhere
to f , i.e., lim fn (t) f (t) = 0 except for t in a set of measure zero. Assume further that
n

there exists a xed function g L1 () such that fn (t) g(t) for all n and almost all t .
Then f L1 (; Z) and
f (t)dt

fn (t)dt =

lim

for all E B.
Proof See theorem 3 on page 45 of Diestel and Uhl [6] or theorem III.3.7 in Dunford and
Schwartz [8].

178

A.5. Integration and dierentiation theory

Theorem A.5.22. Fubinis Theorem. Let (1 , B1 , dt), (2 , B2 , dt) be two Lebesgue measure
spaces with (1 ) < and (2 ) < . We denote by B1 B2 the -algebra of subsets of 1 2
generated by the class of all rectangular sets of the form E F , where E B1 , F B2 and the
product measure is denoted by dt ds.
For f (, ) L1 (1 2 , Z) the functions
1

f (t, )dt and

f (, s)ds
2

are in L1 (2 ; Z) and L1 (1 ; Z), respectively, and

1 2

f (t, s)dt ds =

f (t, s)dt)ds.

f (t, s)dt)ds =

(
1

Proof See Hille and Phillips [12, theorem 3.7.13].

Theorem A.5.23. Let Z1 and Z2 be separable Hilbert spaces, and let A be a closed linear
operator from D(A) Z1 to Z2 . If f L1 (; Z1 ) with f D(A) almost everywhere and
Af L1 (; Z2 ), then
Af (t)dt

f (t)dt =

for all E B.
Proof See Hille and Phillips [12, theorem 3.7.12].
Example A.5.24 Let a < b and let Z be a separable Hilbert space. Assume
further that {fn , n 1} and {em , m 1} are orthonormal bases for L2 (a, b) and Z, respectively.
We show that {n,m , n, m 1} with n,m := fn em is an orthonormal basis for L2 ([a, b]; Z).
From the denition of the inner product on L2 ([a, b]; Z), (A.5), we have that
b

n,m , i,j

fn (t)em , fi (t)ej

Z dt

fn (t)fi (t) em , ej

Z dt

a
b

fn (t)fi (t)mj dt = mj fn , fi

L2 (a,b)

= mj ni .

Thus {n,m , n, m 1} is an orthonormal set. Next we show that it is maximal. If z is orthogonal

to every n,m , then
b

fn (t)em , z(t)

Z dt

=0

for all n, m 1.

If we x m, then we see that for all n 1,

b

fn (t) em , z(t)

Z dt

= 0.

But fn is maximal in L2 (a, b), and so em , z(t) Z = 0 almost everywhere. This holds for all
m 1. Now using the fact that em is maximal in Z, we obtain that z(t) = 0 almost everywhere.
Thus z = 0 in L2 ([a, b]; Z), which concludes the proof.

179

A. Mathematical Background

A.5.2. Dierentiation theory

In the previous subsection, we concentrated on the integration of Hilbert-space-valued functions.
However, as is known from standard calculus, integration is naturally related to dierentiation,
and in this subsection we summarize standard results on dierential calculus for Hilbert-spacevalued functions. We start with the concept of the Frchet derivative.
e
Denition A.5.25. Consider the mapping U from the Banach space X to the Banach space
Y . Given x X, if a linear bounded operator dU (x) exists such that
h

lim

X 0

U (x + h) U (x) dU (x)h
h X

= 0,

then U is Frchet dierentiable at x, and dU (x) is said to be the Frchet dierential at x.

e
e
It is easy to see that if U is identical to a bounded linear operator, then dU (x) = 0 for every
x X.
One of the most important applications of the derivative is the determination of the maxima
and minima of functionals.
Theorem A.5.26. Let O be an open subset of the Banach space X. If the mapping f : O R
has a minimum or a maximum at x O, and df (x) exists, then df (x) = 0.
Proof We shall only give the proof for the case that f has a minimum. The proof for the other
case follows easily by replacing f by f .
For suciently small h we have that x + h and x h are in O. Furthermore, we have that
f (x + h) f (x) df (x)h
and
f (x h) f (x) df (x)h.
Since x is a minimum, the left-hand side of both equations is nonnegative. Looking at the righthand side of these equations we conclude that df (x) must be zero.

Most of the applications of dierential calculus in this book are to functions from R or C to
the Banach space X. Since this is frequently used we shall give a special denition for functions
of this class
Denition A.5.27. A function f : R X is dierentiable if f is Frchet dierentiable
e
lim

h0

f (t + h) f (t) df (t)h
= 0.
|h|

We shall denote the derivative of f at t0 by

df
dt (t0 )

or f(t0 ).

(A.6)

In applications, we apply denition A.5.27 to a function f (x, t) of two variables by considering

it to be a function of t taking its values in an appropriate function space, corresponding to the
Banach space X. However, this Frchet derivative may exist, whereas the usual partial derivative
e
does not, as the following example shows.

180

A.5. Integration and dierentiation theory

Example A.5.28 Consider the function f (x, t) : [0, 1] [1, 1] R dened by
f (x, 0) = 0
t
0

f (x, t) =

1
for |x + [ 1 ] 1 | < 2 |t|
t
t
elsewhere,

where [ 1 ] denotes the integer part of 1 , that is, the largest integer smaller than or equal to
t
t
We have that f (, t) L2 (0, 1) for every t [1, 1] and

1
1
1
min{1, 2 |h|[ h ]+ h }

|f (x, h) f (x, 0)|2 dx =

1
t.

|h|2 dx |h|3 ,

|f (x, h)|2 dx =
1
1
1
max{0, 2 |h|[ h ]+ h }

since the length of the integration interval is smaller than |h|. From this it is easy to see that the
Frchet derivative at t = 0 exists and equals 0. Now we shall show that the partial derivative of
e
f with respect to t at t = 0 does not exist for any x [0, 1]. Let x be an element of [0, 1] and
1
1
consider the sequences {tn , n 1} with tn = n+x and {n , n 1} with n = n . Both sequences
converge to zero, and for suciently large n f (x, tn ) = tn and f (x, n ) = 0. So we have that
lim

f (x, tn ) f (x, 0)
= 1,
tn

and

f (x, n ) f (x, 0)
= 0.
n
Hence the partial derivative with respect to t does not exist at t = 0.
lim

The next theorem concerns dierentiation of integrals.

Theorem A.5.29. Let u : [0, ) Z be such that u L([0, ), Z), where Z is a separable
t
Hilbert space. If v(t) = 0 u(s)ds, then v is dierentiable for almost all t, and
dv
(t) = u(t)
dt

almost everywhere.

Proof See corollary 2 on page 88 in Hille and Phillips [12].

For operator-valued functions we can dene three types of dierentiability.
Denition A.5.30. Let U () be functions from C or R to L(Z1 , Z2 ), where Z1 and Z2 are
Hilbert spaces. Then
a. U () is uniformly dierentiable at t0 if there exists a

lim

h0

U (t0 + h) U (t0 ) h

dU
(t0 )
dt

h0

L(Z1 ,Z2 )

|h|

b. U () is strongly dierentiable at t0 if there exists a

lim

dU
dt (t0 ) L(Z1 , Z2 )

dU
dt (t0 )

U (t0 + h)z1 U (t0 )z1 h

|h|

such that

= 0.

L(Z1 , Z2 ) such that

dU
(t0 )z1
dt

Z2

= 0,

for every z1 Z1

181

A. Mathematical Background
c. U () is weakly dierentiable at t0 if there exists a

lim

dU
dt (t0 )

L(Z1 , Z2 ) such that

| z2 , U (t0 + h)z1 z2 , U (t0 )z1 h z2 ,

|h|

h0

dU
(t0 )z1 |C
dt
= 0,

for every z1 Z1 and z2 Z2

One can easily show that uniform implies strong, which implies weak dierentiability, with
the same derivative. Furthermore, by the Riesz representation Theorem A.3.52 one can easily
show that U () is weakly dierentiable at t0 if and only if the complex-valued functions f (t) =
z2 , U (t)z1 are dierentiable at t0 , for every z1 Z1 and z2 Z2 .
The next example shows that strong dierentiability does not imply uniform dierentiability.
Example A.5.31 Let Z be a Hilbert space with orthonormal basis {en , n 1}, and let Vn
denote the orthogonal complement of span{e1 , . . . , en }. Dene the operator-valued function U ()
by

if t 0 or t 1
0

U (t) =
1
1

tPV
t ,
if
n
n+1
n
where PVn denotes the orthogonal projection on Vn . Then for

1
1+n

h<

1
n

we have

z, ei ei .

U (h)z = h
i=n+1

Thus
U (h)z U (0)z
=
|h|

z, ei ei =
i=n+1

1/2

i=n+1

| z, ei |

and so U () is strongly dierentiable at 0, with derivative 0. However,

U (h) U (0)
|h|

dU
(0)
U (h)
dt
=
= PVn = 1.
|h|

So U () is not uniformly dierentiable at zero.

The situation is dierent for operator-valued functions of a complex variable. As in nite
dimensions, we dene holomorphicity of a complex-valued function as dierentiability.
Denition A.5.32. Let Z1 and Z2 be Hilbert spaces, and let U : L(Z1 , Z2 ), where is a
domain in C. Then U is holomorphic on if U is weakly dierentiable on .

Example A.5.33 Let A be a closed linear operator on the Hilbert space Z. Dene U () :
(A) L(Z) by U () = (I A)1 . We shall prove that this is holomorphic on (A). We have
from the resolvent equation (A.5) that
z1 , (( + h)I A)1 z2 z1 , (I A)1 z2
= z1 , h(I A)1 (( + h)I A)1 z2 .
This implies that U () is weakly dierentiable with dU () = (I A)2 . Thus the resolvent
d
operator is holomorphic, and this proves Lemma A.4.8.c.

182

A.6. Frequency-domain spaces

The following important theorem shows the equivalence of uniform and weak holomorphicity.
Theorem A.5.34. Let Z1 and Z2 be separable Hilbert spaces and let U () : L(Z1 , Z2 ),
where is a domain of C. If U is holomorphic, then U () is uniformly dierentiable in and
furthermore,
d
dU
z2 , U (t)z1 = z2 ,
z1 ,
dt
dt
for every z1 Z1 , z2 Z2 .
Proof See Hille and Phillips [12, theorem 3.10.1].
We remark that the above result is also valid in a general Banach space; for more details
see Hille and Phillips [12]. With this result is easy to extend results that hold for holomorphic
functions f : C C to Hilbert-space-valued holomorphic functions f : C Z.
Example A.5.35 Let be a domain in C, and let be a positively oriented, closed, simple
contour in . Consider a holomorphic function f on with values in a separable Hilbert space
Z. Then we have the following relation
f (s)
ds = f (),
s

1
2

where is any point inside .

First, we have to say what we mean by the integral on the left-hand side. Since is a
rectiable, closed, simple curve there exists a dierentiable mapping k from [0, 1] onto . The
integral is then dened to be
1

1
2
0

f (k(t))
k(t)dt.
k(t)

This is well dened as a Pettis or Lebesgue integral by Lemma A.5.5.

From Theorem A.5.10, for every z Z the following holds:
1
2

1
2

f (s)
ds, z1
s

1
2

f (s)
, z1 ds
s
1
f (s), z1 ds = f (), z1 ,
s

since f (s), z1 is a holomorphic function. This proves the assertion. This result is known as
Cauchys theorem.

A.6. Frequency-domain spaces

A.6.1. Laplace and Fourier transforms
In this book, we consider both state- and frequency-domain representations. The relation between these two representations is provided by the Laplace or Fourier transform. In this section,
we take Z to be a separable Hilbert space.

183

A. Mathematical Background
Denition A.6.1. Let h : [0, ) Z have the property that et h(t) L1 ([0, ); Z) for some

real . We call these Laplace-transformable functions and we dene their Laplace transform h
by

h(s) =

est h(t)dt

(A.1)

for s

C+

:= {s C | Re(s) }.

A good reference for Laplace transforms of scalar functions is Doetsch [7], and for vector-valued
functions a good reference is Hille and Phillips [12], where it is shown that has the following
h
properties.
Proposition A.6.2. Laplace transformable functions h : [0, ) Z have the following properties:
a. If et h(t) L1 ([0, ); Z) for some real , then is holomorphic and bounded on C+ :=
h

+ ) H (Z) (see denition A.6.14); furthermore, the

{s C | Re(s) > } and so h(
following inequality holds:
sup
Re(s)0

h(s + ) e h()

L1 ([0,);Z) ;

(A.2)

b. Uniqueness of the Laplace transform: if h1 and h2 are Laplace transformable functions

such that 1 (s) = h2 (s) in C+ , for some R, then h1 = h2 ;

h

c. If et h(t) L1 ([0, ); Z) for some real , then h( + ) is continuous in for R

+ ) 0 as || ;
and h(
d. If h is dierentiable for t > 0 and

dh
dt

is Laplace-transformable, then

dh

(s) = sh(s) h(0+ );

dt
e. For R it holds that

[et h(t)](s) = h(s + );

(A.3)

(A.4)

f. If et h(t) L1 ([0, ); Z) for some real , then the derivative of h in C+ equals the

Laplace transform of th(t);

g. If et h(t) L1 ([0, ); Z) for some real , then h(s) 0 as |s| in C+ , i.e.,

lim

sup

{sC+ ||s|>}

h(s) = 0.

Proof a. See Doetsch [7] for the scalar case. For the general case, consider the scalar functions

hz (t) := h(t), z . It is easy to see that hz (s) = h(s), z . This function is holomorphic for every
z Z and so by denition A.5.32, h is holomorphic. Furthermore, for Re(s) > 0,

h(s + )

184

e
0

(s+)t

h(t) dt

et h(t) dt,
0

A.6. Frequency-domain spaces

where we have used Theorem A.5.18.
b. See Hille and Phillips [12, theorem 6.2.3].
c. For the scalar case see Bochner and Chandrasekharan [3, theorem 1]. This proof is based on
the denseness of the simple functions that are zero outside some interval. However, this fact also
holds for L1 ([0, ); Z) (see Lemma A.5.20) and so a similar proof is valid for the vector-valued
case.
This property is known as the Riemann-Lebesgue lemma.
d and e. The proof of these properties is similar in the scalar and nonscalar cases; see Doetsch
[7] theorems 9.1 and 7.7, respectively.
f. For the scalar case see theorem 6.1 of Doetsch [7]. The general case is proven again by

introducing the functions hz (t) := h(t), z with Laplace transform z (s) = h, z . From the
h
z equals the Laplace transform of thz (t). Since h

scalar case we know that the derivative of h

d

is holomorphic we know by Theorem A.5.34 that the derivative of hz equals ds h(s), z . Hence
d
ds h(s), z is the Laplace transform of th(t), z . Since this holds for any z Z, the assertion
is proved.
g. This follows essentially from part c, see Doetsch [7, theorem 23.7].
The Laplace transform can be seen as a special case of the Fourier transform.
Denition A.6.3. For h L1 ((, ); Z) we dene the Fourier transform of h by

h() :=

et h(t)dt.

(A.5)

In fact, the Fourier transform can be extended to functions in L2 (R; Z); see Theorem A.6.13.
If h has support on R+ and its Fourier transform exists, then it is equal to its Laplace transform

h() = h().

(A.6)

Sometimes it is convenient to introduce the two-sided Laplace transform for functions h dened

est h(t)dt. This then coincides with the Fourier transform

on all of R, that is,
h(s) :=

using (A.6). This connection with the Fourier transform makes it easy to deduce some further
properties of the Laplace transform.
The reason for using Laplace transforms in linear dierential equations lies in Property A.6.2.d
and in the simple property for the convolution product. In this book we shall only need the scalar
version.
Denition A.6.4. For two functions h, g in L1 (, ), we dene the convolution product

(h g)(t) :=

h(t s)g(s)ds.

(A.7)

Note that if h and g have their support on [0, ), then h g also has its support on [0, ),
and for t 0 the convolution product is given by
t

(h g)(t) :=

h(t s)g(s)ds.

(A.8)

The following are very useful properties of the convolution product.

185

A. Mathematical Background
Lemma A.6.5. For two functions h, g from R to C the following hold:
a. If h L1 (, ), g Lp (, ), then h g Lp (, ) and
hg

(A.9)

for 1 p ;
b. If h L1 ([0, )) and g Lp ([0, )), then h g Lp ([0, )) and
hg

(A.10)

for 1 p ;
c. If h and g are zero on (, 0) and are Laplace transformable, then h g is Laplace
transformable and
g
h g = h;
(A.11)
d. If h and g are in L1 (, ) L2 (, ), then

g
h g= h.

(A.12)

Proof a. See theorem 53 of Bochner and Chandrasekharan [3].

b. This follows from part a by dening extended functions on (, )
h(t) t 0,
0
t<0

he (t) =

and ge similarly. See also the remark made after denition A.6.4.
c and d. See theorems 10.2 and 31.3 of Doetsch [7].
The denition of the convolution product for real- or complex-valued functions as given in
denition A.6.4 can easily be extended to vector-valued functions, and similar results to those
given in Lemma A.6.5 hold.
Lemma A.6.6. Let Z1 and Z2 be separable Hilbert spaces, H P1 ((, ); L(Z1 , Z2 )), and
g Lp ((, ); Z1 ) for a p, 1 p . The convolution product of H and g is dened by

(H g)(t) :=

H(t s)g(s)ds.

(A.13)

Furthermore, H g Lp ((, ); Z2 ) and

H g

g p.

(A.14)

Proof Let z Z2 . Consider the function in s dened by z, H(t s)g(s) . This function
is the same as H (t s)z, g(s) , which is measurable by Lemma A.5.3. Furthermore, since
H(t s)g(s) H(t s) g(s) , we have that integral (A.13) is well dened; see Lemma
A.5.5. Now inequality (A.14) follows directly from Lemma A.6.5.
Inequalities (A.9) and (A.10), together with the Cauchy-Schwarz inequality, are used frequently in this book. Another useful inequality is given in the next lemma.

186

A.6. Frequency-domain spaces

Lemma A.6.7. Gronwalls Lemma. Let a L1 (0, ), a(t) 0. If for some 0 the
function z L (0, ) satises
t

z(t) +

a(s)z(s)ds,
0

then

z(t) exp(

a(s)ds).
0

Proof See the lemma on page 169 and problem 8 on page 178 of Hirsch and Smale [13].

A.6.2. Frequency-domain spaces

In the text we frequently make use of the following Lebesgue frequency-domain spaces (see
Appendix A.5).
Denition A.6.8. Let Z, U , and Y be separable Hilbert spaces. We dene the following
frequency-domain spaces:
P ((, ); L(U, Y ))
:= {G : (, ) L(U, Y ) | y, Gu is
measurable for every u U
and y Y and G is bounded almost
everywhere on (, )};
L2 ((, ); Z) :=

{z : (, ) Z | z, x is
measurable for all x Z and

z()

2
Z d

< }.

(A.15)

(A.16)

We remark that in the case that U and Y are nite-dimensional spaces, we may write
L ((, ); L(U, Y )) instead of P ((, ); L(U, Y )); see also the remark after denition A.5.16.
From Lemma A.5.17 we deduce the following results.
Lemma A.6.9. If Z, U , and Y are separable Hilbert spaces, P ((, ); L(U, Y )) is a
Banach space under the P -norm dened by
G

= ess sup<< G()

L(U,Y )

(A.17)

and L2 ((, ); Z) is a Hilbert space under the inner product

z1 , z2

1
:=
2

z1 (), z2 ()

Z d.

(A.18)

We remark that the inner product on L2 ((, ); Z) diers by a constant from the inner
product dened in Lemma A.5.17. This is important in the theory of Fourier transforms, as can
be seen from equation (A.24).
If U and Y are nite-dimensional, then we have a more explicit expression for the P -norm.

187

A. Mathematical Background
Lemma A.6.10. Let U and Y be nite-dimensional Hilbert spaces. The P -norm for the space
P ((, ); L(U, Y )) is given by
G

= ess sup<< max (G()) = ess sup<< {1/2 (G() G())},

max

(A.19)

where max and max denote the maximum singular value and eigenvalue, respectively (see
Example A.3.13).
Continuous functions in L ((, ); L(U, Y )) can be approximated by strictly proper rational ones. By a strictly proper rational matrix we mean a quotient of polynomials such that it
has the limit zero at innity.
Lemma A.6.11. Let U and Y be nite-dimensional Hilbert spaces. If F L ((, );
L(U, Y )) is continuous and
lim F () = lim F () = 0,

then F can be approximated in the L -norm by strictly proper rational matrices with no poles
on the imaginary axis.
Proof We reduce this to an equivalent problem on the unit circle, by introducing the bilinear
transformation : C C dened by
(z) :=

1+z
1z

for z C\{1}.

It is easy to see that it maps the unit circle excluding the point 1 onto the imaginary axis.
Furthermore, it is easy to see that Fd (z) := F ((z)) is bounded and continuous on the unit
circle except for the point 1. For the point 1, we have
lim

|z|=1,z1

Fd (z) =

lim

R,||

F () = 0.

Hence Fd is continuous on the unit circle.

From the Weierstrass Theorem [22, theorem 7.26], we have that for every > 0 there exists a
polynomial Q such that
sup |Fd (z) Q (z)| < .
|z|=1

Since Fd (1) = 0, |Q (1)| < . Dening P := Q Q (1) gives P (1) = 0 and

sup |Fd (z) P (z)| < 2.

|z|=1

Now using the bilinear transformation again, we see that

sup |F () P (1 ())| = sup |Fd (z) P (z)| < 2.

|z|=1

The function P (1 ()) is a rational function with no poles on the imaginary axis. Furthermore,
we have that
lim
P (1 ()) = lim P (z) = 0,
R,||

and so P (

|z|=1,z1

()) is strictly proper.

The following lemma enables us to reduce many properties of the vector-valued function space
L2 (; Z) to analogous ones of the scalar function space L2 ().

188

A.6. Frequency-domain spaces

Lemma A.6.12. Let Z be a separable Hilbert space with an orthonormal basis {en , n 1}.
For every f L2 (, Z) there exists a sequence of functions fn L2 () such that

f (t) =

fn (t)en

(A.20)

n=1

and

2
L2 (,Z)

fn

2
L2 () .

(A.21)

n=1

On the other hand, if {fn , n 1} is a sequence of functions in L2 () such that

fn
n=1

2
L2 ()

< , then f dened by (A.20) is in L2 (, Z) and (A.21) holds.

Proof For f L2 (, Z), it is easy to see that the function fn dened by fn (t) := fn (t), en
is in L2 (). Furthermore, (A.20) holds, since {en , n 1} is an orthonormal basis for Z (see
denition A.2.33). The norm equality follows from this because

2
L2 (,Z)

2
Z dt

f (t)

fn (t) 2 dt

=
n=1

since the summant is positive

fn (t) 2 dt =

fn

2
L2 () .

n=1

n=1

For the other assertion, dene the sequence of functions gn :=

that gn L2 (, Z), and furthermore, for N > n,

n
k=1

fk ek . It is easy to see

gN gn

2
L2 (,Z)

fk

2
L2 () .

k=n+1

k=1

Since
fk 2 2 () < , this implies that gn is a Cauchy sequence in L2 (, Z). Thus gn
L
converges, and now it is easy to show that (A.21) holds.
We are interested in the relationships between the frequency-domain spaces L2 ((, ); Z)
and P ((, ); L(U, Y )) dened in denition A.6.8 and their time-domain counterparts
L2 ((, ); Z) and P ((, ); L(U, Y )), respectively (see Appendix A.5).
Theorem A.6.13. The frequency-domain space L2 ((, ); Z) is isomorphic to the timedomain space L2 ((, ); Z) via the Fourier transform. So the Fourier transform gives an

isometry from L2 ((, ); Z) to L2 ((, ); Z). If h is the Fourier transform of h, then h

can be recovered via the inversion formula
1
h(t) =
2

et h()d.

(A.22)

From the isometry we know that the norms are equivalent, and this equivalence is usually known
as Parsevals equality

h(t)

2
Z dt

1
=
2

h()

2
Z d.

(A.23)

189

A. Mathematical Background
In addition, for h, g L2 ((, ); Z),

1
h(t), g(t) Z dt =
2

h(), g ()

Z d.

(A.24)

So h, g = h, g 2 , where the rst inner product is dened in Lemma A.5.17 and the second in
Lemma A.6.9.
Proof a. First we have to show that we can extend the Fourier transform to functions in
L2 (, ); Z). Let h L1 ((, ); Z) L2 ((, ); Z) and let {en , n 1} be an orthonormal basis of Z. From (A.5) it is easy to see that

hn () := h(), en = h, en ().
Furthermore, we know from Rudin [24, theorem 9.13] that

1
| h, en (t)| dt =
2

|hn ()|2 d.

So with Lemma A.6.12 we see that h L2 ((, ); Z) and (A.23) holds. This implies that
we can extend the denition of the Fourier transform to functions in L2 ((, ); Z) and that
equality (A.23) still holds.
So we have established that the Fourier transform maps the time-domain space
L2 ((, ); Z) isometrically into the frequency-domain space L2 ((, ); Z). It remains
to show that this mapping is onto, and that inversion formula (A.22) holds.

b. As in part a, one can show that for h L1 ((, ); Z) L2 ((, ); Z) the mapping
1

h
2

et h()d

is well dened and that

h (t)

2
Z dt

1
2

h()

2
Z d.

So this mapping can be extended to hold for all h L2 ((, ); Z). Furthermore, we know
from Rudin [24, theorem 913] that is the inverse Fourier transform for functions of the form

hn en . Since, is linear this implies that is the inverse of the Fourier transform for all h. Thus
the Fourier transform is an isometry between the time- and frequency-domain L2 spaces.
Equality (A.24) follows from (A.23) and property d. after denition A.2.24.

A.6.3. The Hardy spaces

In this subsection, we consider some special classes of frequency-domain functions that are holomorphic on the open half-plane.
Good general references for this section are Kawata [15] and Helson [11] for the scalar case
and Thomas [26] and Rosenblum and Rovnyak [21] for the vector-valued case.

190

A.6. Frequency-domain spaces

Denition A.6.14. For a Banach space X and a separable Hilbert space Z we dene the
following Hardy spaces:
H (X) :=

H2 (Z)

:=

G : C+ X | G is holomorphic, and
0

sup
Re(s)>0

G(s) < ;

f : C+ Z | f is holomorphic; and
0

2
2

= sup(
>0

1
2

f ( + ) 2 d) < .

(A.25)

When the Banach space X or the Hilbert space Z equals C, we shall use the notation H
and H2 for H (C) and H2 (C), respectively. In most of the literature, Hardy spaces on the disc
are usually treated; see, for example, Rosenblum and Rovnyak [21]. The following lemma shows
this is equivalent to considering Hardy spaces on the open right half-plane.
1+z
Lemma A.6.15. Denote by the bilinear transformation (z) = 1z . A function G is an
element of H (X) if and only if f is holomorphic and bounded on the unit disc D. Furthermore,
sup G(s) = sup G((z)) .
sC+
0

zD

Lemma A.6.16. If X is a Banach space, then H (X) from denition A.6.14 is a Banach space
under the H -norm
G := sup G(s) X .
(A.26)
Re(s)>0

Proof Combine Lemma A.6.16 with theorem D of Rosenblum and Rovnyak [21, section 4.7].
We now collect several important results in the following lemma.
Lemma A.6.17. The following are important properties of H (L(U, Y )), where U, Y are separable Hilbert spaces:

a. For every F H (L(U, Y )) there exists a unique function F P ((, ); L(U, Y ))

such that

lim F (x + )u = F ()u

for all u U and almost all R

x0

(i.e., F H (L(U, Y )) has a well dened extension to the boundary);

b. The mapping F F is linear, injective and norm preserving, i.e.,

sup

F (s)

Re(s)>0

L(U,Y )

= ess supR F ()

L(U,Y )

(we identify F H (L(U, Y ) with its boundary function F P ((, ); L(U, Y ))

and we can regard H (L(U, Y )) as a closed subspace of P ((, ); L(U, Y )));

c. Identifying F with F , the following holds:

sup
Re(s)0

F (s)

L(U,Y )

= ess supR F ()

L(U,Y )

< .

191

A. Mathematical Background
Proof Combine Lemma A.6.15 with theorems A of sections 4.6 and 4.7 of Rosenblum and
Rovnyak [21].
We remark that Rosenblum and Rovnyak [21] use L for P . In general, the boundary

function F will not have the property that F is uniformly measurable in the L(U, Y ) topology;
see Rosenblum and Rovnyak [21, exercise 1 of chapter 4] or Thomas [26].
Lemma A.6.18. H2 (Z) is a Banach space under the H2 -norm dened by (A.25), and the following important properties hold:

a. For each f H2 (Z) there exists a unique function f L2 ((, ); Z) such that

lim f (x + ) = f ()

for almost all R

x0

and

lim f (x + ) f ()
x0

L2 ((,);Z)

= 0;

b. The mapping f f is linear, injective, and norm preserving.

(we identify the function f H2 (Z) with its boundary function f L2 ((, );Z) and
regard H2 (Z) as a closed subspace of L2 ((, );Z));
c. For any f H2 (Z) and any > 0 we have that

lim

sup

sC+ ; |s|>

|f (s)| = 0

(A.27)

(sometimes the terminology f (s) 0 as |s| in C+ is used).

Proof a and b. The proof for the scalar case as given by Kawata [15, theorem 6.5.1] is based on
Theorem A.6.13. Since this theorem holds for vector-valued function as well, the proof of parts
a and b is similar to that for the scalar case.
c. See Hille and Phillips [12, theorem 6.4.2].
We remark that in general part c is not true for = 0. From this lemma we deduce the
following result.
Corollary A.6.19. If Z is a separable Hilbert space, then H2 (Z) is a Hilbert space under the
inner product
1
f, g :=
2

f (), g() d.

H2 (Z) is a very special Hilbert space, as is apparent from the following lemma and the PaleyWiener theorem.
Lemma A.6.20. Let Z be a separable Hilbert space and let f H2 (Z) be dierent from the
zero function. Then f is nonzero almost everywhere on the imaginary axis.
Proof Suppose that there is a subset V of the imaginary axis with positive measure such that
f is zero on this set. Then for every z Z, we have that f, z H2 and it is zero on V . This
implies that

| log( f (), z )|
d = .
1 + 2

192

A.6. Frequency-domain spaces

By Theorem 6.6.1 of Kawata [15] this can only happen if f, z is the zero function. Since z Z
was arbitrary, this would imply that f = 0. This is in contradiction to our assumption, and so
the set V cannot have positive measure.
Theorem A.6.21. Paley-Wiener Theorem. If Z is a separable Hilbert space, then under
the Laplace transform L2 ([0, ); Z) is isomorphic to H2 (Z) and it preserves the inner products.
Proof See Thomas [26].
An important consequence is the orthogonal decomposition of L2 ((, ); Z).
Theorem A.6.22. The following holds:
L2 ((, ); Z) = H2 (Z) H2 (Z) ,
where H2 (Z) is the orthogonal complement in L2 ((, ); Z) of H2 (Z) as in denition
A.2.30. H2 (Z) is given by

H2 (Z) = f : C Z | f is holomorphic, and

0

f
Proof It is easy to see that

2
2

= sup(
<0

1
2

f ( + ) 2 d) < .

L2 ((, ); Z) = L2 ((0, ); Z) L2 ((, 0); Z)

(A.28)

(A.29)

and that L2 ((0, ); Z) = L2 ((, 0); Z). Applying the Fourier transform to (A.29) and appealing to Theorems A.6.13 and A.6.21, we obtain that
L2 ((, ); Z) = H2 (Z) F L2 ((0, ); Z) .
By Parsevals equality (A.24) we have that F L2 ((0, ); Z) = H2 (Z) . So it remains to show
equality (A.25).
For h L2 ((0, ); Z) = L2 ((, 0); Z) it is easy to see that the two-sided Laplace transform
of h satises

h(s) = h (s)
for s C ,
0
where h (t) := h(t) and h denotes the one-sided Laplace transform of h . Now h is an
element of L2 ((0, ); Z) and thus by the Paley-Wiener Theorem A.6.21 h (s) is an element of
the space dened in (A.28). On the other hand, if f is an element of the reected H2 (Z) space
dened in (A.28), then f (s) is an element of H2 (Z). Thus by Theorem A.6.21 f (s) = h (s)
for some h L2 ((0, ); Z). Dening h(t) := h (t), it is easy to see that h L((, 0); Z)

and that h(s) = f (s) for s C .

0
Another consequence of the Paley-Wiener Theorem is the following corollary.
Corollary A.6.23. If f (+ ) is in H2 (Z) for some real , then there exists a function h : R+

Z such that e h() L2 ([0, ); Z) and h(s) = f (s) almost everywhere.

193

A. Mathematical Background
Proof By Theorem A.6.21 f () := f ( + ) H2 (Z) is the image under the Laplace transform
of a z L2 ([0, ); Z), z (s) = f (s). Thus from (A.4) we have that h(t) = et z(t) satises

h(s) = f (s).
It is often important to know when a function in H (Z) corresponds to a Laplace-transformable
function. In general, it will only correspond to a distribution; for example, es corresponds to
the delta distribution centred at , (t ). In this direction we have the following result from
Mossaheb [18]; it is a direct consequence of Theorem A.6.21.
Lemma A.6.24. Let g be a holomorphic function on C+ such that sg(s) is bounded on C+ .

on C+ , and e(+)t |h(t)|dt < .

Then for any > 0 there exists an h such that g = h
+

Other sucient conditions for a holomorphic function to be the Laplace transform of a function
in L1 (0, ) can be found in Gripenberg, Londen, and Staans [10].
The Paley-Wiener Theorem A.6.21 gives a complete characterization of all functions with
Laplace transforms in H2 (Z). The following theorem gives a complete characterization in the
frequency domain of all L2 ([0, ); Z) functions with compact support. This result is also called
the Paley-Wiener Theorem.
Theorem A.6.25. Paley-Wiener Theorem. If h L2 ([0, ); Z) has the Laplace transform

h, then necessary and sucient conditions for h to be zero almost everywhere on (t0 , ) are

a. h is an entire function,

b. h H2 (Z), and
c.

h(x + y) Cet0 x for x 0.

Proof Necessity: By the Cauchy-Schwarz inequality (A.1) we see that

t0

e
0

st

h(t)dt

e
0

t0

t0
st

h(t) dt

e2Re(s)t0 1
h
2Re(s)

e2Re(s)t dt
0

h(t) 2 dt
0

L2 ((0,);Z) .

So et h(t) L1 ([0, ); Z) for all R, and by Property A.6.2.a, h is holomorphic on every

+
is entire. Furthermore, we see that for Re(s) < 0, the inequality as stated
C . In other words, h
in c holds. Part b follows directly from Theorem A.6.21.
Suciency: From part b and Theorem A.6.21 we know that is the Laplace transform of an
h
L2 ([0, ); Z) function. Thus it remains to prove that it is zero on (t0 , ). For z Z consider

the scalar function hz := h, z . This clearly satises a, b, and c with H2 (Z) replaced by H2 .
Now by Theorem 19.3 of Rudin [24] it follows that hz is zero almost everywhere on (t0 , ). Since
hz = h, z and Z is separable, we obtain that h is zero almost everywhere on (t0 , ).
The following theorem gives a characterization of bounded operators between frequencydomain spaces.
Theorem A.6.26. Suppose that U and Y are separable Hilbert spaces.

194

A.6. Frequency-domain spaces

a. If F P ((, ); L(U, Y )) and u L2 ((, ); U ), then F u L2 ((, );
Y ). Moreover, the multiplication map F : u F u denes a bounded linear operator
from L2 ((, ); U ) to L2 ((, ); Y ), and
F u
where

L2 ((,);Y )

L2 ((,);U) ,

denotes the norm on P ((, ); L(U, Y )). In fact,

F = sup
u=0

F u L2 ((,);Y )
= F
u L2((,);U)

b. If F H (L(U, Y )) and u H2 (U ), then F u H2 (Y ). Moreover, the multiplication map

F : u F u denes a bounded linear operator from H2 (U ) to H2 (Y ), and
F u
where

H2 (Y )

H2 (U) ,

denotes the norm on H (L(U, Y )). In fact,

F = sup
u=0

F u H2 (Y )
= F
u H2 (U)

c. F P ((, ); L(U, Y )) is in H (L(U, Y )) if and only if F H2 (U ) H2 (Y ).

Proof a. See Thomas [26].
b. It is easy to show that for F H (L(U, Y )) the rst inequality holds. So F F
prove the other inequality, let C+ , y0 Y and f H2 . Consider
0
y0
y0
f,
= F f,
H
H
F
s+ 2
s+ 2
=

1
2
1
2

F ()f (),

y0
+

F ()f (), y0

=
=

F ()f (), y0 Y
f (), F () y0 Y

1
2

f (), F () y0

To

1
d

by Cauchys Theorem A.1.8

1
d

using Cauchys Theorem A.1.8 again

y0
f, F ()
H .
s+ 2

Since the above equality holds for every f H2 , we have that

y0
y0

= F ()
.
F
s+
s+
This implies that F , and Lemma A.3.60 concludes the proof.
F
c. See Thomas [26].
The proof of part b was communicated by George Weiss.
These frequency-domain operators have time-domain counterparts that are isometrically isomorphic under the Fourier or Laplace transform; see, for example, Thomas [26, section 5].

195

A. Mathematical Background
Theorem A.6.27. Suppose that U and Y are separable Hilbert spaces.
a. To every F P ((, ); L(U, Y )) there corresponds a unique shift-invariant, bounded
linear operator F from L2 ((, ); U ) to L2 ((, ); U ). Moreover, F and F are isometrically isomorphic via the Fourier transform

y () = (F u)() = F ()(),

u
which holds for every u L2 ((, ); U ).
b. Moreover, we have in part a that F H (L(U, Y )) if and only if
F L(L2 ([0, ); U ), L2 ([0, ); Y )).
Theorem A.6.28. Suppose that Q is a bounded, shift-invariant, linear operator from L2 (, )
to itself; i.e.,
Q ef () = e Qf ().
Then there exists a function q L (, ) such that (Qf ) () = q()f ().
Proof See theorem 72 of Bochner and Chandrasekharan [3].

196

Bibliography
[1] J-P. Aubin. Applied Functional Analysis. John Wiley and Sons, 1979.
[2] A. V. Balakrishnan. Applied functional analysis. Springer-Verlag, New York, 1976. Applications of Mathematics, No. 3.
[3] S. Bochner and K. Chandrasekharan. Fourier transforms. Number 19 in Annals of mathematics studies. University Press, Princeton, 1949.
[4] H. Cox and E. Zuazua. The rate at which energy decays in a string damped at one end.
Indiana Univ. Math. J., 44(2):545573, 1995.
[5] R.F. Curtain and H.J. Zwart. An Introduction to Innite-Dimensional Linear Systems
Theory. Springer-Verlag, New York, 1995.
[6] J. Diestel and J. J. Uhl, Jr. Vector measures. American Mathematical Society, Providence,
R.I., 1977. With a foreword by B. J. Pettis, Mathematical Surveys, No. 15.
[7] Gustav Doetsch. Introduction to the theory and application of the Laplace transformation.
Springer-Verlag, New York, 1974. Translated from the second German edition by Walter
Nader.
[8] N. Dunford and J.T. Schwartz. Linear Operators, volume 1. John Wiley and Sons, 1959.
[9] Y. Le Gorrec, H. Zwart, and B. Maschke. Dirac structures and boundary control systems
associatesd with skew-symmetric dierential operators. SIAM J. Control Optim., 44:1864
1892, 2005.
[10] G. Gripenberg, S.-O. Londen, and O. Staans. Volterra integral and functional equations,
volume 34 of Encyclopedia of Mathematics and its Applications. Cambridge University
Press, Cambridge, 1990.
[11] H. Helson. Harmonic Analysis. Addison Wesley, 1983.
[12] E. Hille and R.S. Phillips. Functional Analysis and Semigroups, volume 31. American
Mathematical Society Colloquium Publications, Providence, R.I., 1957.
[13] Morris W. Hirsch and Stephen Smale. Dierential equations, dynamical systems, and linear
algebra. Academic Press [A subsidiary of Harcourt Brace Jovanovich, Publishers], New
York-London, 1974. Pure and Applied Mathematics, Vol. 60.
[14] T. Kato. Perturbation Theory of Linear Operators. Springer Verlag, New York, 1966.
[15] Tatsuo Kawata. Fourier analysis in probability theory. Academic Press, New York, 1972.
Probability and Mathematical Statistics, No. 15.
[16] E. Kreyszig. Introductory Functional Analysis with Applications. John Wiley and Sons,
1978.
[17] Norman Levinson and Raymond M. Redheer. Complex variables. Holden-Day Inc., San
Francisco, Calif., 1970.

197

Bibliography
[18] S. Mossaheb. On the existence of right-coprime factorization for functions meromorphic in
a half-plane. IEEE Trans. Automat. Control, 25(3):550551, 1980.
[19] Arch W. Naylor and George R. Sell. Linear operator theory in engineering and science.
Holt, Rinehart and Winston, Inc., New York, 1971.
[20] J.W. Polderman and J.C. Willems. Introduction to Mathematical Systems Theory. Texts in
Applied Mathematics. Springer-Verlag, 1998.
[21] M. Rosenblum and J. Rovnyak. Hardy Classes and Operator Theory. Oxford University
Press, New York, 1985.
[22] W. Rudin. Principles of Mathematical Analysis. International Series in Pure and Applied
Mathematics. McGraw-Hill Book Co., New York, 1976.
[23] Walter Rudin. Functional analysis. McGraw-Hill Book Co., New York, 1973. McGraw-Hill
Series in Higher Mathematics.
[24] Walter Rudin. Real and complex analysis. McGraw-Hill Book Co., New York, second edition,
1974. McGraw-Hill Series in Higher Mathematics.
[25] Angus E. Taylor. Introduction to functional analysis. John Wiley & Sons Inc., New York,
1958.
[26] E.G.F. Thomas. Vector-valued integration with applications to the operator valued H
space. Journal of Mathematical Control and Information, 14:109136, 1997.
[27] G. Erik F. Thomas. Totally summable functions with values in locally convex spaces. In
Measure theory (Proc. Conf., Oberwolfach, 1975), pages 117131. Lecture Notes in Math.,
Vol. 541. Springer, Berlin, 1976.
[28] J.A. Villegas. A Port-Hamiltonian Approach to Distributed Parameter Systems. PhD thesis,
Department of Applied Mathematics, Enschede, The Netherlands, May 2007. Availible at
http://doc.utwente.nl.
[29] J.A. Villegas, H. Zwart, Y. Le Gorrec, and B. Maschke. Exponential stability of a class of
boundary control systems. IEEE Trans. AC, 2009. to appear.
[30] G. Weiss. Regular linear systems with feedback. Mathematics of Control, Signals and
Systems, 7:2357, 1994.
[31] K. Yosida. Functional Analysis. Springer Verlag, 5th edition, 1978.
[32] H. Zwart. Transfer functions for innite-dimensional systems. Systems & Control Letters,
52:247255, 2004.
[33] H. Zwart, Y. Le Gorrec, B. Maschke, and J. Villegas. Well-posedness and regularity of
hyperbolic boundary control systems on a one-dimensional spatial domain. ESAIM, 2009.
to appear.

198

Index
H2 (Z), 191
H , 191
H (X), 191
>, for operators, 157
C[0, 1], 126
C([a, b]; X), 137
C0 -group, 21
C0 -semigroup, 21
D(T ), 133
H 1 ((a, b); Rn ), 28
L(; Z), 176
L2 ((, ); Z), 187
Lp (a, b), 125
Lp (; Z), 176
L (; Z), 176
L (a, b), 125
P (; L(Z1 , Z2 )), 176
Pp (; L(Z1 , Z2 )), 176
P ((, ); L(U, Y )), 187
P (; L(Z1 , Z2 )), 176
Q , 145
V , 130
X , 140
X , 144
Z , 131
, 185, 186

h, 185
p , 124
, 124
, for operators, 157

h, 184
, 137
ker T , 134
, , 127
C+ , 184

H 2 , 191
V , 125
C+ , 184

, 124
, 129
ran T , 134

(A), 159
c (A), 161
r (A), 161
(A), 161
p (A), 161
r (T ), 165
L(X), 135
L(X, Y ), 135
anti-symmetric, 8
approximation
in L , 188
of holomorphic functions, 120
Banach Steinhaus theorem, 138
beam
Euler-Bernoulli, 4
Rayleigh, 4
Timoshenko, 4
Bode plot, 69
bond space, 11
boundary control system, 47
boundary eort, 27
boundary ow, 27
boundary operator, 46
boundary port variables, 14
bounded
operator, 134
set, 125
Burgers equation, 10
Cauchys residue theorem, 119
Cauchys theorem, 119
classical solution, 44
boundary control system, 47
closed
operator, 146
set, 125
closed curve, 118
closed graph theorem, 148
compact
operator, 138

199

Index
set, 125
continuous
strongly, 137
uniformly, 137
continuous on D(F ), 134
contour
closed, 118
positively oriented, 119
simple, 118
contraction mapping theorem, 133
contraction semigroup, 25
convergence
strong, 137
uniform, 137
weak, 144
curve
closed, 118
rectiable, 118
simple, 118
C0 -semigroup, 21
growth bound, 35
measurable, 173
derivative
Frchet, 180
e
dierentiable
strongly, 181
uniformly, 181
weakly, 182
dierential, see Frchet dierential
e
Dirac structure, 12
domain, 24
complex, 117
of an operator, 133
eort, 12
eort space, 12
eigenfunction, see eigenvector
eigenvalue
isolated, 161
multiplicity, 161
order, 161
eigenvector
generalized, 161
Euler-Bernoulli beam, 4
exponential solution, 62
exponentially stable, 93
feed-through, 77
ow, 12

200

ow space, 12
formal adjoint, 106
Fourier transform
inverse, 189
Frchet derivative, 180
e
Fubinis theorem, 179
functional, 132
Gronwalls lemma, 187
group
C0 , 21
strongly continuous, 21
unitary, 25
growth bound, 35
Hahn-Banach theorem, 140
Hamiltonian, 14
Hardy space, 191
heat conduction, 6, 105
Hlder inequality, 142
o
homotopic, 122
ind(g), 121
index, see Nyquist index
innitesimal generator, 24
integral
Bochner, 173
complex, 119
Lebesgue, 172
Pettis, 175
invariant
shift, 196
inverse
algebraic, 134
bounded, 160
inverse Fourier transform, 189
inviscid Burgerss equation, 10
isolated eigenvalue, 161
isomorphic
isometrically, 125
topologically, 125
ker, 134
Laplace transform
two-sided, 185
Lebesgue-dominated convergence theorem, 178
linear, 62
linear functional
bounded, 139

Index
linear space, see linear vector space
normed, 124
Liouvilles theorem, 118
measurable
of semigroups, 173
strong, 172
uniform, 172
weak, 172
mild solution, 45
boundary control system, 48
multiplicity
algebraic, 161
nonzero limit at in C+ , 121
0
norm
equivalent, 125
induced by inner product, 127
operator, 134
Nyquist plot, 69
Nyquist theorem, 121
open mapping theorem, 138
operator
adjoint
bounded, 152
unbounded, 154
algebraic inverse, 134
bounded, 134
closed, 146
coercive, 157
compact, 138
dual
bounded, 145
unbounded, 149
nite rank, 134
inverse, 134, 160
linear, 133
nonnegative, 157
norm, 134
positive, 157
self-adjoint, 156
square root, 157
symmetric, 156
unbounded, 146
order
of a pole, 119
of a zero, 118
orthogonal projection, 158
orthogonal projection lemma, 158

Paley-Wiener theorem, 193, 194

poles, 119
positive real, 69
power, 12
power product, 12
principle of the argument, 120
ran , 134
Rayleigh beam equation, 4
regular, 77
Riemann-Lebesgue lemma, 185
Riesz representation theorem, 149
Rouchs theorem, 118
e
self-adjoint
spectrum, 165
semigroup
C0 , 21
contraction, 25
strongly continuous, 20
set
bounded, 125
closed, 125
compact, 125
dense, 125
maximal, 130
open, 125
orthogonal, 130
relatively compact, 125
Sobolev space, 24
solution, 25, 45
classical, 44
boundary control system, 47
exponential, 62
mild, 45
boundary control systems, 48
spectrum
continuous, 161
point, 161
residual, 161
stable
exponentially, 93
strongly, 93
state, 21
state space, 21, 26
strong convergence, 137
strongly continuous group, 21
strongly continuous semigroup, 20
strongly measurable, 172
strongly stable, 93

201

Index
suspension system, 16
system
general, 62
time-invariant, 62
Timoshenko beam, 4
topological dual space, 140
transfer function, 62
regular, 77
transfer function at s, 62
transmission line, 1, 102
transport equation
controlled, 43
uniform boundedenitioness theorem, 138
uniformly measurable, 172
uniqueness of the Laplace transform, 184
unitary group, 25
variation of constant formula, 44
vector space
complex, 123
linear, 123
real, 123
vibrating string, 3
weak convergence, 144
weakly measurable, 172
well-posed, 75
well-posedness, 73
zero, 118

202